Can the use of older-generation beta-lactam antibiotics in livestock production over-select for beta-lactamases of greatest consequence for human medicine? An in vitro experimental model

Though carbapenems are not licensed for use in food animals in the U.S., carbapenem resistance among Enterobacteriaceae has been identified in farm animals and their environments. The objective of our study was to determine the extent to which older-generation β-lactam antibiotics approved for use in food animals in the U.S. might differentially select for resistance to antibiotics of critical importance to human health, such as carbapenems. Escherichia coli (E. coli) strains from humans, food animals, or the environment bearing a single β-lactamase gene (n = 20 each) for blaTEM-1, blaCMY-2, and blaCTX-M-* or else blaKPC/IMP/NDM (due to limited availability, often in combination with other bla genes), were identified, along with 20 E. coli strains lacking any known beta-lactamase genes. Baseline estimates of intrinsic bacterial fitness were derived from the population growth curves. Effects of ampicillin (32 μg/mL), ceftriaxone (4 μg/mL) and meropenem (4 μg/mL) on each strain and resistance-group also were assessed. Further, in vitro batch cultures were prepared by mixing equal concentrations of 10 representative E. coli strains (two from each resistance gene group), and each mixture was incubated at 37°C for 24 hours in non-antibiotic cation-adjusted Mueller-Hinton II (CAMH-2) broth, ampicillin + CAMH-2 broth (at 2, 4, 8, 16, and 32 μg/mL) and ceftiofur + CAMH-2 broth (at 0.5, 1, 2, 4, and 8μg/mL). Relative and absolute abundance of resistance-groups were estimated phenotypically. Line plots of the raw data were generated, and non-linear Gompertz models and multilevel mixed-effect linear regression models were fitted to the data. The observed strain growth rate distributions were significantly different across the groups. AmpC strains (i.e., blaCMY-2) had distinctly less robust (p < 0.05) growth in ceftriaxone (4 μg/mL) compared to extended-spectrum beta-lactamase (ESBL) producers harboring blaCTX-M-*variants. With increasing beta-lactam antibiotic concentrations, relative proportions of ESBLs and CREs were over-represented in the mixed bacterial communities; importantly, this was more pronounced with ceftiofur than with ampicillin. These results indicate that aminopenicillins and extended-spectrum cephalosporins would be expected to propagate carbapenemase-producing Enterobacteriaceae in food animals if and when Enterobacteriaceae from human health care settings enter the food animal environment.

Introduction respiratory tract, urinary tract), another well-documented but unintended consequence is the direct selection of resistant organisms in the host gastrointestinal tract [14]. Consequently, Enterobacteriaceae, and particularly E. coli, are commonly used indicators of bacterial resistance prevalence reflecting the immediate as well as the cumulative effects of antibiotic selection pressures over the decades [15]. Furthermore, most of these commensal species can cause opportunistic infections when host defense mechanisms are compromised; for instance, Enterobacter cloacae and Klebsiella spp. are major causes of healthcare-associated infections (HAIalso known as nosocomial infections) [16]. The public health importance of Enterobacteriaceae and related gram-negative bacteria (e.g., pseudomonads) is further magnified by their ability to readily exchange mobile genetic elements (MGEs)-e.g., plasmids and transposonsthat carry various antimicrobial resistance genes, thus spreading antimicrobial resistance from intestinal commensals to more highly pathogenic strains [17].
Antimicrobial use in animal agriculture has been reported to contribute to increased AMR prevalence, both in animals and in humans [32]. Some experts have suggested that uses in livestock production account for about two-thirds of the global antibiotic sales and consumption [33]. Although the risk posed to humans by direct acquisition of AMR organisms from food animals is difficult to quantify, ample evidence exists of human acquisition of resistant bacteria via food animal sources [34][35][36][37][38][39].
Carbapenems are the 'last line of defense' beta-lactams; as such, they have never been approved for use in food animals [40]. Nevertheless, carbapenemase-producing Enterobacteriaceae (CPE) have been isolated globally in a variety of food animal systems. The enzyme in Acinetobacter baumannii carrying bla OXA-32 was the first carbapenemase-producing bacterium identified in food animals, and was isolated in dairy cattle in France [28]. In 2011, a year after the finding in France, the German Research Network (RESET) project found E. coli and multiple Salmonella isolates harboring the bla VIM-1 gene in pig feces, the pig farm environment, and in poultry dust [41][42][43]. Other carbapenemase-producing bacteria have been found in food animals in places such as China (A. baumannii: bla NDM-1 ) [44,45] and Lebanon (Pseudomonas aeruginosa: bla VIM-2 , A. baumannii: bla OXA-23 ) [46]. In the High Plains region of Texas and New Mexico, the first U.S. animal agricultural CPE were identified in 2016; of note, one A. baumannii from dairy cattle excreta carried a seemingly novel chromosomal gene bla OXA-497 [47]. Later, plasmid-borne bla IMP-27/64 genes were detected in multiple Enterobacteriaceae species from the environment and fecal samples of a farrow-to-finish swine operation in the Midwest region of the U.S. This latter finding included the observation that the CPE isolates were found in pig barns where a 3GC ceftiofur was used, but not in barns where its use was not recorded [48,49].
