A mouse model of occult intestinal colonization demonstrating antibiotic-induced outgrowth of carbapenem-resistant Enterobacteriaceae

The human intestinal microbiome is a complex community that contributes to host health and disease. In addition to normal microbiota, pathogens like carbapenem-resistant Enterobacteriaceae may be asymptomatically present. When these bacteria are present at very low levels, they are often undetectable in hospital surveillance cultures, known as occult or subclinical colonization. Through the receipt of antibiotics, these subclinical pathogens can increase to sufficiently high levels to become detectable, in a process called outgrowth. However, little is known about the interaction between gut microbiota and Enterobacteriaceae during occult colonization and outgrowth. We developed a clinically relevant mouse model for studying occult colonization. Conventional wild-type mice without antibiotic pre-treatment were exposed to Klebsiella pneumoniae but rapidly tested negative for colonization. This occult colonization was found to perturb the microbiome as detected by both 16S rRNA amplicon and shotgun metagenomic sequencing. Outgrowth of occult K. pneumoniae was induced either by a four-antibiotic cocktail or by individual receipt of ampicillin, vancomycin, or azithromycin, which all reduced overall microbial diversity. Notably, vancomycin was shown to trigger K. pneumoniae outgrowth in only a subset of exposed animals (outgrowth-susceptible). To identify factors that underlie outgrowth susceptibility, we analyzed microbiome-encoded gene functions and were able to classify outgrowth-susceptible microbiomes using pathways associated with mRNA stability. Lastly, an evolutionary approach illuminated the importance of xylose metabolism in K. pneumoniae colonization, supporting xylose abundance as a second susceptibility indicator. We showed that our model is generalizable to other pathogens, including carbapenem-resistant Escherichia coli and Enterobacter cloacae. Our modeling of occult colonization and outgrowth could help the development of strategies to mitigate the risk of subsequent infection and transmission in medical facilities and the wider community. This study suggests that microbiota mRNA and small-molecule metabolites may be used to predict outgrowth-susceptibility. CiXsSNTB417URyJP4k6GoX Video Abstract Video Abstract


Introduction
It is well-established that the gut microbiome plays a critical role in the health of the host [1,2]. Conditions like irritable bowel syndrome and Crohn's disease have been linked to dysbiosis of the microbiome [3], as have other complex conditions like obesity and diabetes [4]. The commensal microbiota also plays a critical role in 'colonization resistance' , or the prevention of pathogenic organisms from taking up residence in the gut, from which they could spread to other organs [5,6]. Factors that disrupt the microbiome, such as antibiotics [7], nutritional changes [8] or competition between microbes [9], can alter the microbiome's population structure, which may allow for colonization or outgrowth by pathogenic organisms.
Outgrowth of pathogenic bacteria, like carbapenemresistant Enterobacteriaceae (CRE), is of particular concern in hospitals and other healthcare settings. CRE such as Klebsiella pneumoniae, Escherichia coli, and Enterobacter cloacae are an important cause of antibiotic-resistant infections worldwide [10]. According to a 2019 report from the US Centers for Disease Control and Prevention, CRE are considered an urgent threat, resulting in 13,100 hospitalizations, 1100 deaths, and an estimated $130 million in healthcare costs in the USA annually [11]. This has led hospitals to implement intake screening protocols to limit the spread of CRE when patients are admitted or transferred from other facilities [12,13]. Screening allows patients with active infection or asymptomatic colonization [14,15] to be detected and properly isolated from vulnerable populations.
Confounding screening efforts, patients with low-level, sub-clinical (occult) colonization are not easily detected by intake screenings or routine surveillance cultures [16]. Outgrowth from occult colonization is hard to differentiate from nosocomial transmission without detailed molecular analyses [17,18]. Robust methods to distinguish between these two scenarios would have important clinical implications. While nosocomial transmission would warrant extensive screening of an entire hospital unit, CRE outgrowth in a single patient could be addressed as an isolated event.
Occult colonization can also play a role in community spread of CRE. CRE can persist in the gastrointestinal tract of healthy individuals and the general population [19]. Travelers to CRE-endemic regions are at risk of being exposed to CRE, even without visiting a medical facility, and they may carry it back to their home country upon return from travel [20,21].
Occult colonization is, by definition, difficult to detect directly, but we hypothesized that there might be other signals that indirectly report on low-level colonization and subsequent risk of outgrowth. In the current study, we constructed a mouse model for occult colonization and used both 16S rRNA and shotgun metagenomic sequencing to demonstrate that occult colonization remodels the microbiome. We then identified vancomycin receipt as a condition under which we could achieve incomplete outgrowth after antibiotic treatment from an undetectable state. This allowed us to identify functions involved in mRNA stability that were enriched in mice where outgrowth was observed. Finally, we identified mutations that evolved in K. pneumoniae under vancomycin selection, suggesting xylose may play a role in outgrowth.

