Ruling resistance out, and reading the blank: genome-based antimicrobial susceptibility prediction in clinical Pseudomonas aeruginosa isolates

Abstract

Genomic prediction of antimicrobial susceptibility aims to identify resistance determinants and thereby to report those antibiotics for which therapeutic failure is to be expected. We investigated what this approach can achieve in Pseudomonas aeruginosa, a species in which resistance often arises through changes in gene expression. For this purpose, 5,749 consecutive clinical isolates from the routine laboratories of two tertiary care centers were sequenced, each with paired phenotypic susceptibility test results for ciprofloxacin, ceftazidime, meropenem, and tobramycin. Genomic variation could be interpreted only after three adaptations. Variants were called against the population-wide major allele rather than against PAO1. This removes markers that are artefacts of the reference. Loss of gene function was captured systematically, as many genetic determinants of resistance act through loss-of-function mechanisms. Furthermore, we introduced a specificity ratio - enabled by testing all four antibiotics on the same isolates - which separated markers associated with resistance mechanisms from those that merely tracked clonal lineage; without this distinction, expanding the catalogue reduced rather than improved predictive performance. In addition to classical resistance determinants, markers of susceptibility were taken into account, such as the functional loss of an efflux pump that exports the antibiotic out of the cell. While resistance could only be assigned with corresponding confidence for a few percent of isolates, the absence of ciprofloxacin and meropenem resistance could be reported for substantially more isolates, in a conservative deduplicated analysis even for 43% and 87%, respectively. The remaining isolates were left blank rather than forced into one of the two categories. The blank is not an absence of information: resistance rates within it lay between those of the two reported groups, and its isolates clustered around the clinical breakpoint. It is the genomic counterpart of a population the phenotype already recognizes and reports as susceptible, increased exposure or as an area of technical uncertainty. For meropenem, the residual risk of a susceptibility finding already meets the accepted threshold of 3%, while for ciprofloxacin it approaches this threshold. For tobramycin, validation in a larger cohort, possibly specifically enriched for resistant isolates, is required before our results can be considered reliable. For ceftazidime, by contrast, it appears that additional markers need to be identified, in particular those associated with increased AmpC expression. Taken together, the present data show that, for two of the four antibiotics investigated here, clinically relevant findings can already be generated today: the genome allows the identification of an isolate group with no indication of resistance, marks a smaller group as resistant, and assigns the remaining isolates to an intermediate range close to the clinical breakpoint.