The increasing isolation of CPEs from food animal agricultural settings poses somewhat of a conundrum. Mollenkopf et al. (2016) proposed that approved beta-lactams in animal agriculture may differentially select for carbapenemase-producing strains in farm animals and their environment [48,50]. Other authors such as Woodford et al., (2014) made a similar but broader suggestion that multiple classes of antimicrobials other than the beta-lactams may provide the selection pressures for carbapenem-resistance Enterobacteriaceae (CRE) isolates in animal agriculture [51]. Multiple genes encoding resistance to different antimicrobial agents are frequently found together on mobile genetic elements-such as plasmids harbored by Enterobacteriaceae-and this can facilitate co-selection of resistance. However, no evidence exists to date to support assertions that early-generation beta-lactams might select disproportionately for higher-level beta-lactam resistance, including CPEs, in the absence of that coselection mechanism.
The objective of this study was to determine the extent to which older-generation β-lactam antibiotics approved in the U.S. for use in food animals (e.g., aminopenicillins and 1st generation cephalosporins) can differentially select for highest priority antibiotic resistance (e.g., to 3rd and 4th generation cephalosporins and carbapenems, respectively) among representative Enterobacteriaceae. In addition to characterizing individual strain fitness in both antibiotic containing and non-antibiotic media, we also aimed to test the differential selection dynamics of a 5-group mixture of host-adapted E. coli strains; as far as possible, with each group bearing one, but not a combination, of beta-lactamase genes: bla TEM-1 , bla CMY-2 , bla CTX-M-� , or bla KPC/ IMP/NDM , along with a group of beta-lactamase-free strains. Through in vitro experiments and statistical modeling, we aimed to narrow the existing knowledge gap concerning the potential for over-selection of ESBLs and CPEs in U.S. food animal production through approved uses of ampicillin and ceftiofur, respectively.

Microbiologic methods
Bacterial strain selection. Bacteria used in this study were sourced from our strain collections (HMS and TEW) as well as from the jointly sponsored U.S. Centers for Disease Control and Prevention (CDC) and the Food and Drug Administration (FDA) Antimicrobial Resistance (AR) Isolate-Bank [52]. Isolates from our own collections were characterized as to antimicrobial resistance, both phenotypically and genotypically, in previous research projects in our (HMS and TEW) groups. Detailed methodologies employed in those characterizations have previously been published [48,49,53]. For carbapenem-producing Enterobacteriaceae (CPE), initial isolation of the bacterial strains from field samples was through supplemented MacConkey broth containing meropenem at 0.5μg/mL and zinc sulfate heptahydrate 70 μg/ mL [48].
Briefly, minimum inhibitory concentrations (MIC) of antimicrobial susceptibility which are routinely monitored in Enterobacteriaceae were determined for the isolates using the broth micro-dilution method. The Sensititre TM system (TREK, Thermo Scientific Microbiology, Oakwood Village, OH) was employed, using gram-negative CMV3AGNF custom panels designed for the U.S. National Antimicrobial Resistance Monitoring System (NARMS); for some isolates, extended-spectrum beta-lactamase ESB1F panels (TREK, Thermo Scientific Microbiology, Oakwood Village, OH) also were used to further characterize the beta-lactam phenotypic susceptibility (i.e., AmpC versus ESBL versus carbapenemase). Bacterial antimicrobial testing was performed in accordance with published NARMS protocols [54]. Outcomes expressed as antimicrobial minimum inhibitory concentrations (MIC) were interpreted according to the clinical interpretative human breakpoint values recommended for E. coli by the Clinical Laboratory Standards Institute (CLSI) as listed below, or else as their raw values befitting the type of the statistical analysis performed [55].
Bacterial genotypes were determined in our laboratories through short-read whole genome sequencing. In brief, bacterial genomic DNA was extracted with the QIAamp DNA extraction kit on the QIAcube HT automated platform (QIAGEN, Valencia, CA) while library preparation was with the Illumina Nextera XT or DNA Flex Kits (Illumina Inc, San Diego, CA). Sequencing runs were performed with the MiSeq Reagent Kit v3 paired-end reads (2 x 300 bp) on the Illumina MiSeq instrument (Illumina Inc., San Diego, California). Post-run bioinformatic analyses were performed on the BaseSpace Sequence Hub (Illumina Inc., San Diego, California). Depending on the source of the isolate, nucleotide sequence reads were assembled with either Velvet de novo or SPAdes genome assembler software [56,57]; bacteria sequence types were determined with the SRST2 Basespace application (Illumina Inc., San Diego, California). Antimicrobial resistant gene annotation was with the Antibiotic Resistance Gene-ANNOTation database (ARG-Annot) or ResFinder [58] and plasmid type annotation was via PlasmidFinder [59]. Genotypic and phenotypic characterizations of isolates obtained from the CDC/FDA AR Isolate Bank were as published on the agency website and accessible from the National Center for Biotechnology Information (NCBI) portal [52]. For the purpose of uniformity in data presentation, and to further ascertain prior strain annotations from the aforementioned sources, raw reads of strains used in this study were again, either pulled from NCBI (National Center for Biotechnology Information, U.S. National Library of Medicine) SRA (sequence read archive) [60] or from our archive. Nucleotide sequence reads were assembled using SPAdes ver.3.11.1 genome assembler softwares [57]; bacteria sequence types were determined with the SRST2ver.0.2.0 [61] and Multi-Locus Sequence Typing (MLST) database of Escherichia coli (accessed on May, 2020) [62]. Antimicrobial resistant gene and plasmid annotation was performed using ABRicate ver.0.8.7 [63] and ResFinder [58] and PlasmidFinder [64], databases (both accessed on May, 2020). Corresponding sequence type, genotypic and phenotypic antimicrobial resistance profile, and plasmidal information of the isolates included in this study are provided under the given bioproject and biosample accession numbers in supplemental material (S1 Table).
