Lead Informed Artificial Intelligence Mining of Antitubercular Host Defense Peptides.

Abstract

Identifying host defense peptides (HDPs) that are effective against drug-resistant infections is challenging due to their vast sequence space. Artificial intelligence (AI)-guided design can accelerate HDP discovery, but it traditionally requires large data sets to operationalize. We report an AI workflow that utilizes limited data sets (∼100 peptides) to uncover potent, selective, and safe HDPs by informing selection through lead candidate mutational scanning. This approach, referred to as Lead Informed Machine Interrogation of Therapeutic Sequences (LIMITS), is applied against the exemplary pathogen Mycobacterium tuberculosis. Experimental validation of predicted sequences shows nearly an order of magnitude improvement in potency, selectivity, and safety, relative to the initial template. Post hoc analysis suggests sequence length may be a unique and underappreciated driver of antitubercular HDP activity. These results demonstrate that, with continued development, the LIMITS approach can identify selective HDPs under data-limited conditions and elucidate structure-function-performance relationships previously hidden in biologic complexity.