Proximity-ligation metagenomics reveals differential plasmid and chromosomal antimicrobial resistance gene carriage in disrupted gut ecosystems

Summary

Distinct ecological pressures shape accumulation of antimicrobial resistance genes (ARGs) and virulence genes in the gut microbiome. In this study, we used proximity ligation shotgun metagenomics to characterize bacterial host-mobilome relationships associated with antimicrobial resistance and virulence genes in two disease cohorts with microbiome dysbiosis: recurrent Clostridioides difficile infection (rCDI) and cirrhosis. Microbiome dysbiosis in patients with rCDI is primarily driven by prolonged antibiotic exposure, whereas dysbiosis in patients with cirrhosis results from altered gut physiology. We found an increased relative abundance of chromosomally linked antibiotic resistance determinants in both disease cohorts compared with healthy controls. The rCDI cohorts additionally exhibited increased relative abundance of plasmid-mediated ARGs.

Graphical abstract

Highlights

• •Both rCDI and cirrhosis show increased chromosomal Enterobacteriaceae ARGs vs. controls
• •Plasmid-borne Enterobacteriaceae ARGs were higher only in rCDI, not cirrhosis
• •Vancomycin ARGs were found in both cohorts despite no vancomycin use in cirrhosis
• •ARG/VF burden was not linked with clinical history, rifaximin or other antibiotics

Genomics; Microbiome; Systems biology

Introduction

The gastrointestinal microbiota plays a critical role in regulating host physiology and conferring colonization resistance against pathogenic microorganisms.^1,2 Simultaneously, the gut microbiome serves as a major reservoir of antimicrobial resistance genes (ARGs) and virulence factors (VFs), which may be transferred between strains through horizontal gene transfer (HGT).^3,4,5 Ecological conditions within the gut, including high bacterial densities, spatial organization, and exposure to environmental stressors such as antibiotics, create a highly conducive environment for HGT.^4,6,7 As a result, the gut microbiome serves as a central arena for microbial evolution, where HGT drives rapid emergence of drug-resistant opportunistic pathogens that contribute to a growing threat to global public health.⁸

Members of the Enterobacteriaceae and Enterococcaceae families frequently cause infections in medically complex patients and are increasingly associated with rising rates of antimicrobial resistance.^8,9,10 Both cirrhosis and recurrent Clostridioides difficile are major conditions associated with gut microbial dysbiosis,^11,12,13 which can drive the expansion of these opportunistic pathogens and ARGs.^14,15 The underlying cause of dysbiosis between these conditions differs as rCDI microbiomes experience prolonged antibiotic exposure and inflammatory perturbations, whereas cirrhosis is characterized by liver and immune dysfunction, altered gut physiology, and impaired barrier integrity, typically before extensive antibiotic use.^9,16,17 Comparison of these two distinct disease cohorts provides an opportunity to determine the impact of antibiotic-driven versus physiological-driven ecological disruptions on the accumulation of ARGs and VFs in Enterobacteriaceae and Enterococcaceae.

While typical shotgun metagenomic sequencing can identify ARGs across complex microbial communities, it generally lacks the ability to link extrachromosomal DNA to specific microbial genomes.^18,19 This can limit our understanding of these interactions and designing precision therapies to reduce pathogen burden. A powerful approach to overcome this limitation is proximity ligation (also known as Hi-C) shotgun metagenomics (PLSM). This assay cross-links DNA molecules co-localized within cells via fixation, thus enabling identification of mobile genetic elements (MGEs) associated with host genomes.²⁰ PLSM can therefore be applied to provide strain-level resolution of mobile ARGs and VFs in complex microbial communities. This is especially important in the study of Enterobacteriaceae and Enterococcaceae, as the spread of plasmid-borne extended spectrum β-lactamases (ESBLs) and vancomycin resistance genes (VRGs) is of urgent clinical relevance.²¹

In this study, we applied PLSM to study the fecal microbiomes from healthy individuals and two disease states associated with contrasting disrupted gut ecosystems: recurrent Clostridioides difficile infection (rCDI) and cirrhosis. We identified that both rCDI and cirrhosis patient microbiomes are associated with an increase in chromosomal Enterobacteriaceae ARGs compared to healthy controls, while plasmid-linked Enterobacteriaceae ARGs were associated primarily with rCDI microbiomes. We also show that a subset of patients in both cohorts harbored VRGs, despite cirrhosis patients having no known long-term exposure to vancomycin. We observed limited impact of clinical data and drugs such as rifaximin in defining the ARG and VF burden, collectively highlighting the complexity of pathobiome dynamics in the human gut.