When possible, the choice of E. coli strains of swine origin was prioritized over other bacteria sources, swine being our exemplar food animal host. Pig being monogastric possess numerous intestinal anatomical, physiological and microbiota similarities with the human digestive tract; these similarities have informed the many use of swine in modeling human intestinal parameters [65,66]; additionally, the first description of CPEs in the United States was in a swine production facility [48]. When this was not possible, strains from other livestock, human patients, or else the environment were selected. E. coli strains were initially selected according to the presence of plasmid-borne genes encoding beta-lactamases, or else a complete lack thereof. Isolates identified from this screening process were then stratified by the presence of their beta-lactamase genes; that is, bla TEM-1 or bla CMY-2 or bla CTX-M-� , or bla KPC/IMP/NDM , or none, but avoiding to the extent possible those strains with a combination of bla genes. Each group of the selected resistance genes corresponded to a known spectrum of hydrolytic activity against beta-lactam antibiotics.
Twenty isolates were selected per bla-positive group or bla-negative control group. Ideally, 30 strains of each group would have been the sample size to achieve near-normal log 10 transformed distribution of target quantitative outcomes under the Central Limit Theorem [67,68]; however, strains with only a single bla gene were limited in the sources accessible to us. Even more so for the CPEs; consequently, presence of a single bla gene was achieved only for ten percent of the carbapenemase producing strains (see S1 Table). Study strains were allowed to possess genes encoding resistance to other classes of antimicrobials. These genes were not expected to impact the beta-lactam antimicrobial resistance of the strains in the absence of coselection pressures; for example, genes conferring tetracycline, aminoglycoside and sulfonamide resistance were commonly identified among all strain groups.
Bacterial growth curve estimation. To assess the growth rates of each bacterial strain absent and under different antibiotics and their concentrations, and to estimate the withinand among-bla-group growth fitness parameter differences, bacterial growth curves were estimated with the Bioscreen C TM Automated Microbiology Growth Curve Analysis System (Growth Curve Ltd, Helsinki, Finland). Bacteria from pure culture and preserved on cryobeads at -80˚C were streaked onto Trypticase™ soy agar with 5% sheep blood (Becton, Dickinson and Company, Sparks, MD) and incubated at 37˚C overnight. A 0.5 McFarland standard (Sensititre TM Thermo Fisher Scientific, Waltham, MA) bacterial suspension was made for each isolate (i.e., to a bacterial concentration of~1.5 x 10 8 CFU/mL), by suspending one or two colonies selected from the overnight plate growth in demineralized water (Thermo Fisher Scientific, Waltham, MA). A 1:10 mixture of the bacterial suspension (120 ul) in cation-adjusted Mueller-Hinton II (CAMH-2) broth (Thermo Fisher Scientific, Waltham, MA) (1,080 ul) in a 1.5 mL black sample tube was then made (final bacterial concentration of~1.5 x 10 7 CFU/mL). From this bulk mixture, 300ul aliquots in triplicate were dispensed into each 10 � 10 honeycomb plate well (Growth Curves USA, Piscataway, New Jersey, USA) for each strain; thereafter, automated optical density (OD) estimates were obtained at 420-580nm (wideband) over 48 hours at 37˚C. Measurements (OD) were taken every 10 minutes following moderate agitation of the incubating cultures. For each experiment, a single QC strain E. coli ATCC 25922 (American Type Culture Collection, Manassas, VA) in triplicate wells, and two wells with plain CAMH-2 broth, were included as positive and negative controls, respectively.
The effects of ampicillin (an aminopenicillin), ceftriaxone (a third-generation cephalosporin (3GC)), and meropenem (a carbapenem) on the growth parameters of each strain in the corresponding bla-groups were estimated. The tested concentration of each antimicrobial corresponded to the human clinical resistance breakpoint value for the drug-bacteria species MIC as recommended by the CLSI [55]. Specifically, the concentrations were 32 μg/mL of ampicillin sodium, 4 μg/mL of ceftriaxone disodium, or 4μg/mL of meropenem trihydrate; high purity forms of the antimicrobials (Sigma-Aldrich Inc., St. Louis, MO) dissolved in CAMH-2 broth were used to prepare the tested concentrations. Ceftriaxone (a commonly used 3GC in human medicine) was chosen for this in vitro assay to represent potential selection risks in the human host.
Bacterial competition assay. From each resistance group (i.e., no bla genes,-but with a tetracycline resistance gene, tet(B)-bla TEM-1 , bla CMY-2 , bla CTX-M-� , or bla KPC/IMP/NDM ), two representative strains were selected for competition studies. From the WGS data, isolates with the least number of co-resistance genes were selected as representative strains per strain-group to control for potential fitness constraint imposition, while equally adjusting for antimicrobial co-selection potential. A 0.5 McFarland (Sensititre TM Thermo Fisher Scientific, Waltham, MA) standardized bacterial suspension was prepared for each selected isolate as described above. An equal mixture of all ten strains (1 mL of each suspension) was prepared (expected density of each strain in the mixture was~1.5 x 10 7 CFU/mL). A 1:10 dilution of the mixture in CAMH-2 broth, as well as in CAMH-2 broth with various antibiotic concentrations (expected density of each strain~1.5 x 10 6 CFU/mL) were made. The antimicrobial concentrations were ampicillin sodium (Sigma-Aldrich, St. Louis, MO) at 2, 4, 8, 16 or 32 μg/mL and ceftiofur hydrochloride (Zoetis Inc., Parsippany-Troy Hills, NJ) at 0.5, 1, 2, 4, 8 μg/mL. Ceftiofur was chosen for this assay to represent the 3GC used in veterinary medicine; that is, to directly mimic the selection pressure occurring in vivo in farm animals.
