Multi-Property Optimization of Antimicrobial Peptides Using Reinforcement Learning and Conditional Independence Regularization.

Abstract

Antimicrobial peptides (AMPs) have emerged as a compelling alternative to conventional antibiotics, offering broad-spectrum activity against bacterial and viral infections while addressing antimicrobial resistance. However, the design of AMPs that simultaneously satisfy multiple functional properties remains a critical challenge, hindered by the high-dimensional complexity representation of their latent space. In this study, we propose CAMP-RL, a conditional generative model for de novo AMP generation and optimization. CAMP-RL integrates reinforcement learning (RL) and conditional independence regularization (CIR) within a conditional VAE framework to enhance controllability of AMP generation over the multi-properties optimization. By incorporating RL rewards for property-guided generation and CIR to enforce orthogonality among latent dimensions, CAMP-RL enables decomposition of AMP latent space into conditionally independent subspaces, facilitating precise multi-objective optimization. Comprehensive evaluations demonstrate that CAMP-RL outperforms existing state-of-the-art methods in both generating novel AMPs and optimizing existing AMPs. This work advances computational peptide design, offering a promising solution for developing novel antimicrobial therapeutics.