Results

Study participants

The rCDI cohort consisted of 34 patients (Table S1). Most patients were older (mean age ± SD, 64.1 ± 16.4 years) and female (82.3%). The mean duration of their treatment with cycles of anti-CDI antibiotics was 6.9 ± 3.8 months. The most common inciting antibiotics preceding CDI were administered in the context of surgeries, followed by urinary tract infections, diverticulitis, skin or wound infections, dental procedures, sinusitis, pneumonias, and other. In some of the analyses, we stratified the rCDI cohort into three levels of healthcare burden based on the duration of antibiotic treatments, the number of different classes of antibiotics, and intensity of care (e.g., prolonged hospitalization, intensive care unit).

The liver disease cohort included five patients with compensated cirrhosis and 14 patients with decompensated cirrhosis (Table S2). The majority were male (84.2%) and 65.3 ± 10.4 years of age. None had yet experienced extensive use of antibiotics for complications of liver disease or other indications. However, half of the patients with decompensated cirrhosis were on rifaximin for prophylaxis of hepatic encephalopathy.

Cirrhosis and rCDI patients have distinct bacterial communities compared to healthy controls

PLSM sequencing was performed on a total of 66 stool samples (19 cirrhosis, 34 rCDI, and 13 healthy individuals) generating an average of ≈61 million reads per sample (range 18.6–143.0 million reads) (Table S3). Sequence reads were assembled into a total of 2070 bacterial MAGs (metagenome-assembled genomes) with ≥70% completeness and ≤5% contamination (Table S4). As expected, healthy control individuals exhibited microbiomes dominated by reads assigned to obligate anaerobic Clostridia, including Lachnospirales (29.08%, average relative abundance) and Oscillospirales (25.88%) (Figure 1A). In contrast, both rCDI and decompensated cirrhosis microbiomes show an increased proportion of reads mapped to Enterobacteriaceae (7.59% and 9.22%, respectively), although the abundance varied greatly between samples. To further assess differences in bacterial communities across treatment groups, we performed differential abundance testing using ANCOM-BC on aggregated family-level counts with healthy individuals serving as the baseline. The differential abundance analysis indicated a significant enrichment in Aristaeellaceae, CAG-508 (Clostridia), and UBA1381 (Clostridia) MAGs compared to decompensated liver and rCDI samples (p < 0.05, Benjamini-Hochberg correction) (Figures 1B and 1C). In contrast, we observed a significant enrichment of Enterobacteriaceae and Enterococcaceae MAGs in both decompensated cirrhosis and rCDI patients compared to the healthy controls. Community disruption was most pronounced in rCDI samples, which were enriched in MAGs absent from healthy individuals, including Veillonellaceae and Fusobacteriaceae.

Figure 1.

Differences in bacterial community composition

(A) Relative abundance of bacterial taxa in fecal samples at the order level. For each sample, bars represent the proportion of reads contributed by high-quality bacterial MAGs (≥70% completeness, <5% contamination), colored by taxonomic order. Bacterial classes with less than 8% average relative abundance across all samples are shown in gray. Sample types are marked above each bar as follows: green (healthy), yellow (compensated liver cirrhosis), dark pink (decompensated liver cirrhosis), and blue (rCDI).

(B and C) Differential abundance analysis showing bacterial families significantly associated with decompensated cirrhosis (B) and rCDI (C) compared to healthy controls (ANCOM-BC, BH-corrected p value < 0.05). The log-fold change (lfc) indicates the magnitude of the effect size, with positive values corresponding to higher abundance in cirrhosis or rCDI patient fecal samples.

