Computationally designing antimicrobial peptides for Acinectobacter Baumannii

Abstract

Acinetobacter Baumannii, which is mostly contracted in hospital stays, has been developing resistance to all available antibiotics, including the last line of drugs, such as carbapenem. Because of its quick adaptation there is an immediate need to design new antibiotics, possibly antimicrobial peptides (AMPs) to which bacteria do not develop resistance easily. Our threefold goal was to curate the available activity of AMPs on the same strain of A. Baumannii, build a neural network model for predicting their activity and use it to rationally pre-screen for lead generation from the thousands of naturally occurring AMPs. By curating and analyzing the recent activity data from 81 AMPs on ATCC 19606 strain, we develop a quantitative AMP activity prediction model. We selected three other models with comparable performance against a test set with known activities. With the goal of inspiring further studies on AMP drug candidates and their rational shortlisting, we made activity predictions for the entire database of AMPs using all the models. To handle the uncertainty of training with a small data set, highlighted peptides which had consistent results from all models.