Development of a murine model of occult colonization
We established a murine model to identify markers of occult intestinal colonization by multidrug-resistant Enterobacteriaceae. We chose C57BL/6N wild-type mice and a dominant lineage, sequence type 258 (ST258) strain of K. pneumoniae (KPNIH1) [18,22] for this model. Furthermore, ST258 has been demonstrated to colonize patients long-term [23], and it has a well-curated transposon library of knockout mutants [24]. We initially studied 3 variables: inoculum size, location of occult colonization, and time until antibiotics (Fig. 1a). Most infection models use a very large inoculum (>10 7 CFU), but as we were interested in occult colonization, we chose a more conservative range of inoculum sizes. We titrated from 10 3 to 10 5 CFU and observed at days 4-6 post-gavage that 10 5 CFU inoculum was rarely detectable, whereas the 10 4 and 10 3 CFU inocula were never detectable in stool (Fig. 1b).
To determine if mice were colonized at levels below our detection threshold, we administered antibiotics, as this has been shown to trigger outgrowth of resistant strains [25]. Initially, we chose an antibiotic cocktail, comprising ampicillin, vancomycin, neomycin, and metronidazole. The KPNIH1 strain we chose to work with is resistant to all 4 of these antibiotics (Additional file 1 a), and this cocktail is widely used in murine studies to ablate microbiota [26]. Administering the antibiotic cocktail 1 week postgavage allowed K. pneumoniae to grow from an occult state to high levels in stool (Fig. 1b). Of note, because we chose to singly house mice, K. pneumoniae outgrowth was not a result of mouse-to-mouse transmission.
Next, we determined the lowest inoculation dose possible for K. pneumoniae outgrowth from an occult state. While outgrowth was observed for the 10 5 CFU inoculum in all mice, 10 4 CFU resulted in outgrowth in most but not all mice. 10 3 CFU very rarely resulted in outgrowth, suggesting that K. pneumoniae is typically unable to colonize occultly from such a starting inoculum (Fig. 1b). Thus, we chose 10 5 CFU as the inoculum for subsequent experiments.
K. pneumoniae is capable of colonizing many body sites, and some strains can translocate across the gut barrier [27]. To identify where in the host K. pneumoniae establishes occult colonization, we examined the gastrointestinal tract as well as peripheral tissues at day 7 post-gavage (Fig. 1a, middle). We identified K. pneumoniae in the cecum of the majority of mice, but not the small intestines, mesenteric lymph nodes, liver, or blood plasma (Fig. 1c), suggesting that the primary reservoir of K. pneumoniae occult colonization is the cecum with no systemic spread. Furthermore, to determine which of these body sites show K. pneumoniae outgrowth following antibiotic cocktail treatment, we also examined them at day 21 post-gavage using the mice from the titration experiment from (Fig. 1a, top, 10 5 CFU inoculum). We found that aside from stool, the cecum showed the most abundant K. pneumoniae outgrowth (Additional file 1 b).
To determine the duration of K. pneumoniae occult colonization, we initiated receipt of antibiotics at 1, 2, 4, or 8 weeks post-inoculation (Fig. 1a, bottom). We observed that K. pneumoniae consistently showed outgrowth when antibiotic was administered at 1 or 2 weeks but rarely if a Study designs to test inoculation size, occult location, and time until antibiotic. (Top) Mice were orally gavaged with 10 3 , 10 4 , or 10 5 CFU of K. pneumoniae (Kp) or the PBS vehicle control at day 0 (triangle). An antibiotic cocktail was given from day 7 to day 21. Days of stool collection are marked by circles. Mice were euthanized at day 21, and cecal contents collected. (Middle) Mice were orally gavaged with 10 5 CFU of K. pneumoniae, and tissues were plated at day 7. (Bottom) Mice were orally gavaged with 10 5 CFU of K. pneumoniae, and different groups of mice were given antibiotic for 2 weeks starting at t = 1, 2, 4, or 8 weeks post-gavage. Stool and cecal contents were collected at (t+2) th week. b CFU of K. pneumoniae in stool and cecal contents (cec). N=6. The dotted line marks the limit of detection for CFU. c CFU of K. pneumoniae in stool, cecal, and small intestinal (SI) contents and the mesenteric lymph nodes (MLN), liver, and blood plasma. N=10. d CFU of K. pneumoniae at (t+2) th weeks in stool and cecal content. N=6. For b, Fisher's exact test with Hommel correction for multiple testing was used to compare numbers of mice showing outgrowth versus no-outgrowth in inoculated mice to those in the PBS control group. Outgrowth is defined as CFU > 0. *p value < 0.05, **p value < 0.01 at all at 4 or 8 weeks, suggesting that, in our model, K. pneumoniae occultly colonizes the gut for up to 4 weeks (Fig. 1d). Our optimized occult colonization model using an inoculation dose of 10 5 CFU and administration of antibiotic at 1 week post-gavage for 2 weeks duration (Fig. 2a), mimics a clinically relevant timecourse of bacterial outgrowth. Having a robust model is our first step to finding markers for occult colonization and outgrowth.