Plasmid composition parallels changes in the bacterial community structure in rCDI and cirrhosis

We next evaluated the mobilome structure across treatment groups, and assembled sequence reads into a total of 705 plasmid and 2521 phage contig bins with ≥50% completeness and ≤5% contamination (Tables S5 and S6). Analysis of read counts assigned to different sequence types (genomic/chromosomal, plasmid, phage, unbinned) showed that the average proportion of reads assigned to plasmid contigs was greater in both rCDI and decompensated cirrhosis patient groups (2.86% and 2.22%, respectively) compared to compensated cirrhosis patients or healthy controls (0.74% and 0.32%, respectively) (Figure S1).

Given the importance of plasmid-borne genes in the carriage of ARGs, we examined the family-level taxonomic composition of plasmid-linked reads across all samples. On average, Enterobacteriaceae plasmid genomes accounted for a greater relative abundance of mapped plasmid reads in rCDI patients (56.37%) and cirrhosis patients (35.28% compensated and 28.00% decompensated, respectively) compared to healthy individuals (18.73%) (Figure 2A). We observed substantial variability between individuals, which included several rCDI/cirrhosis patients with no detectable Enterobacteriaceae plasmid reads and one healthy individual with a very high abundance of Enterobacteriaceae plasmid reads. Enterococcaceae plasmid reads were also detected in two decompensated cirrhosis and three rCDI samples, while no healthy controls carried detectable levels of Enterococcaceae plasmids. To further analyze potential differences in relative plasmid loads, we examined the TPM (transcripts per million) abundance of high-quality plasmid contig bins (normalized by the TPM of across all contig bins). Total plasmid TPM differences did not reach statistical significance after correction (Figure 2B). rCDI patients showed significantly lower loads of Bacteroidaceae plasmids (Figure 2C) and higher loads of Enterobacteriaceae plasmids (Figure 2D) compared to healthy controls and cirrhosis patients (Wilcoxon rank-sum test and Benjamini-Hochberg correction).

Figure 2.

Differences in plasmid composition and abundance

(A) Bars represent the proportion of total reads contributed by plasmid contig bins (≥50% completeness, <5% contamination), colored by host taxonomic family. Plasmids with mixed host sources (multiple families) were grouped together as mixed. Other represents host families with less than 10% average relative abundance.

(B) TPM normalized abundance of plasmids reads in each sample.

(C) TPM normalized abundance of Bacteroidaceae plasmids in each sample.

(D) TPM normalized abundance of Enterobacteriaceae plasmids in each sample. Statistical differences between healthy control, decompensated cirrhosis, and rCDI samples were determined via Wilcoxon rank-sum test (BH-adjusted p values).

Differential patterns of ARG and VF acquisition by Enterobacteriaceae in liver cirrhosis and rCDI

Identification of both ARGs and VFs was determined through mapping metagenomic reads against the NCBI AMRFinderPlus database. Linkage of ARGs to bacterial host genomes revealed substantial ARG burdens across several bacterial families, including Enterobacteriaceae, Lachnospiraceae, Bacteroidaceae, and Enterococcaceae (Table S7). Enterobacteriaceae encoded the highest overall abundance of ARGs and were the primary carriers of genes related to β-lactam, efflux pump, and fosfomycin resistance mechanisms (Figure 3A).

Figure 3.

Distribution and abundance of Enterobacteriaceae ARGs and VFs

(A) Relative abundance of ARG classes encoded by Enterobacteriaceae contigs.

(B) TPM-normalized abundance of chromosome-linked Enterobacteriaceae ARGs across sample types.

(C) TPM-normalized abundances of plasmid-borne Enterobacteriaceae ARGs across sample types. Statistical differences between healthy control, decompensated cirrhosis, and rCDI samples were determined via Wilcoxon rank-sum test (BH-adjusted p values).

(D) Scatterplot showing the relationship between ARG and VF abundance across samples.

