B1633
Title: An ensemble approach to feature identification and prediction of antimicrobial peptide activity
Authors: Nawisa Jullapech - University of Reading (United Kingdom) [presenting]
Fazil Baksh - University of Reading (United Kingdom)
Zuowei Wang - University of Reading (United Kingdom)
Abstract: Antimicrobial peptides (AMPs) are attracting continuous attention for their biocompatibility and strong potential in inhibiting the growth and replication of bacterial cells and combating multidrug-resistant pathogens. Crucial to effective AMP design is identifying the intrinsic relationships between peptide sequences, resulting physicochemical features, and their antibacterial activities. Machine learning (ML) has emerged as a powerful tool in tackling this target. A novel ensemble feature selection method is reported that combines a diverse group of ML algorithms with the best subset selection for identifying key physicochemical features governing the antimicrobial activity of AMPs. Using the DBAASP database with target pathogen Gram-negative bacteria E. coli ATCC 25922, the approach achieves prediction accuracy above 85\% on the antimicrobial activities of both 10-16 and 18-27 aa peptides, higher than those reported in previous ML studies of the same types of AMPs. The results further reveal that hydrophobicity, net charge and isoelectric point are essential physicochemical properties in determining AMPs antimicrobial activities. This finding is consistent with, and so confirms experimental suggestions. Finally, the ML models developed are used to construct a complete antimicrobial activity phase diagram over the multi-dimensional physiochemical factor space potentially accessible by experiments, providing useful information to guide the design of novel AMPs.