Clinically relevant antibiotics ampicillin, vancomycin, and azithromycin induce outgrowth of K. pneumoniae
We reasoned that determining which individual antibiotic(s) in the 4-drug cocktail is responsible for driving outgrowth could provide insight into the underlying biology. To this end, we systematically tested each of the antibiotics in the cocktail (Additional file 1 c) for whether they could individually induce outgrowth (Fig. 2b). Among these, ampicillin and vancomycin are known clinically to predispose individuals to K. pneumoniae colonization [28,29]. Accordingly, we observed significant outgrowth of K. pneumoniae after treatment with ampicillin (p < 0.01), a trend towards outgrowth after vancomycin (p = 0.10 at day 14), but no outgrowth with metronidazole or neomycin. We also tested the clinically relevant antibiotics ciprofloxacin, rifaximin, and azithromycin ( Fig. 2b and Additional file 1 c) since these Fig. 2 Ampicillin, vancomycin, and azithromycin induce outgrowth of occult K. pneumoniae. a Mice were orally gavaged with 10 5 CFU of K. pneumoniae at day 0 (triangle) and treated with antibiotic at day 7 for 2 weeks (orange region). Days of stool collection are marked by circles. b, c CFU of K. pneumoniae in b stool at days 0, 7, 14, and 21 and c cecal contents at day 21. The dotted line marks the limit of detection for CFU. N=5-11. d Alpha diversity (number of amplicon sequence variants or ASVs) of antibiotic-treated microbiota at day 14. e Same experiment as (a), but the bacteria gavaged were other carbapenemase producing strains of K. pneumoniae, E. cloacae, or E. coli, followed by ampicillin treatment. Levels of Enterobacteriaceae in stool at day 21. Measurement was done by qPCR using primers at bla KPC gene. N=4-9. For b and e, Fisher's exact test with Hommel correction for multiple testing was used to compare numbers of mice showing outgrowth versus no-outgrowth in inoculated mice to those in the PBS control group. Outgrowth is defined by detected CFU > 0. For d, Kruskal-Wallis H test was performed to test significance, followed by pairwise Bonferroni-corrected Mann-Whitney U tests to compare each treatment group against the control group. *p value < 0.05, **p value < 0.01, ***p value < 0.001 drugs are commonly used by travelers in CRE-endemic regions [30]. Among these, only azithromycin produced a trend towards K. pneumoniae outgrowth (p =0.075 at day 14; Fig. 2b). Importantly, the resistance of K. pneumoniae to all the antibiotics tested (Additional file 1 a) and the retention of K. pneumoniae in the cecum of many no-outgrowth mice (Fig. 2c) ruled out the possibility that K. pneumoniae was killed by the antibiotics or that occult colonization failed. More likely, outgrowth is influenced by antibiotic-specific perturbations to the microbial community in the gut.
To probe the general effect of 1 week of antibiotics on the microbiota, we used 16S rRNA gene sequencing to measure microbial alpha diversity (measure of the ecological richness within a single sample). In agreement with past studies [7,26,31,32], low alpha diversity was associated with the antibiotic cocktail as well as individual ampicillin, vancomycin, and azithromycin treatments (p < 0.05; Fig. 2d). Together, these results suggest that antibiotics that disrupt microbial community structure tend to promote K. pneumoniae outgrowth.
The experiments up to this point were performed on female mice. To examine whether occult colonization and antibiotic outgrowth are affected by the gender of the mouse, we subjected male mice to ampicillin outgrowth using identical model parameters from Fig. 2a. We observed similar absence of K. pneumoniae in stool after inoculation followed by robust outgrowth after ampicillin treatment (Additional file 2 a), suggesting that host gender is not a contributing factor to occult colonization and outgrowth.
Next, to assess whether antibiotic outgrowth is a general feature of Enterobacteriaceae occult colonization, we tested whether ampicillin treatment would result in outgrowth from an occult state by other carbapenemase producing strains of K. pneumoniae as well as carbapenemase producing strains of Enterobacter cloacae and Escherichia coli [17,33,34]. As it is difficult to identify a precise combination/dose of antibiotics that would select for all these carbapenemase producing strains from stool plating, we instead performed quantitative PCR using a primer set that binds to bla KPC gene. We validated this approach by showing that the two methods produced similar KPNIH1 abundance (Additional file 2 b). Using qPCR, there was significant outgrowth of all strains tested upon ampicillin receipt (p < 0.01 at day 21; Fig. 2e), demonstrating that occult colonization (Additional file 2 c) and antibiotic-induced outgrowth (Fig. 2e) are generalizable to the Enterobacteriaceae family. We recognize that bla KPC gene is on a plasmid where the copy number may influence the qPCR result; thus, we used another primer set that binds to a chromosomal gene at the Klebsiella capsular polysaccharide synthesis gene cluster.
This Klebsiella-specific primer showed markedly similar abundance of K. pneumoniae strains compared with the bla KPC primer (Additional file 2 d). The result agrees with our previous work that showed that the plasmid containing bla KPC gene has a low copy number (1.5x chromosomal for KPNIH1; 3.4x for KPNIH27) [33] in the gut environment, suggesting that the bla KPC primer set is a reasonable choice for detecting KPC+ bacteria using qPCR-based assays.