TPM quantification of Enterobacteriaceae ARGs indicated that chromosomally linked ARGs were significantly more abundant in both rCDI and decompensated cirrhosis patients compared to healthy controls (Figure 3B) (p < 0.05, Wilcoxon rank-sum test, Benjamini-Hochberg correction). The most common chromosomal Enterobacteriaceae ARGs identified were efflux pumps, including emrD (n = 46 samples), oqxA (n = 33), and oqxB (n = 31). β-lactamases were the most diverse group of chromosomally encoded ARGs, with the most prevalent families being blaEC, blaOXY, blaSHV, and blaCMY (Figure S2). The rCDI patient group additionally showed significantly greater abundance of plasmid-borne ARGs compared to healthy controls and decompensated cirrhosis patients (Figure 3C) (p = 8.4e-4 and p = 0.011, Wilcoxon rank-sum test, Benjamini-Hochberg correction). Across all samples, we identified that the Enterobacteriaceae ARG abundance correlated strongly with VF abundance (p < 2.2e-16, ρ = 0.81, Spearman rank correlation test), with rCDI patients and cirrhosis patients generally exhibiting higher loads of both VFs and ARGs than healthy individuals. (Figure 3D).

PLSM reveals Enterobacteriaceae MAGs encoding plasmid-borne β-lactamases and pathotype-associated VFs

We identified several notable plasmid-borne β-lactamase genes—including blaTEM-1, blaOXA-1, blaSHV-27, and blaCTX-M—in rCDI samples; however, no plasmid-borne β-lactamase genes (adjusted coverage >90%) were detected in any compensated cirrhosis, decompensated cirrhosis, or healthy patient samples (Figure 4).The plasmid-borne β-lactamases in rCDI patients were found on MAGs classified as Morganella (n = 1), Klebsiella (n = 3), and Escherichia (n = 10), which often co-encoded other classes of ARGs on plasmids (Figure 5). These Escherichia MAGs often harbored multiple types of VFs, including plasmid-borne siderophore systems such as salmochelin (iroBCDEN), aerobactin (iucABCD, iutA), and yersiniabactin (ybtPQ) genes as well as plasmid-borne colicin genes (cvaC). Adhesion factors were also identified across these strains, including P fimbriae (pap genes), S fimbriae (sfaFS), and long polar fimbriae (lpfA). Acquired VFs were absent from the Klebsiella and Morganella MAGs encoding plasmid-borne β-lactamases. Consistent with this, we observed a positive correlation between the number of encoded ARGs and VFs across Escherichia MAGs (Spearman rank correlation test, p < 0.0002, ρ = 0.57), whereas no statistically significant correlation was detected for Klebsiella MAGs (p = 0.59, ρ = 0.16) (Figure S3). VFs were rare across K. pneumoniae strains that did not encode plasmid-borne β-lactamases; however, we notably identified a single K. pneumoniae MAG harboring plasmid-borne siderophore genes (iroB and iroD) and mucoid phenotype genes (rmpC and rmpD) which was present in a rCDI patient (CD23) who also carried an E. coli MAG encoding plasmid-linked β-lactamase genes(Tables S7 and S8).

Figure 4.

Heatmap showing the abundance of Enterobacteriaceae plasmid-borne ARGs across samples

Only samples with at least one Enterobacteriaceae plasmid-borne ARG are shown. The black dots represent samples where a given ARG was identified (>90% adjusted coverage), and box color intensity reflects the TPM-normalized abundance of the contig the ARG was found on. The bar plot indicates the taxonomic distribution of genomes each ARG was linked to.

Figure 5.

ARG and VF distribution of Enterobacteriaceae MAGs with plasmid-borne β-lactamases

Heatmap showing the presence of different plasmid-encoded antimicrobial resistance genes (ARGs) and virulence factors (VFs) present in Enterobacteriaceae MAGs that harbor at least one plasmid-borne β-lactamase. VFs were grouped by functional category, with the number in each box indicates the number of genes detected. Boxes without a number represent a single gene. Color intensity reflects the percentage of VFs located on plasmid contigs. Asterisk next to genome bin (∗) indicates Klebsiella strains not classified as K. pneumoniae or Escherichia strains not classified as E. coli.

