Core Antibiotic-Induced Transcriptional Signatures Reflect Susceptibility to All Members of an Antibiotic Class

ABSTRACT

Current growth-based antibiotic susceptibility testing (AST) is too slow to guide early therapy. We previously developed a diagnostic approach that quantifies antibiotic-induced transcriptional signatures to distinguish susceptible from resistant isolates, providing phenotypic AST 24 to 36 h faster than current methods. Here, we show that 10 transcripts optimized for AST of one fluoroquinolone, aminoglycoside, or beta-lactam reflect susceptibility when the organism is exposed to other members of that class. This finding will streamline development and implementation of this strategy, facilitating efficient antibiotic deployment.

INTRODUCTION

Worldwide, more than 700,000 people die annually from antibiotic-resistant infections (1), and gaps in global antibiotic resistance tracking suggest that this burden is severely underestimated (2, 3). Antibiotic-resistant infections lead to higher medical costs, longer hospital stays, and increased mortality (4, 5). Current growth-based antibiotic susceptibility testing (AST) is too slow to guide therapy in real time (6–8). This diagnostic delay causes overreliance on empiric broad-spectrum antimicrobials, contributing to the emergence of resistance and poor patient outcomes (9, 10).

Antibiotic-induced transcriptional signatures predict susceptibility.

We recently developed a novel microbial diagnostic assay called GoPhAST-R (combined genotypic and phenotypic AST through RNA detection) that can provide AST in <4 h directly from a positive blood culture bottle, 24 to 36 h faster than standard growth-based methods (11). GoPhAST-R measures specific antibiotic-responsive mRNA expression signatures using a commercially available hybridization-based RNA detection platform, NanoString, which is quantitative and multiplexed and works directly from crude lysate with minimal hands-on time (11, 12). After brief antibiotic exposure at the clinical breakpoint concentration, susceptible cells become stressed and exhibit rapid transcriptional changes that distinguish them from unharmed resistant cells. By targeting markers of susceptibility, GoPhAST-R allows accurate and robust phenotypic AST classification, agnostic to resistance mechanism. In addition to detecting these antibiotic-induced transcripts for phenotypic AST, GoPhAST-R can simultaneously target known genetic resistance markers to improve accuracy of resistance detection (13) and facilitate molecular epidemiology by tracking the emergence and spread of specific resistance mechanisms (14), with no need for additional testing. This simultaneous integration of genotype and phenotype sets GoPhAST-R apart from purely phenotypic rapid systems, either commercially available (15) or under development (16), that provide growth-based AST within <7 h from positive blood culture bottles.

We previously demonstrated GoPhAST-R for three antibiotics representing different classes, across five pathogens. We first performed transcriptome sequencing (RNA-Seq) to identify candidate genes whose antibiotic-induced expression best distinguishes susceptible from resistant isolates and then refined these candidate transcripts on the NanoString assay platform, using machine-learning algorithms to select the top 10 antibiotic-responsive genes, and validated their AST prediction accuracy for each pathogen-antibiotic pair (11). We further validated the simultaneous detection of selected genotypic carbapenem resistance determinants (11).

To enhance flexibility of development and implementation of this diagnostic approach, we aimed here to test the generalizability of these transcriptional signatures of susceptibility across an antibiotic class. Since antibiotics elicit transcriptional responses related to their mechanism of action (17–20), we hypothesized that antibiotic-responsive genes optimized for one antibiotic could reflect susceptibility when an isolate is exposed to other members of the same drug class. Here, we confirm this hypothesis for two common pathogens with a propensity for multidrug resistance—Escherichia coli and Klebsiella pneumoniae—treated with multiple members of three major antibiotic classes in regular clinical use: fluoroquinolones, aminoglycosides, and beta-lactams.

Experimental design.