Occult colonization with K. pneumoniae perturbs the structure of the microbiome
We hypothesized that the occultly colonized state could be accompanied by a shift in other members of the microbial community. We used 16S rRNA gene sequencing to compare baseline (day 0) stool samples with stool from mice occultly colonized with K. pneumoniae at day 7. We observed a significant reduction in the abundance of the Erysipelotrichaceae family (p = 8 x 10 -8 ; Additional file 3 a) and its component genus Dubosiella (p = 0.0002; Additional file 3 b) in the stool of occultly K. pneumoniae colonized mice. Indeed, Erysipelotrichaceae consistently decreased in the vast majority of occultly colonized microbiota (Additional file 3 c), and this pattern could be attributed to decreases in Dubosiella (Additional file 3 d).
To determine the reproducibility of these changes in the microbiome, we repeated the occult colonization experiment and this time performed shotgun metagenomic sequencing to obtain species-level resolution. Taxonomic classification was performed using Kraken2 [35] and the Genome Taxonomy database (GTDB) [36] as reference since it was found to be superior to RefSeq, the standard database (Additional file 4 a). First, we looked at beta diversity and found that baseline microbiota formed a separate cluster from occultly colonized microbiota, suggesting differences in microbiome structure (Fig. 3a). Similar to 16S rRNA results, there was a significant reduction in the abundance of Erysipelotrichaceae (p = 0.0006; Fig. 3b and Additional file 7), its component species Faecalibaculum rodentium (p = 0.0003), and a trend in Dubosiella sp004793885 (p = 0.071) and Dubosiella newyorkensis (p = 0.071) ( Fig. 3c and Additional file 8). While the relative abundance of Erysipelotrichaceae dropped in nearly every occultly colonized animal (Fig. 3d), control mice that were mock gavaged with PBS did not show this reduction, suggesting that the difference is not due to random drift over time or technical confounders but rather a feature of K. pneumoniae occult colonization. However, we did observe a different Erysipelotrichaceae abundance in the two groups at baseline. The difference in starting levels could be partially explained by the larger sample size for the experimental animals (N=28) compared to control animals (N=8) but likely also reflects natural variation in the abundance of this taxon in the mouse microbiome.

Vancomycin-induced outgrowth is incomplete and predicted by the functional potential of the microbiome
We observed in the single drug experiments (Fig. 2b) that vancomycin treatment resulted in K. pneumoniae outgrowth in only 40% of animals, whereas the 4-drug cocktail or ampicillin resulted in 100% of animals showing outgrowth. This observation led us to hypothesize that the microbiota at day 7, just before vancomycin treatment, were taxonomically or functionally different between susceptible and non-susceptible mice. First, we reproduced the stochastic outgrowth phenotype with a larger sample size (Fig. 4a). Next, we investigated taxonomic composition using shotgun metagenomic sequencing. At the genus level, we found a statistically significant enrichment of Akkermansia in non-susceptible mice at day 7 (p = 0.043) (Additional file 4 b); however, significance was not observed for any specific species (lowest p = 0.156) (Additional file 4 c). As this was inconclusive, we performed a second metagenomic analysis to identify enriched gene functions (Additional file 4 d). We constructed a machine learning classifier to measure the predictive value of functional features for susceptibility. This model achieved an areaunder-the-curve (AUC) score of 0.939 (highest score = 1) for distinguishing susceptible microbiota from nonsusceptible microbiota (Fig. 4b). The model chose 4 genes as the optimal marker set (Fig. 4c). The gene with the largest effect size encoded polyribonucleotide nucleotidyltransferase, which is involved in mRNA degradation. Another gene in the marker set encoded 50S ribosomal protein L16, which is involved in mRNA translation. Both genes were enriched in nonsusceptible microbiota (Fig. 4c).
Thus, it appears, from a gene-based perspective, that functions related to mRNA, particularly its degradation, are predictors of nonsusceptibility.

Vancomycin treatment selects for mutations in a K. pneumoniae transcriptional regulator of xylose metabolism
Up to this point, we had focused on characteristics of the microbiome that contribute to occult colonization and antibiotic-induced outgrowth. We next sought to perform a complementary experiment looking for genes in K. pneumoniae that contribute to colonization and growth. To this end, we passaged K. pneumoniae under antibiotic selection in mice to identify genes important for its outgrowth. Our rationale was that acquired mutations would elucidate the landscape of selective pressures in the antibiotic-disrupted gut, allowing us to predict what metabolites are important for outgrowth and susceptibility. Specifically, we evolved K. pneumoniae by iterative passages through the gut of vancomycintreated mice (Fig. 5a). At the 8th iteration, 3 evolved K. pneumoniae strains from each lineage were picked for whole-genome sequencing and analyzed for SNP variants (Additional file 5 a). Notably, 5 out of the 6 lineages contained an isolate with a missense variant in the xylose-binding domain of the xylR gene (Fig. 5b). XylR is a transcriptional activator of the xylose operon. Structurally, XylR contains a xylose-binding domain that is critical to its transcriptional activity [37], and we noted that all five xylR SNP mutations were located in this exact domain, with one SNP being precisely located in a residue (Arg240) critical for functional activity [37]. Xylose can be fermented to provide energy for growth [38]. Thus, it is possible that reduced xylose may be a selective pressure in the antibiotic-treated gut causing K. pneumoniae non-susceptibility.
We next tested the 5 independently evolved xylR mutants for altered antibiotic resistance but found no Mice were pre-treated with vancomycin for 7 days, orally gavaged with 10 5 CFU of K. pneumoniae and maintained on vancomycin for 7 more days to allow K. pneumoniae growth. Stool was diluted 1000-fold to achieve ~10 5 K. pneumoniae CFU and gavaged into another vancomycin pre-treated mice. Such transfer of stool microbiota was performed for 8 iterations into singly housed mice to avoid any cross-contamination. K. pneumoniae colonies were cultured from the stool of the 8th iteration mouse. Colonies from independent lineages were whole-genome sequenced. Plate swipes containing aggregate pools of K. pneumoniae cultured directly from stool were also sequenced for each lineage. b Summary of XylR mutations derived from 5 separate lineages together with their allelic frequencies obtained from plate swipes. c In vitro anaerobic growth assays of evolved strains xylR-L and transposon strains xylR-Tn in minimal media containing xylose or glucose after 24 h. N=4. d Levels of xylose in stool before and after vancomycin or ampicillin treatments. N=10. e, f In vivo competition between e the wild type and xylR-Tn1 transposon strain, or f the wild type and xylR-L4 evolved strain in mice undergoing vancomycin or ampicillin treatment. Mice were orally gavaged with 10 5 CFU of each K. pneumoniae strain at day 0 (triangle). Strain levels at days 3 and 7 were quantified by qPCR using specific primers to each xylR allele. N=5. g Levels of xylose in stool at day 0, 7, or 14. Stool was obtained from the mice shown in Fig. 4a. N=13-14. For c, 1 a). This suggests that the xylR variants did not confer altered resistance to vancomycin or ampicillin. We next assessed whether the variants alter xylose utilization by comparing the growth of the wild type and 5 xylR mutants in minimal media with xylose or glucose as the sole source of carbon (Fig. 5c). While the wild type and all 5 mutants showed similar growth on glucose, the mutants showed significantly higher growth on xylose (p < 0.001), suggesting that the SNPs are gain-of-function mutations that enhance the utilization of xylose. Supporting this, loss-of-function xylR transposon mutants [24] were incapable of utilizing xylose as a sole carbon source (Fig. 5c). These data suggest that the murine evolved strains are gain-offunction with superior growth on xylose compared with the wild type.

