Estimating the effect of antimicrobial resistance genes on minimum inhibitory concentration in Escherichia coli

Background Surveillance and prediction of antibiotic resistance in Escherichia coli relies on curated databases of genes and mutations. Such databases currently lack quantitative data estimating the effect on MIC caused by the acquisition of any given element for a particular antibiotic-species combination. Methods Using a collection of 2875 E. coli isolates with linked whole genome sequencing and MIC data, we used multivariable interval regression models to estimate the change in MIC for specific antibiotics associated with the acquisition of genes and mutations in the AMRFinder database with and without an adjustment for population structure. We then tested the ability of these models to predict MIC and binary resistance/susceptibility using leave-one-out cross validation. Findings We provide quantitative estimates (with confidence intervals) of the change in MIC associated with the acquisition of genes/mutations in the NCBI-AMRFinder database. Whilst the majority of genes and mutations (89/111 (80.2%) were associated with an increased MIC, a much smaller number (27/111, 24.3%) were found to be putatively independently resistance conferring (i.e. associated with an MIC above the EUCAST breakpoint) when acquired in isolation. We found evidence of differential effects of acquired genes and mutations between different generations of cephalosporin antibiotics and demonstrated that sub-breakpoint variation in MIC can be linked to genetic mechanisms of resistance. 20,697/24,858 (83.3%, range 52.9-97.7 across all antibiotics) of MICs were correctly exactly predicted and 23,677/24,858 (95.2%, range 87.3-97.7) to within +/-1 doubling dilution. Interpretation Quantitative estimates of the independent effect on MIC of the acquisition of antibiotic resistance genes add to the interpretability and utility of existing databases. Using these estimates to predict antibiotic resistance phenotype demonstrates performance that is comparable to or better than approaches utilising machine learning models and crucially more readily interpretable. The methods outlined here could be readily applied to other antibiotic/pathogen combinations. Funding This work was funded by the NIHR and the MRC. RESEARCH IN CONTEXT Evidence before this study We searched PubMed from inception to 05/04/2024 using the terms (( Escherichia coli OR E. coli ) AND ((MIC) OR (minimum inhibitory concentration))) AND (predict*) AND (whole genome sequencing). Of the 56 articles identified by these search terms, eight were of direct relevance to this study. These studies generally focused on single antibiotics (3 studies), had relatively small datasets (6 studies ¡1000 isolates) or used machine learning approaches on pan-genomes to predict binary (i.e. susceptible/resistant) phenotypes (2 studies). Only one study attempted to predict ciprofloxacin MICs in 704 E. coli isolates using a machine learning approach with known resistance conferring genes/mutations as features. To our knowledge, there are no studies estimating the independent effect (as opposed to the total effect of all elements present) of the acquisition of specific antibiotic resistance genes (ARGs) or resistance-associated mutations on MICs of different antibiotics in E. coli more generally. What this study adds In this study we estimate the change in MIC for particular antibiotics associated with the acquisition of specific ARGs or resistance-associated mutations, adjusting for the presence of other relevant genes and population structure. In doing so we provide an approach to greatly enhance the information provided by existing ARG databases and approaches based on predicting binary susceptible/resistant phenotypes, for example by demonstrating differential effects of ARGs on resistance to antibiotics of the same class, enriching our understanding of the relationship between genotype and phenotype in a way that is easily interpretable. Using more “parsimonious” models for prediction, we demonstrate high overall accuracy comparable to or better, and crucially more readily interpretable, than recent machine learning models. We also demonstrate a genetic basis behind sub-breakpoint variation in MIC for some antibiotics, demonstrating the value of non-dichotomised phenotypes for identifying wildtype isolates (i.e. those carrying no ARGs) with greater confidence. Implications of all available evidence Whole genome sequencing data can be used to predict MICs for most commonly used antibiotics for managing E. coli infections with accuracy approaching that of conventional phenotyping techniques, though very major error rates remain too high for deployment in routine clinical practice. Further studies focusing on genotypes with high phenotypic heterogeneity should investigate the phenotypic replicability, genetic heritability and clinical outcomes associated with these isolates.