AI-guided design of new antimicrobial peptides

Antimicrobial resistance (AMR) is a growing threat to global health, driven by the overuse of antibiotics and the lack of novel therapeutic agents. Antimicrobial peptides (AMPs) have emerged as promising alternative to traditional small molecule-based antibiotics due to their broad-spectrum activity and reduced susceptibility to resistance. Recent advances in generative deep learning provided new tools for the design of AMPs allowing rationalization and speeding up the process of their development. While showing efficacy in generating AMPs, such methods often lack interpretability required to efficient used by domain experts. In this study, we build upon a previously established workflow that combines a Wasserstein Autoencoder (WAE) with Generative Topographic Mapping (GTM) — a visualization technique that enables an interpretable sampling from the latent space. The method was applied for the design of AMPs against Staphylococcus aureus, achieving a high experimental hit rate. However, the original method was limited to in-house data on anti-S. aureus peptides containing 10 to 14 residues. In this work, the approach was extended to 14 bacterial species and 15 bacterial strains, using peptide sequences with an extended range of lengths (6-25 amino acid residues). In addition to that, various sequence representation methods, including 2-mers, 3-mers, protein large language model (pLLM) were investigated. The extended workflow enables to identify the zones associated with broad-spectrum antimicrobial activity on GTM landscapes and to guide targeted sampling from these zones. For in silico predictions of peptides activity, Support Vector Machine (SVM), CatBoost and Random Forest (RF) machine learning models were used. Based on the activity predictions, 10 peptides were selected and validated in vitro by a partner laboratory. In total 5 out of 10 peptides possessed at least 1 activity: 3 of these peptides were active against Staphylococcus epidermidis, and 2 of them showed broad-spectrum antimicrobial activity against S. epidermidis and S. aureus. This approach has demonstrated the potential of this interpretable generative framework for the discovery of broad-spectrum AMPs.