Membranes and Machine Learning: Designing a Model of Antibiotic Activity to Bypass Gram Negative Membranes and Efflux Pumps

Due to the growing prevalence of antibiotic-resistant bacteria, there is a pressing need for rapid design of new antibiotics with unique modes of action. Gram negative bacteria in particular pose a thorny problem for antibiotic design due to the combined effects of their impermeable outer membranes and their antibiotic-removing efflux pumps. We employ a combined theoretical and experimental approach to understand the limiting factors on antibiotic efficacy in p. aeruginosa and rationally design experiments to rapidly zero in on promising antibiotic candidates. By using machine learning to identify a relevant subset of descriptors produced by simulation and experiment for predicting experimentally-measured minimum inhibitory concentrations of different antibiotic candidates, we are able to learn a self-consistent model of antibiotic retention based on a simple kinetic approximation. We separately employ a Gaussian process regressor to direct a search for antibiotic candidates with optimal experimental and simulated properties. This work leads to heightened understanding of the qualities important to antibiotic retention in Gram negative bacteria and offers a simple way to narrow the experimental design space of antibiotic candidates, allowing for rapid, high-throughput screening.