Triplicates of the mixed cultures were incubated at 37˚C for 24 hours in the Bioscreen C TM Automated Microbiology Growth Curve Analysis System, as previously outlined. The postgrowth density of each strain group was estimated phenotypically by the spiral-plate method (Eddy Jet 2™ Spiral-Plater, Neutec Group Inc., NY); to achieve this, after 24 hours of incubation the mixed cultures were spiral-plated onto MacConkey agar plates infused with one of tetracycline-16 μg/mL, ampicillin-32 μg/mL, cefoxitin-32 μg/mL, ceftriaxone-4 μg/mL, ceftriaxone-4 μg/mL with clavulanate-4 μg/mL, cefepime-8 μg/mL, or meropenem-4 μg/mL [55]. The bacterial culture dilutions for spiral-plating were performed as necessary on ice beds (to halt bacterial growth). The agar plates were read on an automated colony counter (Flash & Go TM , Neutec Group Inc., NY) to obtain estimates of the bacterial density in colony forming units (CFU)/mL of the mixed culture as of 24 hours of incubation. The strain group(s) whose density was assessed using each of the selective agars is listed in Table 1. A fresh preparation of experimental bacteria mixture was also plated as outlined above.

Statistical methods
Growth curve analyses. Data from the EZexperiment™ software (Growth Curves Ltd, Helsinki, Finland) were transformed from wide to long format based on the time of incubation (i.e., in 10-minute intervals from time = 0) in Stata version 15.1 software (Stata Corp, College Station, TX). Triplicate OD measurements were collapsed to their mean by strain at each time point to obtain a single representative value. Raw and mean measured OD of each bacterial strain was plotted against time. The isolates were stratified by resistance group and the maximal bacterial population growth rates were estimated, as well as the growth rates in presence of each of the antibiotic concentrations.

PLOS ONE
Can older-generation beta-lactams over-select for newer beta-lactamases of consequence for human medicine?
To more precisely estimate the growth parameters, four nonlinear regression models were fitted to the OD-based growth curve of the bacterial groups grown in non-selective media, using the least squares method implemented in the Stata 1 15.1 software (Stata Corp., College Station, TX). The four models were: 3-parameter Gompertz, 4-parameter Gompertz [69], 3-parameter logistic [70] and 3-parameter exponential [71]. Further, the relative fit of the models across all the bacterial resistance groups and experimental conditions (i.e., with and without antibiotics) was explored. The 3-parameter Gompertz (Eq 1) was found to most consistently provide the best fit (highest adjusted coefficient of determination (R 2 )) for the growth curves, across all antibiotic concentration/bacteria-group combinations. Consequently, the bacterial growth rate (OD/hour) estimates, after transforming time units from minutes to hours, were extracted from the 3-parameter Gompertz models fitted to these growth curves: Where, t is time in hours, β 1 is the peak bacterial density (OD), β 2 is the estimated maximum growth rate (Δ OD per hour) and β 3 is the estimated bacterial growth lag period (in hours). Models were stratified by resistance gene group and the type of growth media (i.e., without and with different antibiotics at breakpoint concentrations). To obtain graphical representations of modeled growth rates, post-analysis non-linear OD marginal predictions were generated. Predicted OD were subsequently graphed against time in hours.
Strain group growth in the mixed cultures. Mixed-effect nonlinear regression models were fitted to the estimated density of each resistance group at 24 hours of incubation of the mixedstrain culture (the CFU/mL readings from the selective antibiotic MAC plates were log 10 transformed to normalize the data prior to the analysis), Modeled fixed effects were the antibiotic type, the various concentrations of each antibiotic used in the growth media for the mixed culture, and the selective antimicrobial plate type. The experimental replicate was modeled as a random effect factor. Marginal mean predictions of the strain group densities from the model output with 95% confidence intervals were determined and represented graphically. The analysis also was performed in Stata 1 version 15.1 software (Stata Corp., College Station, TX).

Strain group fitness
In all strain groups, numerous strains demonstrated biphasic exponential growth phase-suggesting a switch in energy source utilization from glucose to a different compound [72][73][74] when nutrient supplies became limited-while others did not (Fig 1). All groups exhibited a lag period prior to the exponential amplification of the OD value. Within-group similarity in the first phase of exponential growth was observed for strains with and without a diauxic growth pattern. The onset of the second phase of exponential growth-when present-exhibited withinand among-group variation, as seen in Fig 1 with selected representative strains. The maximum OD value attained by any strain across all resistance gene groups was roughly the same (�1.4), but the time taken to reach this point differed within and among groups.
The 4-parameter Gompertz and the 3-parameter exponential models showed the lowest R 2 values when fitted to the data (0.34 to 0.84 and 0.34 to 0.85 across the strains, respectively). The 3-parameter logistic (0.85-0.93) and 3-parameter Gompertz (0.86-0.93) models demonstrated the highest R 2 values. Thus, the 3-parameter Gompertz model was fitted to all the growth curves, and the model predictions were plotted and evaluated. When modeled in the absence of beta-lactam antibiotics, the estimated peak growth rate of the beta-lactam susceptible group [0.159 (OD/hour), 95% CI: 0.152-0.166] was significantly higher than for the bla-positive groups (see Table 2 and Fig 2). Among the bla producers, the estimated growth rate of the carbapenemase producing E. coli in non-antibiotic media was highest [0.142 (OD/hour), 95% CI: 0.136-0.149]; however, this estimate was not statistically significantly different from that of the TEM-type beta-lactamase group [0.139 (OD/hour), 95% CI: 0.135-0.143], as indicated by the overlap of their respective 95% CIs. The AmpC-type beta-lactamase group and the ESBL producers exhibited statistically similar peak growth rate estimates [0.120 (OD/ hour), 95% CI: 0.116-0.124] and [0.127 (OD/hour), 95% CI: 0.121-0.132]. Although the betalactam susceptible group showed a higher growth rate, among the beta-lactam resistant groups no fitness cost pattern along the gradient of the encoded resistance (from non-extended-toextended or from extended-to-non-extended beta-lactam resistance) was observed.