rCDI and cirrhosis patient gut microbiomes may serve as reservoirs of VRGs

VRGs were detected in a subset of both cirrhosis (n = 3 compensated, n = 9 decompensated) and rCDI (n = 15) patients but were absent across all healthy control samples. Linkage analysis to microbial genomes further revealed that the presence of 9 Enterococcus MAGs encoding VRGs, which were primarily found on chromosomal elements (Figure 6A). The remaining chromosomal VRGs were distributed across multiple taxonomic groups, with Enterocloster (Lachnospiraceae) as the most abundant carrier (Figure 6B). Notably, we also identified a single C. difficile MAG in an rCDI sample (CD28) with chromosomally encoded VanG, VanR, VanS, VanT, VanZ1 genes, potentially enabling survival during ongoing vancomycin therapy.

Figure 6.

Distribution of vancomycin resistance genes in rCDI and cirrhosis patients

(A) Distribution of Enterococcus MAGs encoding VRGs. Color intensity reflects the contig-level abundance of the VRG. Boxes marked with “P” indicate VRGs encoded on plasmid sequences.

(B) Bar plot showing the TPM normalized abundances of chromosomal and plasmid-linked contigs encoding VRGs, colored by bacterial taxonomy. Enterococcus and Clostridioides are shown at the genus level; all other bacteria are shown at the family level.

No detectable association between clinical history and ARG/VF burden within study cohorts

Despite enrichment of ARGs and VFs in disease cohorts, we observed no consistent relationship between resistance burden and conventional healthcare exposure metrics or patient demographics (Figure S4). Stratification of the rCDI cohort by cumulative anti-C. difficile antibiotic treatments, length and intensity of hospital care, and patient age failed to correlate with ARG/VF burden. Similarly, no correlation was observed between ARG/VF burden and age in cirrhosis patients. Rifaximin use, which has been linked to increased resistance to the last-resort antibiotic daptomycin,²² was not significantly associated with increased Enterobacteriaceae ARG or VF abundance in cirrhosis patients (Figure S5). When we compared the 10 rifaximin users (3 rCDI+7 decompensated cirrhosis) to non-users (decompensated and rCDI), we found no statistical differences in Enterobacteriaceae VF, Enterobacteriaceae ARG, and VRG abundance across samples.

Discussion

In this study, we used PLSM to determine strain-resolved differences in the accumulation of ARGs and VFs in two cohorts of disease patients with disrupted gut ecosystems. The use of proximity ligation before performing high-depth sequencing of fecal samples offers several advantages over typical shotgun metagenomics, including improved MAG reconstruction and the ability to link ARG carriage to specific microbial genomes or mobile elements. This approach is particularly valuable for understanding antimicrobial resistance dynamics in Enterobacteriaceae, as many ARGs are carried on conjugative plasmids capable of transferring across different bacterial genera.²³ Using this approach, we were able to identify differential changes in the gut microbiome and mobilome structure across patients with rCDI or cirrhosis.

We chose rCDI and cirrhosis as two major diseases where microbial alterations, especially related to resistance and virulence, can affect outcomes. We first identified significant changes in the abundance of key bacterial taxa associated with the gut microbiome of rCDI and cirrhosis patients. Consistent with previous studies of rCDI and cirrhosis fecal microbiomes,^11,24,25,26 we observed significantly higher normalized abundances of Enterobacteriaceae and Enterococcaceae bacteria, which are strongly associated with infections within healthcare settings.^2,27 Notably, enriched taxa in rCDI patients included Fusobacteriaceae and Veillonellaceae, which have previously been linked to increased intestinal inflammation and the promotion of C. difficile colonization.^28,29,30 While common commensal-associated taxa (e.g., Clostridia bacteria) were depleted in both conditions compared to healthy controls, there appeared to be larger differences in the microbiome structure of rCDI patients. This finding is likely explained by the substantial exposure to vancomycin among this patient group, which results in a strong disruption in gut microbial communities,³¹ and enrolling patients with cirrhosis not on antibiotics beyond rifaximin.

