Uncovering Key Characteristics of Antibacterial Peptides through Machine Learning

Antimicrobial peptides (AMPs) have emerged as promising alternatives to traditional antibiotics in addressing the growing threat of multi-drug-resistant (MDR) bacteria-a crisis that could lead to millions of deaths over the next three decades if left unaddressed. While the general role of cationic and hydrophobic interactions in AMP-mediated bacterial killing is well established, the distinctions between structural characteristics of AMPs targeting different types of bacteria remain underexplored. To address this issue and streamline the design of potent AMPs depending on the bacterial structure, machine learning (ML) models were employed on AMPs targeting Gram-negative bacteria (Pseudomonas aeruginosa PAO1), Gram-positive bacteria (Staphylococcus aureus ATCC 29213), and mycobacteria (Mycobacterium tuberculosis H37Rv and Mycobacterium smegmatis mc² 155) to derive important features that determine antimicrobial efficacy. The Random Forest models mainly reveal that AMPs with a cLogP of less than -6 and a net-charge limited to +4 with variations of the hydrophobic composition in between 20%-50% (20%-40% against P. aeruginosa PAO1, 30%-50% against S. aureus ATCC 29213, 35%-45% against M. tuberculosis H37Rv and M. smegmatis mc² 155) and variations of the cationic composition in between 10%-40% (10%-20% against P. aeruginosa PAO1, 30%-40% against S. aureus ATCC 29213, 10%-30% against M. tuberculosis H37Rv and M. smegmatis mc² 155) predicts significant antibacterial activity. The feature characteristics of the three bacterial types may directly relate to their distinct cell envelope structures and aid in mode of action postulation. This work demonstrates how ML can effectively inform AMP design by accounting for microbial structural differences and underscores its broader potential in customizing peptides for specific bacterial strains.