A generative artificial intelligence approach for peptide antibiotic optimization
A generative artificial intelligence approach for peptide antibiotic optimization
https://www.nature.com/articles/s42256-026-01237-5
Publish Date: 2026-05-13 05:18:00
Source Domain: www.nature.com
- ApexGO is an optimization framework composed of three components: a generative model to produce peptides, an oracle model called APEX to predict Minimal Inhibitory Concentrations (MIC) of peptides against multiple bacterial pathogens, and an optimization algorithm based on Bayesian Optimization (BO).
- APEX 1.1 oracle model predicts MICs for 11 bacterial pathogens and guides the BO to produce peptides with high antimicrobial activity.
- The optimization leverages a deep generative model, specifically a VAE, trained on large sets of amino acid sequences to map peptides onto a continuous latent space for BO in a discrete domain.
- They employ Trust Region BO to limit over-exploration and improve efficiency by restricting the search to a hyper-rectangular trust region in the latent space.
- The optimization is constrained to maintain at least 75% similarity to the initial template peptide to derive peptides previously validated in terms of antimicrobial activity.
- In addition to optimizing for single peptides, they simultaneously optimize a set of 20 unique peptides within different trust regions to increase chances of success after laboratory validation.
- The optimization aims to find peptides with low MIC values as predicted by APEX, and they benchmark their framework against existing generative BO approaches and deep generative AMP models.
- They demonstrate that their optimizer can efficiently discover peptides with antimicrobial activity through experiments in bacterial cultures and mouse models.