Xylose metabolism contributes to K. pneumoniae growth in the context of antibiotic therapy
Based on these data, we hypothesized that antibiotic treatment reduces the level of xylose, driving K. pneumoniae to evolve mutations that enhance xylose utilization.
To explore this idea, we quantified the amount of xylose in stool before and after antibiotic treatment and indeed found significantly reduced levels of xylose in both vancomycin-treated (p = 6 x 10 -5 ) and ampicillin-treated stool (p = 2 x 10 -5 ) (Fig. 5d). The adaptation of K. pneumoniae to a changing xylose supply suggests the importance of xylose metabolism for K. pneumoniae growth. We tested this by gavaging mice with equal amounts of wild-type K. pneumoniae and a xylR transposon knockout. We found lower levels of the xylR transposon strain compared to the wild-type strain in vancomycin-treated mice (p = 0.012 at day 7), suggesting that the induction of xylose utilization via XylR provides a competitive growth benefit (Fig. 5e). To test whether the xylR variants in the evolved strains provide enhanced growth in vivo, we colonized vancomycintreated mice with equal amounts of the wild-type strain and an evolved strain (Fig. 5f ), the latter of which possessed a SNP only in the xylR gene (Additional file 5 a). We found that the evolved strain significantly outcompeted the wild-type strain (p = 0.012). To determine if the growth superiority of the evolved strain is limited to vancomycin-treated mice, we performed a similar experiment in ampicillin-treated mice (Fig. 5f ) and observed that the evolved strain also showed enhanced growth compared to wild-type strain (p = 0.012). These results demonstrate that xylose metabolism is important for the growth of K. pneumoniae in the context of vancomycin and ampicillin receipt.

Elevated intestinal xylose is a marker of susceptibility to K. pneumoniae outgrowth
These results prompted us to revisit the incomplete vancomycin-induced outgrowth of occultly colonized mice (Fig. 4a). We hypothesized that xylose may be associated with susceptibility prior to antibiotic treatment. We found that xylose was significantly higher in susceptible compared to nonsusceptible stool at day 7 (p = 0.026) and, interestingly, at the day 0 baseline as well (p =0.038; Fig. 5g). These results support the use of pre-antibiotic xylose levels as a marker of susceptibility.

