Reinforcement learning with low-rank adaptation for targeted antimicrobial peptide design

Abstract

Antimicrobial peptides (AMPs) are emerging as promising alternatives to traditional antibiotics, offering solutions to antimicrobial resistance through diverse mechanisms. Despite their potential, current computational approaches for AMP design rarely address strain-specific targeting, limiting their clinical efficacy as bacterial strains exhibit unique membrane compositions and susceptibility profiles requiring tailored interventions. Furthermore, the limited availability of strain-specific training data presents a significant challenge, necessitating parameter-efficient learning approaches that can optimize AMP properties with minimal overfitting. This study introduces a novel AMP generation framework that integrates reinforcement learning with a Generative Pre-trained Transformer (GPT) model enhanced by Low-Rank Adaptation (LoRA) parameter-efficient fine-tuning. This approach enables the design of peptides optimized for multiple objectives, specifically antimicrobial activity and toxicity, tailored to individual pathogen strains. Our framework employs a two-stage learning process: pretraining on a large-scale peptide and AMP database to capture linguistic and contextual features, followed by reinforcement learning that leverages MIC (Minimum Inhibitory Concentration) and hemolysis prediction models to optimize antimicrobial potency and safety profiles. The integration of LoRA is crucial for efficiently adapting the model to strain-specific characteristics while addressing limited training data. The comparative analysis demonstrated our model’s superior performance over existing AMP generation approaches in both activity and hemolytic toxicity metrics. An ablation study confirmed the contributions of reinforcement learning and LoRA. Furthermore, the model can generate peptides satisfying both activity and toxicity conditions for unseen strains, highlighting its capability to design AMPs for emerging pathogens. In addition, molecular dynamics (MDs) simulations confirmed that the generated peptides penetrate bacterial membranes, supporting antimicrobial activity. © The Author(s) 2025. Published by Oxford University Press.