The next goal of our study was to characterize microbial host carriage of ARGs across sample types to identify potential reservoirs of clinically relevant genes. We found that both rCDI and cirrhosis patients exhibited a greater load of chromosomal Enterobacteriaceae ARGs compared to healthy individuals. This was not surprising given the general increase in Enterobacteriaceae, some of which naturally carry chromosomal variants of β-lactamases.³² Interestingly, we observed an increase in plasmid-borne ARGs exclusively in rCDI patients, suggesting that the gut ecological conditions associated with rCDI may create a permissive environment for HGT within Enterobacteriaceae lineages. This trend may be driven by a combination of pre-CDI antibiotic use, increased Enterobacteriaceae abundance, and inflammation experienced during rCDI infections, which are conditions linked to increased HGT rates in the gut.^15,33

Among Enterobacteriaceae plasmid-borne ARGs, we identified several highly clinically relevant β-lactamases in rCDI patient feces, including bla-CTX-M, bla-TEM-1, and bla-SHV-27 genes. Of these, the CTX-M enzymes are of particular concern, as most variants are ESBLs associated with high levels of resistance to critically important third-generation cephalosporins.³⁴ However, the most common plasmid-borne β-lactamase was TEM-1, which is associated with resistance to penicillin (including amoxicillin) and early-generation cephalosporins.³⁴ Enterobacteriaceae plasmids carrying narrow- or extended-spectrum β-lactamases can co-encode genes conferring resistance to other antibiotics including tetracyclines, aminoglycosides, sulfonamides, and fluoroquinolones.²³ Despite the use of PLSM, we were unable to construct these full Enterobacteriaceae plasmids, which can be hundreds of kilobases long.²³ Nevertheless, we still observed a pattern of Enterobacteriaceae MAGs encoding multiple classes of plasmid-linked resistance genes, although multidrug phenotypes of these strains cannot be determined from sequence data alone. The spread of drug-resistance to virulent strains of E. coli or K. pneumoniae represents a major public health challenge as it limits treatment options in vulnerable individuals. Our findings suggest that the accumulation of ARGs within E. coli is associated with an increase in virulence potential, a pattern which has been previously observed by other studies.^35,36 As a species, E. coli maintains a large open pangenome, including numerous VFs that allow for adaptation to diverse hosts and ecological niches.³⁷ Notably, we identified E. coli MAGs encoding plasmid-borne β-lactamases in rCDI patients, which also encoded plasmid-borne aerobactin, salmochelin, and yersiniabactin siderophore-mediated iron acquisition systems linked increased Enterobacteriaceae colonization of the mammalian gastrointestinal tract.^38,39 These Siderophore systems, together with P fimbriae and S fimbriae also detected in a subset of strains, are well-characterized VFs of extraintestinal pathogenic E. coli (ExPEC), which commonly cause bloodstream and urinary tract infections.^40,41

In contrast to Escherichia, we did not observe the same correlation between VF and ARG accumulation for Klebsiella. This may reflect shared core VFs among K. pneumoniae, including variable capsular polysaccharide (K antigen) and lipopolysaccharide (O antigen) loci, which aid in host immune evasion.^42,43 We identified several acquired VFs across K. pneumoniae strains; however, none of them also co-encoded plasmid-borne β-lactamases. We did identify a K. pneumoniae MAG in an rCDI patient encoding plasmid-borne salmochelin siderophore and mucoid phenotype genes, which have been previously linked to increased virulence potential in K. pneumoniae.^42,44

Despite our study indicating that VFs and ARGs are more abundant in rCDI patients, Enterobacteriaceae infections are known to pose a severe risk in patients with cirrhosis.^45,46 We did not enroll any patient with cirrhosis with recent or current absorbable antibiotics, which is different from the usual decompensated patients. Although our study included only a small sample of compensated cirrhosis patients, we still identified a visual progression in Enterobacteriaceae loads between compensated and more advanced decompensated cirrhosis patients. The major changes in microbiome structure and an increase in VFs/ARGs abundance in cirrhosis likely reflect ongoing changes in gut physiology and mucosal immunity, which impair the intestinal barrier and predispose patients to Gram-negative and -positive infections.^45,46 Future longitudinal studies on cirrhosis patient microbiomes may provide valuable insight into the dynamics of Enterobacteriaceae population changes in high-risk individuals.