Discussion
In this work, we present an optimized mouse model for use in occult gut colonization studies. It is clinically relevant, uses common immune-competent mice without antibiotic pre-treatment, and is stable and reproducible. We observed consistent antibiotic-induced outgrowth patterns over a total of N=330 mice, including incomplete vancomycin-associated outgrowth (Additional file 5 b). We successfully reproduced these essential features in a different mouse facility (data not shown). Similar to previous mouse models of K. pneumoniae colonization [39], we chose as our occult colonizer a genetically tractable, human-associated ST258 strain of K. pneumoniae, but demonstrated that other Enterobacteriaceae may also be used. Using this system, we observed that occult colonization remodels the microbiota and that genes regulating mRNA stability and xylose metabolism influence subsequent antibiotic-triggered outgrowth. Previous studies on occult colonization have shown the ability of exogenous strains to adhere to the intestinal mucosa [40], but mechanistic understanding is fairly limited. In other studies involving a somewhat similar process of persistent colonization, it was found that some of the important factors are iron scavenging [41], utilization of sialic acid from mucus layer [42], and bacteriocins to compete against other strains [43]. It may be interesting to see whether these factors are also important in occult colonization. Similar approaches could be conducted for other strains of interest to study their occult properties.
We studied the impact of low-level K. pneumoniae colonization on the microbiome and observed a reduced abundance of Erysipelotrichaceae. Due to the large variability in baseline taxa abundances, the experiments presented here did not have the power to identify marker taxa of occult colonization from the complex native gut microbiome. Future experiments using more timepoints, additional animals or a defined gut community could be used to identify these marker taxa.
We also investigated the functional potential of susceptible microbiota and found genes regulating mRNA stability as a predictor of susceptibility. Response to stress caused by antibiotics involves changes in the transcriptional state of the bacteria [44], particularly elevated synthesis of selected stress response proteins [45,46]. To achieve this regulation, it is thought that keeping a low general mRNA abundance allows crucial mRNA encoding stress-related proteins to gain access to ribosomes for translation [47,48]. Our metagenomic analysis of nonsusceptible microbiota potentially supports this paradigm of translational resource reallocation, in which mRNAs not related to stress responses are selectively degraded. Our findings of enhanced mRNA degradation, along with enhanced ribosomal capacity, may allow the microbiota to better cope with antibiotic stress and resist K. pneumoniae outgrowth. It is a goal of future studies to probe how the dynamics and regulation of mRNA stability affect susceptibility.
We also explored the impact of small molecules on K. pneumoniae outgrowth and identified xylose as a potential marker molecule. An advantage of using xylose as a marker of susceptibility is that xylose is also present in human stool as a breakdown product of dietary fiber [49,50]. Plant-derived dietary xylose is a major carbon source for the gut microbiota, with different microbiota members playing roles in branched-chain xylan degradation and the cross-feeding of other members of the community [51][52][53]. We found that K. pneumoniae relies on this source of free xylose for growth, in agreement with another recent study [54], and we went a step further to link xylose levels with susceptibility. Another strength of xylose as a marker of susceptibility is that the level is relatively stable over time, as both day 0 and day 7 in our model showed significant differences in xylose abundance between susceptible and nonsusceptible mice. Moreover, xylose abundance can be measured using a commercial kit in a matter of hours and therefore does not require costly metabolomics. Future work to improve health outcomes may investigate how to therapeutically reduce levels of free xylose, such as by catalyzing its fermentation to short chain fatty acids [38], which have numerous health benefits for the host [55,56].
We acknowledge that singly housing of mice is not in the best interests of animal welfare due to stress arising from social isolation, but it is important for the mouse model to eliminate any chance of K. pneumoniae transmission through stool coprophagy. In group housing, for example, we have observed that during antibiotic-induced outgrowth, when a mouse in a cage becomes positive for K. pneumoniae, all mice in that cage would also become positive for K. pneumoniae within 1 or 2 days. Without an accurate statistical count of the frequency of K. pneumoniae-positive animals through single housing, it would not have been possible to unravel the difference in outgrowth frequency between ampicillin and vancomycin. Thus, single housing of animals was crucial in the mouse model given our study design. Nonetheless, we do not dismiss the possibility of its utility in group housing as long as it is clear that outgrowth should only be determined as a binary readout in such case.
There are a number of limitations in the current work. Xylose abundance is less effective for predicting susceptibility to ampicillin-induced outgrowth suggesting that, during ampicillin treatment, xylose may be compensated for by other carbon sources, such as fucose or sialic acid [42,57], thus allowing outgrowth regardless of xylose availability. The lack of a strict requirement for xylose in ampicillin-induced growth agrees with the observation that the xylR transposon mutation did not decrease fitness (Fig. 5e). Second, we did not test whether Akkermansia may be protective against K. pneumoniae outgrowth. Although from metagenomic sequencing Akkermansia showed higher abundance in nonsusceptible mice, more work is needed to examine whether an initial colonization with Akkermansia may reduce subsequent K. pneumoniae outgrowth. Third, to further elucidate the role of mRNA stability in susceptible animals, it may be helpful to study RNA turnover or ribosome profiling of the microbiota. Fourth, we identified xylose as a marker of susceptibility but did not examine their causal relationship. Feeding mice with a high-xylose diet may further clarify on this issue. Lastly, the microbiome of mouse and human have considerable differences [58], hence further work is needed to confirm whether our findings in the mouse microbiome extend to human microbiome.

Conclusion
CRE infections are responsible for high patient mortality and comorbidities and are difficult to treat because of their resistance to broad classes of antibiotics. The epidemiology of CRE transmission is complicated to track because of frequent patient transfers between healthcare facilities and increasing global travel including leisure or medical tourism. Occult CRE colonization compounds this difficulty by allowing an undetectable reservoir in carriers and causing an underestimation of CRE prevalence in healthcare systems and in the community. It is clinically important to understand which individuals are predisposed to outgrowth in order to estimate the risks of different antibiotic treatments. We identified two susceptibility factors, mRNA stability pathways and intestinal xylose abundance, using two complementary approaches. These results provide taxonomic, functional, and genetic insight into CRE occult colonization and outgrowth, and pave the way for mechanistic studies to come.

Mouse husbandry
C57BL/6N wild-type mice were bred under specificpathogen-free conditions at an American Association for the Accreditation of Laboratory Animal Care-accredited animal facility at the NHGRI and housed in accordance with procedures outlined in the Guide for the Care and Use of Laboratory Animals. All experiments were performed under an animal study proposal approved by the NHGRI Animal Care and Use Committee. Cages, water, and feed were pre-sterilized by autoclaving before use. Gender-and age-matched mice between 7 and 12 weeks old were used in all experiments. Cross-contamination between cages was minimized during cage handling by frequent changes of gloves and disinfection of hands with 200 ppm chloride dioxide, and this was practiced by both the user and animal caretaker. For all experiments, mice were singly housed to prevent transmission of K. pneumoniae by coprophagy.

