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

Area of Science:

• Computational chemistry
• Bioinformatics
• Artificial intelligence in drug discovery

Background:

• Antimicrobial peptides (AMPs) show promise against antibiotic-resistant bacteria but lack strain-specific design.
• Current computational methods struggle with limited strain-specific data and optimizing multiple peptide properties.
• Tailored AMPs are needed to address unique bacterial membrane compositions and susceptibility profiles.

Purpose of the Study:

• To develop a novel AMP generation framework for strain-specific optimization of antimicrobial activity and toxicity.
• To address the challenge of limited strain-specific training data using parameter-efficient learning.
• To design peptides with improved efficacy and reduced hemolytic toxicity for targeted bacterial infections.

Main Methods:

• Integration of reinforcement learning with a Generative Pre-trained Transformer (GPT) model.
• Enhancement of the GPT model using Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning.
• Two-stage learning: pretraining on peptide/AMP data, followed by RL using MIC and hemolysis prediction models.

Main Results:

• The novel framework demonstrated superior performance in antimicrobial activity and reduced hemolytic toxicity compared to existing methods.
• Reinforcement learning and LoRA were confirmed as key contributors to the model's effectiveness via ablation studies.
• The model successfully generated peptides with desired activity and toxicity profiles for previously unseen bacterial strains.
• Molecular dynamics simulations validated the generated peptides' ability to penetrate bacterial membranes.

Conclusions:

• The proposed framework enables efficient, strain-specific design of antimicrobial peptides with optimized activity and safety.
• This approach effectively overcomes data limitations for designing targeted antimicrobial therapies.
• The study highlights a significant advancement in computational drug discovery for combating emerging infectious diseases.