A generative artificial intelligence approach for peptide antibiotic optimization

TL;DR

ApexGO is a generative AI method that optimizes antimicrobial activity by modifying the sequences of existing peptide antibiotics.

Problem

With the rise of antibiotic-resistant bacteria, existing AI methods are biased toward fixed library screening or random generation, limiting their ability to optimize existing peptide scaffolds under practical design constraints.

Approach

A transformer variational autoencoder embeds peptide sequences in a continuous latent space, and Bayesian optimization efficiently proposes sequence edits to boost antimicrobial potency. Optimized derivatives are generated through modifications of template peptides.

Results & Contribution

From 10 template peptides, 100 derivatives were synthesized and tested, achieving an 85% hit rate and a 72% success rate in activity enhancement. It outperformed previous methods against Gram-negative bacteria, and in mouse infection models, it showed stronger anti-infective effects than template controls and comparable or superior effects to last-resort antibiotics.