In a similar fashion, the combined growth rates of the representative pair of strains of each group used in the mixed cultures were predicted (Table 3). Select representative raw data line plots of OD-based growth curves from cation-adjusted Mueller-Hinton broth for 10 study strains (color-coded by the 5 resistance gene groups) are shown in Fig 1. These estimates show a 95% CI overlap among multiple group-based strain pairs; the pair of beta-lactam susceptible strains demonstrated the highest predicted growth rates [0.16 (OD/hour), 95% CI: 0.15-0.17], an estimate non-statistically different from that of the carbapenemase producers [0.13 (OD/ hour), 95% CI: 0.12-0.15]. The predicted growth rates of the ESBL-producing pair [0.12 (OD/ hour), 95% CI: 0.12-0.12] were observed to be non-statistically different from that of the carbapenemase producers but statistically different from the beta-lactam susceptible strains. The

Effect of beta-lactam antibiotics on bacterial growth rates
The relative growth rates of the resistance-gene groups in the presence of beta-lactams of different generations (i.e., at concentrations corresponding to the human clinical interpretive breakpoints for MICs of these drugs for E. coli) were assessed, to predict preferential selection by the antibiotics for the strains with studied resistance gene groups. As expected, the beta-lactamase-free strains registered no discernable growth in media with any of the three tested beta-lactam antibiotics. All strains from the four beta-lactam resistance gene groups showed no significant growth impairment in 32 μg/mL ampicillin (though with substantive 95% CI overlap of ampicillin-containing versus plain CAMH-2 broth growth rates, respectively:   Table 2 and Fig 3).

PLOS ONE
Can older-generation beta-lactams over-select for newer beta-lactamases of consequence for human medicine?

Effect of beta-lactam antibiotics on mixed bacterial populations
Initial resistance-gene group populations grown on the selective antimicrobial MAC plates (Table 1) using the starting mixture of select representative strains (see Table 1 (Table 4). Assuming a simple substitution model with no compensatory growth through

PLOS ONE
Can older-generation beta-lactams over-select for newer beta-lactamases of consequence for human medicine?
The constituent proportions of beta-lactamase producers in the ampicillin containing cultures, as estimated by the CFU counts on the MAC+AMP agar plates relative to the non-selective plate, were 29%, 34.7%, 59.1%, and 100% for 2, 4, 8, and 16 μg/mL, respectively, in ascending order of ampicillin concentration. In contrast to this trend, the beta-lactamase producers were estimated to grow only to 90% of the total expected CFU in the 32 μg/mL ampicillin broth culture, a substantial drop from the preceding concentration. Nonetheless, the absolute count of the beta-lactamase producers increased steadily across ampicillin concentrations from 9.07 (log 10 CFU/mL) [95% CI: 8.97-9.17] in the 2 μg/mL ampicillin broth culture to 9.28 (log 10 CFU/mL) [95% CI: 9.16-9.39] in the 32 μg/mL ampicillin broth culture. Although this result exhibits a seeming paradoxical increase in relative proportions of non-beta-lactamase producers in 32 μg/mL ampicillin broth culture, the overall trend observed was a steep decline in the susceptible bacteria sub-population and a steady increase in the component beta-lactamase producing strains' population as in vitro ampicillin concentrations increased. The estimated proportion of CPE strains in the mixed culture quadrupled from the 2 μg/mL ampicillin broth to the 16 μg/mL ampicillin broth, increasing from 0.04% of the estimated total density to 0.16% (peak). Bacterial counts on the MAC+AXOCLAV plates also generally increased in proportion along with ampicillin broth concentration: 6.8%, 5.8%, 23.8%, 43.9% and 48.7% (for 2, 4, 8, 16, and 32 μg/mL, respectively). Likewise, bacterial count estimates on the MAC+AXO plates increased in proportion with doubling ampicillin concentrations, peaking at ampicillin concentrations of 16 μg/mL: 27.5%, 30.5%, 51.5%, 96.3% and 86.7%, and in ascending order of ampicillin concentration. The marginal mean estimates from the MAC +AXO plates, MAC+AXOCLAV plates and the MAC+AMP plates suggest the ESBL-producing strains constitute the bulk of beta-lactamase producers following selective pressures of ampicillin in a competitive mixed culture (see Table 4 and Fig 4).