We used NanoString probe sets targeting the top 10 genes we previously identified (11) for each of three individual antibiotics—ciprofloxacin (a fluoroquinolone [FQ]), gentamicin (an aminoglycoside [AG]), and meropenem (a beta-lactam [BL] in the carbapenem subclass)—and assessed whether they reflect susceptibility when exposed to other members of their respective class (Table 1). All strains were obtained from clinical or reference microbiological laboratories, representing diverse geographic locations and resistance mechanisms when possible, and all MICs were verified by broth microdilution (6) (see Data Set S1 in the supplemental material). For the FQ and AG classes, strains were grown in Mueller-Hinton broth (MHB) and treated at early log phase (optical density at 600 nm [OD₆₀₀], ∼0.2). For BLs, we previously found that strains with prominent inoculum effects (21–23) may exhibit decreased induction of transcriptional susceptibility signatures at high cell density (11). To account for this effect for the BL class, we treated isolates at the Clinical and Laboratory Standards Institute (CLSI)-recommended MIC inoculum range of 2 × 10⁵ to 8 × 10⁵ CFU ml⁻¹ by growing them in MHB to early log phase, back-diluting to 2 × 10⁵ CFU ml⁻¹ in fresh MHB, and incubating for 60 min prior to treatment.

TABLE 1.

Classification of antibiotics used in this study with respective treatment conditions

Antibiotic	Class	Subclass	Antibiotic treatment
			Concna (mg/liter)	Duration (min)
Ciprofloxacin	FQ	2nd-generation FQ	0.25	60
Levofloxacin	FQ	3rd-generation FQ	0.5	60
Moxifloxacin	FQ	3rd-generation FQ	0.25	60
Gentamicin	AG	4	60
Tobramycin	AG	4	60
Amikacin	AG	16	60
Ampicillin	BL	Aminopenicillin	8	120
Cefazolin	BL	1st-generation cephalosporin	2	120
Ceftriaxone	BL	3rd-generation cephalosporin	1	120
Aztreonam	BL	Monobactam	4	120
Piperacillin-tazobactam	BL	Ureidopenicillin/beta-lactamase inhibitor	16/4	120
Ertapenem	BL	Carbapenem	0.5	120
Meropenem	BL	Carbapenem	1	120

^aThe concentration for antibiotic treatment was chosen as the clinical susceptibility breakpoint established by the CLSI for all antibiotics whenever available (24) or as the European Committee on Antimicrobial Susceptibility Testing (EUCAST) breakpoint for moxifloxacin (29).

After brief antibiotic exposure (60 min for FQs and AGs, 120 min for BLs) at their respective clinical susceptibility breakpoint concentrations (6, 11, 24), or an equivalent control incubation without antibiotics, samples were mechanically lysed and used as input for NanoString assays as previously described (11). Using the NanoString platform (12), we quantified expression of the top 10 responsive and 10 control genes we previously identified for individual pathogen-antibiotic pairs in multiplexed fashion (11). Control genes, whose expression is unaffected by antibiotics, were used to scale for cell number at lysis, resulting in a normalized fold induction of each responsive gene in antibiotic-treated versus untreated samples for each strain, calculated as previously described (11). To assess assay performance, these normalized fold inductions were compared with the known AST classification of each isolate. For all analyses, we used the code we previously wrote and executed, as described elsewhere (11). Code and NanoString raw data files can be made available upon request.

To further characterize the isolates used in this study, we also sequenced those not previously deposited in the NCBI Sequence Read Archive (SRA). Genomic DNA was extracted using a Qiagen DNeasy blood and tissue kit, prepared using an Illumina Nextera XT DNA library prep kit, sequenced on an Illumina MiSeq (or NextSeq, for strain 16) platform, and analyzed for known resistance determinants using ResFinder (25) (Data Set S1).

Generalizability of 10-transcript signatures across FQs and AGs.

We tested the ciprofloxacin 10-transcript signatures across the FQs ciprofloxacin, levofloxacin, and moxifloxacin and the gentamicin 10-transcript signatures across the AGs gentamicin, tobramycin, and amikacin. For each class, we selected six isolates of each species: three susceptible and three resistant to all class members (Data Set S1). For both species, heat maps illustrate that the top 10 genes identified for AST of ciprofloxacin and gentamicin showed similar normalized fold induction upon treatment with three FQs and three AGs, respectively (Fig. 1a and b). One-dimensional projections summarizing these transcriptional data (11, 26) show robust distinction of susceptible and resistant isolates across each class (Fig. S1a and b).

FIG 1.

