Simple birth-death-mutation models predict some—but not all—aspects of the experimental evolution of antibiotic resistance

Mathematical modelling of antibiotic resistance plays an important role in understanding the mechanisms of resistance emergence and spreading, testing the feasibility of new treatment protocols, and antimicrobial stewardship. However, many assumptions underlying some of the most commonly used mathematical models have not been rigorously tested experimentally. We verify whether one of these models - a birth-death-mutation process - is able to quantitatively predict the outcome of laboratory experiments. We grow bacteria in a bioreactor in conditions that closely resemble the assumptions of the model, and compare the model predictions with experimental observables such as the probability and time to resistance evolution, mutant number distribution, and the genetic composition of the evolved populations. We show that the model fails to reproduce some aspects of the experiments (failing differently for different antibiotics) but that simple modifications of the model significantly improve its predictive power. These modifications give insight into the population dynamics of resistant mutants for each antibiotic tested, and highlight the importance of quantitative modelling for accurate prediction of antibiotic resistance evolution.Author summary: Mathematical models are central to the study of antimicrobial resistance (AMR) because they can test mechanistic concepts, reveal processes that are difficult to observe experimentally, and predict the evolution of resistance. Surprisingly, however, these models are rarely tested against quantitative data. Here we rigorously test the predictions of one of the most widely used AMR models—the birth-death process with mutations. Using a bioreactor, we perform carefully controlled experiments in which bacteria evolve resistance to three different antibiotics. The standard, widely-used birth-death-mutation model cannot reproduce our data but augmenting it with antibiotic-specific mechanistic information can significantly increase its predictive power. Our work highlights the necessity of quantitative comparisons between AMR models and experiments. It also shows what additional factors must be considered to improve the predictions of such models, ultimately helping researchers develop new treatments that benefit from rather than being thwarted by microbial evolution.