Quantitative Analysis of the Population Dynamics of Antibiotic Responses

Abstract

Antibiotics are a cornerstone of modern medicine. In the eighty years since they became widely available, broad-spectrum antibiotics have allowed clinicians to administer them empirically, treating a wide variety of infectious diseases with one-size-fits-all approaches. However, the availability of antibiotics has accelerated the prevalence of antibiotic resistance. The practice of antimicrobial stewardship, intended to slow resistance development, has also further muddied already limited incentives for new antibiotic development, contributing to an ever-thinning pipeline of new treatments. In this context, learning how to deploy antibiotics in ways that minimize resistance development is crucial for maximizing the effectiveness of existing antibiotics and guiding the design of new ones. Given the many ways that bacterial populations can respond to antibiotics, a quantitative understanding of their dynamics is necessary to identify the strains, drugs, and situations for which this is possible. In this dissertation, I used high-throughput growth curve generation, mathematical modeling, and population composition measurements to gain insight into the population dynamics of antibiotic treatment. I first developed a pipeline for high-throughput measurement of bacterial responses using liquid handling. Focusing on responses to treatment with beta-lactams, the most commonly prescribed antibiotics, I collected two datasets comprising over 20,000 growth curves in total: one measuring the response of 311 clinical isolates to a set concentration of amoxicillin with and without the Bla inhibitor clavulanic acid, and another collecting dose-response matrices of laboratory strains treated with amoxicillin in combination with three different Bla inhibitors. I found a wide variety in antibiotic responses across strains. The information content of these responses is sufficient for applications such as strain identification and the prediction of responses to novel antibiotics or drug concentrations. Focusing on the response to beta-lactam/beta-lactamase (Bla) inhibitor combinations, I developed a mathematical model of a mixed population with sensitive and resistant subpopulations, responding to treatment with a beta-lactam and an inhibitor of beta-lactam-degrading Bla enzymes. Because Bla-mediated antibiotic degradation benefits both the resistant Bla-producing cells and sensitive subpopulations, but is localized to resistant cells, it functions as a partially private public good. Inhibiting it can thus produce non-intuitive selection dynamics. Using this model, I found that different dose ratios of antibiotic to inhibitor were sufficient to generate different selection dynamics. However, I also identified strain- and drug-specific factors that governed selection dynamics, including the burden, private benefit, and inhibition of the private benefit of producing Bla. I derived a criterion, depending on these factors, that can predict the selection dynamics of different simulated parameter sets.I then use engineered laboratory strains to verify that the privatization of Bla-mediated antibiotic degradation increases selection for resistance in synthetic strains. Using my dataset of clinical isolates, I then used nonlinear optimization to fit model parameters that could recapture the experimental data for each isolate. By conducting competition experiments of clinical isolates with laboratory strains, I found that isolates with higher estimated private benefit were generally selected for during combination treatment, while strains with lower private benefit were generally selected against. This work underscores how quantitative phenotypic characterization can yield clues for guiding treatment strategy.Finally, in the Appendix, I apply some of these approaches to collaborative work understanding the selection dynamics of a resistance-targeting prodrug currently under investigation and characterizing communities in a hospital sink environment.