Differential expression of the same antibiotic-induced 10-transcript signatures reflect susceptibility for all drugs within a class. The figure shows heat maps of normalized, log-transformed fold induction (treated/untreated) of the top 10 antibiotic-responsive transcripts we previously identified for E. coli (top panels) and K. pneumoniae (bottom panels) treated with (a) ciprofloxacin, (b) gentamicin, and (c) meropenem upon exposure at CLSI breakpoint concentrations to other fluoroquinolones, aminoglycosides, and beta-lactams, respectively. Color scales indicate range of log₂(fold induction) for transcripts in the respective heat map(s), symmetrically scaled to each plot. Strains are indicated by numbers over the heat map columns (Data Set S1), with CLSI classifications of each strain based on broth microdilution shown below (S, susceptible; I, intermediate; R, resistant). Gene identifiers for antibiotic-responsive transcripts are listed on the left, as defined for NCBI reference sequences NC_000913 (E. coli; top panels) and NC_009648 (K. pneumoniae; bottom panels).

Generalizability of a 10-transcript signature across BLs.

We next tested BLs, a large class of diverse compounds comprising multiple subclasses that span a wide spectrum of antibacterial activity. The diversity of the BLs challenges the generalizability of transcriptional signatures of susceptibility within a drug class and offers the most clinical benefit from rapid and efficient antibiotic deployment if successful. We tested the 10-transcript signatures identified for meropenem across treatments of seven BL antibiotics spanning multiple subclasses and ranging in spectrum of activity: ampicillin, cefazolin, ceftriaxone, aztreonam, piperacillin-tazobactam, ertapenem, and meropenem. We selected eight clinical isolates for each species that vary in susceptibility across the different BLs, from pan-susceptible to pan-resistant (Data Set S1). Note that although K. pneumoniae is intrinsically resistant to ampicillin, strain 23 has an MIC that falls in the susceptible range; whole-genome sequencing (WGS) revealed a premature stop codon (E88*) in the chromosomal bla_SHV-89 that may explain this unusual phenotype. Normalized fold induction of the meropenem-responsive genes tested across the seven BLs is shown as heat maps (Fig. 1c) and summarized as one-dimensional projections (Fig. S1c). In both species, each BL induced the top 10 meropenem-responsive transcripts in only the susceptible isolates, allowing susceptibility distinction despite variability in expression levels of certain genes across the class. Strains with MICs closer to the breakpoint exhibited partial induction, consistent with our previous finding that the magnitude of transcriptional response to antibiotic exposure at the breakpoint correlates with MIC (11). Specific carbapenemases from our panel of common carbapenemases and extended-spectrum beta-lactamases were found in the same assay, explaining resistance phenotypes in most carbapenem-resistant isolates (Fig. S2). Notably, strain 14 is correctly distinguishable from susceptible isolates upon ertapenem exposure, despite lacking genotypic markers from our panel. WGS of this isolate revealed three beta-lactamases, including bla_CMY-2 (27, 28), as well as polymorphisms in the OmpC and OmpF porins (Data Set S1), suggesting a possible atypical basis for ertapenem resistance that would have been difficult to predict by genotype alone. This exemplifies GoPhAST-R’s ability to determine phenotypic resistance to each antibiotic, independent of resistance mechanism.

GoPhAST-R as a robust diagnostic platform.

This work demonstrates that the same antibiotic-induced 10-transcript signatures reflect antibiotic susceptibility for all drugs within a class, consistent with a conserved core transcriptional response to each of three major antibiotic classes. Despite their diversity, even BL compounds share a susceptibility signature across all subclasses, implying a common stress response to their similar cellular targets. The 10-transcript signatures used in this study, previously designed for individual antibiotics, may not represent the top 10 genes for the whole class or for each individual member, but they clearly report on susceptibility for each compound in a given class. Moreover, WGS of isolates used in this study revealed diverse genotypic resistance mechanisms (target site mutations, efflux pumps, porin mutations, or antibiotic-modifying enzymes [see Data Set S1]), further illustrating how our phenotypic AST approach is fundamentally agnostic to resistance mechanism. Thus, this study provides further evidence for transcriptional profiling as a robust phenotypic measure of antimicrobial stress, underscoring the flexibility and breadth of GoPhAST-R: the same minimal gene set derived for one specific antibiotic can assess susceptibility not only across diverse strains but also across drugs with a shared mechanism of action. This finding will streamline GoPhAST-R implementation, contributing to the critical effort to employ rapid AST diagnostics to guide upfront selection of the narrowest effective agent against a given pathogen. Efficient, informed deployment of antibiotics will improve patient outcomes while minimizing selection for resistance.

Data availability.

Raw sequencing files of strains not previously deposited in the NCBI Sequence Read Archive (SRA) were deposited in the SRA under project no. PRJNA707347.