Artificial Intelligence based de novo cyclic antimicrobial peptide generation

Materials and methods:

A generative model will be trained on a publicly available peptide corpora, then adapted toward antimicrobial space using curated activity data from public antimicrobial peptide databases. Different strategies will be compared for conditioning, which will be implemented through a scorer that combines predicted antimicrobial activity, predicted haemolysis and ring closure feasibility. Generated candidates (total: 50,000) will be screened based on novelty, sequence and scaffold diversity, distance to known cyclic antimicrobial peptides, synthetic accessibility and conformational plausibility to give a prioritised set of 50 candidates. This prioritised set will be forwarded to molecular dynamics simulation of peptide-membrane interaction, and highest-ranked candidates will be sent for chemical synthesis and MIC determination against representative Gram-positive and Gram-negative strains.