Ceftiofur. In contrast to the starting ampicillin broth concentration (2 μg/mL), the starting ceftiofur broth concentration (0.5 μg/mL) demonstrated a significant suppression of the estimated total bacterial count relative to estimates from the non-selective broth culture [9.17 (log 10 CFU/mL) 95% CI: 9.08-9.26 compared to 9.58 (log 10 CFU/mL) 95% CI: 9.44-9.72], respectively (Table 4). Further significant (P < 0.05) suppressions of the estimated total bacterial CFU in the mixed-strain culture were not observed until the broth concentration doubled from 2 μg/mL to 4 μg/mL of ceftiofur [9.11 (log 10 CFU/mL) 95% CI: 9.02-9.20 compared to 8.87 (log 10 CFU/mL) 95% CI: 8.78-8.96], respectively. The proportion of the maximum capacity (that is, in absence of the antibiotic) achieved by beta-lactamase producers in the ceftiofur broth culture (as estimated by the CFU counts on the MAC+AMP agar plates relative to the non-selective plate) were 115%, 108%, 98%, 87% and 108%, in ascending order of ceftiofur concentration. The proportions above or near 100% suggest dominance of the culture by such strains and are likely artifacts introduced by bacterial counting methods when they exceed 100% of growth on plain media.
The absolute count of the beta-lactamase producers as estimated on these plates also decreased; most notably, they were suppressed in the broths with 4 μg/mL and 8 μg/mL of ceftiofur. The overlap in estimated CFU proportions on the MAC+AMP and MAC+AXO selective agar plates suggests the TEM-1-type beta-lactamase strains constituted a less important component of the community. The estimated proportion of CPE strains isolated via the MAC +MERO selective media plates showed a steady increase from 0.2%, through 0.1%, 0.3%, 0.8%, and 1.9%, along the ceftiofur concentration gradient outlined in Table 1. The CFUs selected by the MAC+AXOCLAV plates also increased in proportion along with ceftiofur concentrations: 56.6%, 52.9%, 60.3%, 78.3% and 108.5% (see Table 4 and Fig 5).

Discussion
Often, and for good reasons, regulatory and policy approaches to the challenge of AMR tend to be reactive rather than proactive. These are exemplified by restrictions on the use of specific antimicrobial drugs, classes, or drug indications for specific diseases, often long after they have been approved and labeled for such use. The current absence of carbapenems and 4GC use in the U.S. in livestock production has offered unique opportunities for assessing the value of proactive policies to addressing the question of the roles of direct and indirect selection of bacterial resistance types to these antimicrobials in the presence and absence of co-selection by other antimicrobials.
Though more consistently seen with chromosomally located genes, the carriage of AMR genes on horizontally transferable genetic materials has also been shown to impose a fitness cost (often estimated using the maximal growth rate as a surrogate fitness measure) on bacterial strains [75,76]. In our study, mammalian-derived commensal E. coli strains with similar beta-lactamase genotypes were found to possess a spectrum of fitness levels as reflected in the growth rates. This underscores the importance of a global genotypic evaluation across all AMR profiles, virulence profiles and other stress adaptation mechanisms, in order to estimate the fitness impact of a particular variable. In this study, unlike in prior studies [77], we characterized the growth of a large number of bacterial strains for each gene group to assess and adjust for this variation among strains of the same species exhibiting similar genotypic and phenotypic characteristics. When analyzed based on the bla-gene classifications (n = 20 per gene group),

PLOS ONE
our study found the bla-negative group to be the most fit in the absence of antibiotic selection pressures, suggesting a fitness advantage may exist in association with the lack of a bla gene. A similar model comparing beta-lactamase producers, surprisingly, showed a similar level of fitness between the high-potency carbapenemase bla gene group and the lower potency bla gene group (bla TEM-1 ). Consequently, our results do not suggest a trend of increasing fitness cost (as estimated by the maximum growth rate) in lock step with increasing resistance severity of the bla gene groups. On the other hand, a relatively impaired maximum bacterial density (as implied by the highest OD as a surrogate measure) was observed for the carbapenemase-producing group; taken together, this may constitute a more important indicator of relative fitness cost in an enteric environment with limited nutrient-based carrying capacity and intense competition with other strains utilizing the same resource. Diminished group total bacterial density estimates in the antibiotic-supplemented media may indicate a different form of impaired bacterial fitness that could also impact population dynamics. Indeed, the biphasic exponential growth phase and the prolong lag phase duration seen in raw data plots (Fig 1) were not reflected in the model predictions nor the fitted plots of the growth curves respectively.
When bacteria of mixed origin are in a community, competition is known to occur, either due to limitations in resource availability or else as strains adapt and obtain an evolutionary advantage [78,79]. For instance, Ushijima and Seto demonstrated in an in vitro study that a group of normal intestinal commensals-E. coli, E. aerogenes, Enterococcus faecalis, Bacteroides ovatus and Fusobacterium varium-are capable of suppressing the growth of an intestinal pathogen-Salmonella Typhimurium-under certain conditions [80]. In addition, bacterial ecologists have noted that when members of a bacterial community are genetically similar, antagonism/inhibition is less likely to occur; in contrast, cooperation may ensue [81]. Given the similarity (same species) of our limited number of study strains, it is reasonable to expect that the fitness differential would predict the population dynamics of the bacteria community, especially given the similar profile of nutrient requirements and ecological niche. The fitness pattern of the five sets of representative pairs of strains from the five bla-gene-based groups in the mixed populations followed an overall similar trend to the estimates with twenty strains per group, each grown as mono-culture (Figs 1 and 2). That said, the margins of fitness advantage exhibited by the bla-negative pair and the pair of carbapenemase producers over the other pairs chosen for the competitive assay were slightly more pronounced (Tables 2 and 3). Our results showed a preponderance of the bla-negative strains in the bacterial community after 24-hour incubation in broth without a selection pressure, an expected outcome in the absence of mutual growth interference by the component E. coli strains. The data also showed a seeming suppression of the carbapenemase-producers beyond the limitations conferred by their growth rates; however, this may instead reflect CPE density under-estimation using the agar plates supplemented with 4 μg/mL meropenem (and perhaps point to a need for additional micro-nutrient supplementation such as zinc). Negri et al., in their in vitro mixed culture study of beta-lactam resistant and susceptible Streptococcus pneumoniae, similarly found no interference among S. pneumoniae strains with different profiles of resistance to beta-lactams [77].
