Prediction of Antibiotic Resistance Evolution by Growth Measurement of All Proximal Mutants of Beta-Lactamase

The antibiotic resistance crisis continues to threaten human health. Better predictions of the evolution of antibiotic resistance genes could contribute to the design of more sustainable treatment strategies. However, comprehensive prediction of antibiotic resistance gene evolution via laboratory approaches remains challenging. By combining site-specific integration and high-throughput sequencing, we quantified relative growth under the respective selection of cefotaxime or ceftazidime selection in similar to 23,000 Escherichia coli MG1655 strains that each carried a unique, single-copy variant of the extended-spectrum beta-lactamase gene bla(CTX-M-14) at the chromosomal att HK022 site. Significant synergistic pleiotropy was observed within four subgenic regions, suggesting key regions for the evolution of resistance to both antibiotics. Moreover, we propose PEAR(P) and PEA(R), two deep-learning models with strong clinical correlations, for the prospective and retrospective prediction of bla(CTX-M-14) evolution, respectively. Single to quintuple mutations of bla(CTX-M-14) predicted to confer resistance by PEAR(P) were significantly enriched among the clinical isolates harboring bla(CTX-M-14) variants, and the PEAR(R) scores matched the minimal inhibitory concentrations obtained for the 31 intermediates in all hypothetical trajectories. Altogether, we conclude that the measurement of local fitness landscape enables prediction of the evolutionary trajectories of antibiotic resistance genes, which could be useful for a broad range of clinical applications, from resistance prediction to designing novel treatment strategies.