Inoculation of K. pneumoniae, E. coli, and E. cloacae
K. pneumoniae, E. coli, and E. cloacae were cultured overnight in LB at 37°C under aerobic conditions with shaking at 200 rpm. Unless otherwise stated, mice were orally gavaged with 10 5 CFU, and all experiments involving K. pneumoniae used MKP103 strain as the wild type. MKP103 is a derivative of ST258 KPNIH1 in which the bla KPC gene was deleted [24]. All other K. pneumoniae, E. coli, and E. cloacae strains used in the study harbor the bla KPC gene, are resistant to carbapenem and ampicillin, and are derived from human sources. More details of each strain are given in Additional file 6 b.
For in vivo competition of K. pneumoniae strains, mice received vancomycin or ampicillin for 4 days prior to being orally gavaged with 10 5 CFU of an evolved/transposon strain and 10 5 CFU of MKP103. Mice were maintained on antibiotics for 7 more days.

Collection of stool and cecal content
Stool pellets were collected from mice at days 0, 7, 14, and 21 unless otherwise stated. For each time point, a fresh pellet was used for CFU counting, while another pellet was snap-frozen on dry ice and stored at −80°C for 16S rRNA or metagenomic sequencing. Cecal contents were obtained by making a cut along the cecal pouch. Small intestinal contents were flushed out with PBS using a syringe and oral gavage tube. Mesenteric lymph nodes and livers were homogenized through 70 μm strainers prior to plating. Whole blood was collected into blood plasma tubes and plated directly.

Quantification of CFU
To determine CFU for K. pneumoniae MKP103, samples were serially diluted and plated onto LB agar containing 100 μg/ml vancomycin, 100 μg/ml metronidazole, 100 μg/ml neomycin, and 50 μg/ml ampicillin. Plates were incubated for 18 h at 37°C under aerobic conditions. Negative control plating of wild-type stool typically showed zero colonies. Five to 10 randomly selected colonies were verified per experiment with PCR using Klebsiella-specific primers (see Additional file 6 a for primer sequences).

In vitro growth assays
In vitro growth assays were performed in M9 minimal medium (BD) supplemented with 20 mM xylose or glucose (Sigma). Wild-type and xylR evolved/transposon strains were first cultured at 37°C for 24 h in an anaerobic box (AnaeroPack system, Mitsubishi Gas Chemical Company) with disposable single-use anaerobic sachets (Gas-Pak EZ Anaerobe Container System, BD). Then, strains were washed and resuspended at 10 5 CFU/ml in M9 with 20 mM sugar and incubated anaerobically in 96-well plates in 100 μl final volumes for 24 h. After 24 h, the OD at 600 nm was measured with a plate reader (Biotek).

Xylose quantification
Free xylose levels were quantified using a colorimetricbased D-xylose assay kit (Megazyme). Stool samples were weighed, resuspended in 1 ml ddH2O, vortexed for 30 s, and then centrifuged at 13,000 g for 2 min to remove large particles. Next, supernatants were transferred to new tubes, and the manufacturer's instructions were closely followed to measure xylose levels, including the step of Carrez purification to remove protein inhibitors.

Determinaton of minimum inhibitory concentrations
MICs for vancomycin, metronidazole, ampicillin, ciprofloxacin, and azithromycin were determined using MIC test strips (Liofilchem) according to the manufacturer's instructions. MICs for rifaximin and neomycin were determined for two-fold serial dilutions of the antibiotic from 0.5 to 256 μg/ml. Bacteria were inoculated at 5 x 10 4 CFU in 100 μl in 96-well plates. After overnight incubation at 37°C, OD600 was measured. MIC was defined as the lowest concentration with no bacterial growth.

DNA extraction
DNA was extracted from stool using the DNeasy Powersoil Kit (Qiagen) according to the manufacturer's instructions except that, prior to any processing, 200 μl of the solution in the Powerbead tube was replaced with 200 μl of phenol to chloroform to isoamyl alcohol (25:24:1, Thermo Fisher).

Quantitative PCR
qPCR was used to quantify levels of carbapenemase producing clinical strains and xylR mutants. 2 μl of the DNA that was extracted from stool was used in 10 μl reactions with Powerup SYBR Green PCR master mix (Applied Biosystems) in 384-well plates. Carbapenemase producing clinical strains were detected by a pair of primers binding to bla KPC gene. Additionally, K. pneumoniae clinical strains were also detected by a Klebsiella-specific primer set that bind to a chromosomal gene at the Klebsiella capsular polysaccharide synthesis gene cluster. Primer sequences are provided in Additional file 6 a. qPCR reactions were run using a Quant Studio 6 Flex (Applied Biosystems). C T values from qPCR were converted to number of genomes using a standard curve made from dilutions of known DNA mass.

16S rRNA sequencing analysis
Taxonomic classification was performed on the first 200 bp (quality score >30) of read 1 using the R package dada2 v1.12.1 [59] against the Silva database (release 132) [60]. The dada2 output containing 1186 amplicon sequence variants (ASVs) was imported into the R package Phyloseq v1.28.0 [61] for downstream analysis. Alpha diversity (number of ASVs) was calculated using the esti-mate_richness function with option "Observed" in Phyloseq. Volcano plots were generated using the R package Enhanced volcano [62]. Differential abundance testing between timepoints was performed on relative abundances using paired Wilcoxon signed-rank test with Benjamini-Hochberg correction.