Mollenkopf et al., in their CPE surveillance study of a swine production facility, found a 16.5% detection risk across environmental and fecal samples [49]. Notably, virtually all (~100%) of their CPE (bla IMP-64 /IncQ1) positive samples originated from the farrowing barns; importantly, the authors attributed this finding to the use of ceftiofur in the sows (on-label treatments as needed) and in the piglets (extra-label use for infection prevention and control such as following castration of male piglets) [49]. As mentioned earlier, when estimated across twenty strains, the carbapenemase-producers did not appear to exhibit a high fitness cost relative to other bla-positive strains; importantly, this suggests that even absent beta-lactam selection pressure and with a sufficiently high initial population, CPEs could persist at a low prevalence along with other beta-lactam-resistant Enterobacteriaceae. Instead, the relatively lower maximum OD of the CPE cultures may represent the real factor associated with lower levels of CPE currently found in food animal production environments.
The introduction into the bacterial growth media of different generations of beta-lactam antibiotics and at increasing concentrations was observed to change the population dynamics of the in vitro bacterial consortium. Increasing concentrations of ampicillin sodium gradually inhibited the bla-negative sub-population while selecting for the beta-lactamase resistant strains, including the full spectrum from TEM-1-type beta-lactamase producers through CPE strains. In contrast, low concentrations of ceftiofur hydrochloride (0.5 μg/mL) completely suppressed the susceptible populations of bla-negative and TEM-1-type beta-lactamase producers, thus effectively selecting for the CMY-2, CTX-M, and CPE resistant sub-populations. Increasing concentrations of ceftiofur hydrochloride suppressed the total bacterial density while further increasing the proportions of the higher-level beta-lactamase producing strains; for example, the carbapenemase producers showed about a forty-fold jump in relative proportion from non-selective broth through to 8 μg/mL of ceftiofur hydrochloride broth. Our results show that a 3 rd generation cephalosporin (3GC) such as ceftiofur provides a more than adequate selection advantage for carbapenemase producers, even in the absence of direct selection (i.e., carbapenem use) and minimal indirect co-selection. The presence of co-selected bla genes on the same plasmids could further aggravate the selection observed, given the carriage of many plasmid borne resistance genes has been observed to impose little or no additional fitness cost to the organism [82,83]. In our own study, it was difficult to find strains harboring either AmpC or ESBL genes that did not also harbor bla TEM-1 . Among the CPE strains, it was even more difficult to find strains that lacked not only bla TEM-1 , but also AmpC or ESBL genes. Genes are generally added to an existing arsenal of resistance and virulence factors in the strains. These findings agree with the observations in the surveillance study of the swine production facility by Mollenkopf et al. that ceftiofur provided adequate selection advantage for carbapenemase-producers to emerge into detectable levels [49].
In both ampicillin-and ceftiofur-supplemented CAMH-2 broth, the ESBL sub-population appeared to increase in dominance with increasing antimicrobial concentrations. This was likely facilitated by their relatively high minimum inhibitory concentrations (MIC) for these antimicrobials [84] and their fitness advantage over other bla-positive strains (Table 3). Overall, our findings among Enterobacteriaceae are remarkably similar to those of Negri et al.in gram-positive bacteria, which tested the effect of varying concentrations of amoxicillin, cefixime, cefuroxime and cefotaxime on a mixture of S. pneumoniae strains with MIC values ranging from susceptible to resistant across the antimicrobial agents. The newer generation betalactams completely suppressed the susceptible strains while selecting for higher resistance strains; in contrast, the less potent amoxicillin, at lower concentrations, mildly suppressed the susceptible strains while selecting effectively for the low-level beta-lactamase producing strains [77]. Ambler class C beta-lactamase enzymes, such as encoded by the plasmid-borne bla CMY-2 gene and the Ambler class A enzyme encoded by the plasmid-borne bla CTX-M gene have both been determined to be effective against 3GC antibiotics [26,84]. However, in 4 μg/mL ceftriaxone broth the CMY-2 type beta-lactamase strains demonstrated impaired group growth rates compared to the CTX-M � type beta-lactamase strains ( Table 2 and Fig 3). This observed difference in their growth potential suggests that with any given mixed bacterial community exposed to a similar or higher concentration of ceftriaxone, the strains harboring the bla CTX-M genes would be favored over the CMY-2 type beta-lactamase producing strains. Despite this theoretical outcome, in our mixed-strain cultures grown in ceftiofur broth, the relative proportion of the AmpC-type bla strains continued to increase along with the concentration of ceftiofur all the way to the highest experimental concentration. This occurred, rather than a plateauing or reduction in the relative proportion of these strains at the highest ceftiofur concentrations.
In the 10-strain competition assay, there was a preponderance of susceptible strains in the bacterial community in the absence of antibiotics, when compared with the beta-lactamase producers, as would be expected giving the superior fitness of the susceptible strains (Table 3). This is also consistent with contemporary estimates of beta-lactam resistance prevalence among indicator organisms such as E. coli, especially absent antibiotic selection-pressure. In contrast to estimated fitness values, the proportion of carbapenemase-producers, as measured by way of the MAC+MERO plates, was not comparable with the AmpC-type beta-lactamase strains. It should be noted that the component CPE population was estimated at 1 μg/mL meropenem (MAC+MERO) agar plates for the starting mixture and 4 μg/mL meropenem (MAC +MERO) agar plates for the post-incubation mixture, due to serious strain inhibition below the limits of detection at the higher meropenem concentration for the starting mixture; consequently, the CPE strain population was likely to be underestimated at the higher meropenem concentration. However, as the resistance breakpoint established by CLSI [46] is at 4 μg/mL meropenem (MAC+MERO) this was how the study was designed.
Importantly, the limitations inherent in models such as employed in this research must be considered in evaluating their usefulness. For instance, it should be noted that in nature the starting populations of bacteria with various resistance profiles are unlikely to be equal as was modeled in this study; therefore, the post-exposure changes in the prevalence of the resistant bacterial strains may not be as remarkable as we determined. Furthermore, in vivo antimicrobial concentrations are not constant, as was the case with our in vitro model; rather, the drug and active drug metabolite concentrations dynamically rise and fall, based on the dosing regimen and the drug distribution, metabolism and excretion. Still, recommended antimicrobial regimens are known to expand resistant coliform populations in livestock intestinal flora, sometimes for several weeks after the final drug administration [14]. Also, in nature, a distinct separation of bacterial groups by beta-lactamase enzyme profile is unlikely; that is, resistant bacteria frequently harbor multiple resistance determinants against a single class of antimicrobial agent, as well as to different classes of agents. Overlapping sets of resistance genes, as was frequently encountered during our isolate selection for this study, would be expected to add layers of complexity to the selection dynamics in nature. Although, our strain-group sample size was limited to twenty strains each (due to restricted availability), instead of thirty strains as projected by statistical theorem; Gompertz-3 model estimated standard errors of growth parameters were however very narrow, hence suggesting a rather robust estimate with the employed sample size. Consequently, we do not expect the strain-group sample size to negatively impact the parameter estimates. Similarly, due to availability constraint, several members of the carbapenemase producing strain-group bore supplementary bla-type genes. Given that the selection process for these strains (CPEs) are occurring in hospital environments rather than in the community where pan susceptible strains are more prevalent, their carriage of multiple resistance genes seems unavoidable due to the cumulative addition of novel genes to an existing array of resistance genes rather than de novo acquisition by susceptible strains. These may surely cloud the interpretation of the reported results, in particular the group-fitness estimates, and potential co-selection effects by test antimicrobials. In our mixed-strain experiments however, only a member of the pair of CPE representative strains bore a add-on bla gene (bla-CMY-2 ); we importantly note that, our data did not indicate a relative fitness cost or an amplified co-selection of these representative strains.
Our study did not factor in the role of innate bacteria resistance mechanisms, such as nonspecific efflux pumps and membrane-porin down regulation [85], because the presence of such mechanisms can be expected to exert a relatively uniform non-specific effect across study strains; however, for beta-lactam antibiotics this is not likely to interfere with interpretations since most resistance is enzymatic and the damage inflicted by the antibiotic is to the cell wall. Lastly, the influence of far more abundant and niche-competing anaerobic intestinal commensals, or free-living environmental strains, in antimicrobial selection was not modeled in this study. Although the conditions of these in vitro models may not perfectly approximate in vivo or environmental realities, they nonetheless constitute a reasonable first step in a systematic and order approach to this challenge.
Overall, our hypothesis that older generation beta-lactam antibiotics of lesser priority-such as ampicillin-can also provide a selection advantage to highest priority resistance types-such as 3rd/4th generation cephalosporins and carbapenems-albeit less efficiently when compared to 3 rd generation cephalosporins, was supported by the results of these two in vitro experimental studies. Currently, beta-lactam antimicrobial resistance due to carbapenemase producing Enterobacteriaceae is not a known existential therapeutic threat in animal agriculture; in contrast, human health care infections caused by the K. pneumoniae producing KPC-type bla enzyme are a global concern [1]. These strains, along with other Enterobacteriaceae bearing less prevalent carbapenemase encoding genes (e.g., bla NDM , bla VIM & bla IMP ) have been reported in many U.S. states [16]. Presently, the challenge posed by carbapenemase-producing bacteria is limited to specific settings in the human healthcare system. Although E. coli strains with carbapenemase-encoding genes (including the bla KPC gene) have been identified [86,87], community acquired infections with strains bearing this order of beta-lactam resistance are still relatively rare [88].
Our study suggests that if introduced into food-animal populations, perhaps through surface water downstream from hospital and wastewater treatment plant effluent discharge [89], CPEs along with other high priority beta-lactamase resistance profiles could be expanded due to lower-priority beta-lactam use in food animals, and subsequently spread back to the human community through food-animal products and via the livestock environment. To further clarify and characterize these findings, we suggest additional studies, both observational and experimental. For example, an in vitro continuous anaerobic medium such as employed by Ushijima and Seto might provide a better model of the selection dynamics expected in a monogastric mammalian bowel compared to the 10-strain batch culture we used [80]. Further, pharmacokinetics-pharmacodynamics (PK-PD) mathematical modeling of these selection dynamics and ultimately in vivo animal studies would be logical next steps.
In conclusion, this study showed that the absence of direct carbapenem selection pressure in food-animal production cannot be relied upon alone to reduce the spread of bacterial strains with reduced carbapenem susceptibility. Use of commonly prescribed older-generation beta-lactams such as ampicillin and ceftiofur can expand both ESBL Enterobacteriaceae and CPE prevalence in commensal and pathogenic bacterial communities.
Supporting information S1