Shotgun metagenomics community composition analysis
Metagenomic reads were taxonomically classified using Kraken2 [35] with option "--confidence 0.1. " Taxonomic classification was performed using either the built-in standard RefSeq database or a custom GTDB database (release 95) [36] built using the Struo workflow [66]. A total of 27,130 GTDB bacterial and archaeal genomes from a total of 194,600 were selected to build the database using the following criteria: > 50% completeness, < 5% contamination, and being a representative genome. No fungi, viral, or protozoa genome were included. Taxonomy was further speciated using Bracken [67]. In total, 706 species were classified. Bracken output was converted into the BIOM format using Krakenbiom (https:// github. com/ smdab doub/ kraken-biom) and imported into Phyloseq for downstream analyses. To generate Bray-Curtis PCoA plots, ASV abundances were variance-stabilized using Hellinger transformation via the "decostand" command in the R package vegan [68] and plotted using the "ordinate" command in Phyloseq. Differential abundance testing between unpaired samples was performed in MaAsLin2 [69] using the linear model method with Benjamini-Hochberg correction.

Shotgun metagenomics functional analysis
Metagenomic datasets were assembled using SPAdes v3.14.0 [70] with the option --meta. Protein-coding sequences (CDS) for assembled contigs were predicted and annotated with Prokka v1.14.0 [71] with the option --metagenome. Protein clustering was performed using the "linclust" function in MMseqs2 [72] with options "--cov-mode 1 -c 0.8" (minimum coverage threshold of 80% the length of the shortest sequence) and "--kmerper-seq 80. " The "--min-seq-id" option was set at 0.95 to generate the catalogues at 95% protein identity. A total of 467,873 proteins were obtained in the catalog. Functional characterization of all protein sequences was performed with eggNOG-mapper [73]. The clustered protein sequences were converted to a DIAMOND database [74]. The metagenomic reads were aligned against the protein catalogue using DIAMOND blastx with the options --id 90, --evalue 1e-6, -k 1, --maxhsps 1, and --unal 0. Proteins were filtered to remove low-abundance counts and converted to relative abundance by normalizing to total reads. Machine learning of proteins was performed using SIAMCAT v1.10.0 [75] with leave-one-out cross-validation (i.e., 28-fold crossvalidation for n=28 samples) and the "lasso" method. Robustness refers to the proportion of machine learning models that have incorporated the feature weights. Heatmap was plotted after converting relative abundance to z-score.

Whole-genome and metagenomic sequencing of evolved strains
The evolution strategy described in Fig. 5a produced 3 evolved K. pneumoniae isolates from each of 6 independent lineages. DNA was extracted from fresh cultures of these 18 strains on an automated platform (Promega Maxwell) as previously described [23]. To examine K. pneumoniae evolution on a metagenomic level, DNA was also extracted from pooled K. pneumoniae colonies cultured directly from the stool of each mouse lineage. Culture plates with 100-150 colonies were swiped together (plate swipes) and stored in 15% glycerol at −80°C until DNA extraction. Illumina libraries were created using Nextera library chemistry. Sequencing was performed on an Illumina MiSeq at the NIH Intramural Sequencing Center.

SNP calling
SNP calling for individual K. pneumoniae isolates was performed by aligning the whole-genome sequencing reads for these isolates to the KPNIH1 reference genome and calling SNPs using SNIPPY v3.2 with default settings (https:// github. com/ tseem ann/ snippy). Identified SNPs were validated by Sanger sequencing. In addition to the SNPs reported in Additional file 5 a, there were 3 SNPs inherent to MKP103 that presumably occurred during the growth and selection for deletion of the bla KPC gene from KPNIH1. These 3 SNPs, which were present in all evolved strains, included a missense SNP in KPNIH1_08215 (hypothetical protein), a synonymous SNP in KPNIH1_09140 (lipid transporter ATP-binding/ permease), and a synonymous SNP in KPNIH1_09495 (gamma-glutamylputrescine synthetase). SNPs were named according to the Human Genome Variation Society (HGVS) nomenclature [76].
For metagenomic SNP calling, allelic frequencies of SNPs from the sequencing of plate swipes were determined using bam-readcount (https:// github. com/ genome/ bam-readc ount). Briefly, reads were aligned to the KPNIH1 reference genome using bowtie2 [77], and the resulting sam file was converted into a bam file, sorted, indexed, and input into bam-readcount. Additionally, the known SNP locations for each lineage that had been observed using SNIPPY for isolate genomes were also input into bam-readcount to extract allelic frequencies for these SNPs in each plate swipe. Sequencing depth of the 5 xylR SNP locations was between 209 and 293 reads.

Statistics
For normally distributed data, a two-tailed t-test was used to compare 2 groups, whereas multiple groups were compared using one-way ANOVA followed by a post-hoc Dunnett's test to compare treatment groups against a control group. For non-normally distributed data, an unpaired Mann-Whitney U test or paired Wilcoxon signed-rank test was used to compare 2 groups, whereas multiple groups were compared using the Kruskal-Wallis H test followed by pairwise Mann-Whitney U tests to compare each treatment group against a control group. Correction for multiple testing was performed using the Benjamini-Hochberg method when the number of tests exceeded 20; otherwise, Bonferroni correction was used. Significance for categorical variables (e.g., outgrowth versus no outgrowth) was determined using Fisher's exact test with Hommel correction for multiple testing against a control group. PERMANOVA was calculated using the Adonis function in the R package vegan. Normality was assessed using the Shapiro-Wilk test. All statistical analyses were performed in R.