The increase in VRG abundance observed in rCDI patients is likely primarily driven by the usage of vancomycin to treat C. difficile. Interestingly, we observed a single C. difficile MAG in an rCDI patient encoding VRGs. While the VanG cluster is usually only associated with low-level resistance,⁴⁷ it is possible that these genes conferred sufficient protection to allow survival despite continued vancomycin administration. Of great concern is the spread of plasmid-borne VRGs to Enterococcus species, which are an important cause of hospital-acquired infections.⁴⁸ Unexpectedly, we identified VRGs in a subset of cirrhosis patients despite no history of vancomycin usage. While two patients carried VRGs encoded on Enterococcus MAGs, the majority of the identified VRGs were linked to Lachnospiraceae (Enterocloster and Anaerostipes). Since none of the rifaximin non-users were ever on rifaximin previously, other absorbable longer-term antibiotic exposure could have affected this but the clinical significance of VRGs in Lachnospiraceae is unclear. Further experimental validation and longitudinal studies are needed to determine whether Lachnospiraceae can exhibit phenotypic vancomycin resistance or act as a potential reservoir of VRGs within the gut.

Despite the overall increase in VFs and ARGs in rCDI patients, we did not detect any associations with conventional healthcare metrics or patient demographics. While the rCDI patients received prolonged vancomycin treatment, this exposure alone does not fully account for the observed variation in pathobiome structure. The differential abundance of ARGs and VFs across patients likely reflects differences in microbial community structure, opportunistic ecologic vacancy, proximity to pathobionts in the environment, host immunity, and gut physiology. Detection of these complex interactions is limited by sample size and incomplete patient history. Within cirrhosis patients, we did identify a visual progression of Enterobacteriaceae loads between less severe compensated and decompensated cirrhosis patients, suggesting that disease state and increased dysbiosis plays an important role in pathogen accumulation. The use of rifaximin in cirrhosis patients has been previously linked to increased resistance to last resort drugs used to treat Enterococcus infections.²² In our dataset, we did not identify any increase in VRGs in rCDI and cirrhosis patients who were previously administered rifaximin, suggesting that increases in VRGs were not directly related to rifaximin prophylaxis within our cohorts.

In conclusion, we utilized PLSM to identify plasmid and chromosomal ARGs and VFs within two cohorts of patients with disease-associated gut dysbiosis. We notably identified that both rCDI and cirrhosis patients harbor a greater abundance of Enterobacteriaceae chromosomal ARGs compared to healthy controls, while rCDI patients also exhibit an increased relative abundance of plasmid-borne Enterobacteriaceae ARGs. Previously, we reported that microbiota transplant therapy (MTT) [also known as fecal microbiota transplantation] can lead to a significant reduction in ARG burden in rCDI and cirrhosis patients.^49,50 Expansion of these treatments may therefore be key to reducing patient mortality, and controlling the increasing global burden of antimicrobial resistant pathogens.

Limitations of the study

Although PLSM is a powerful tool for reconstruction of microbial communities and mobilome structure, we acknowledge several limitations of this study. Our analysis was based on the detection and assembly of high-quality MAGs, which likely do not represent all microbial strains within a complex community. Likewise, plasmid reconstruction from Hi-C linkage is providing a limited resolution of plasmid structure and host associations, in part due to the difficulty of assembling large plasmids in sequencing data.⁵¹ Metagenomics does not provide phenotypic data on any MAGs, and therefore we are unable to directly assess multidrug resistance or virulence phenotypes. Future studies should include a larger sample size of compensated cirrhosis patients and longitudinal sampling within individual patients, which would allow for finer scale detection of ARG/VF accumulation associated with host physiological changes.

Resource availability

Lead contact

Requests for further information should be directed toward Jasmohan S. Bajaj (jasmohan.bajaj@vcuhealth.org).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •Host-removed sequence reads are available on the NCBI Sequence Read Archive (SRA) under NCBI BioProject: PRJNA1438274.
• •R code to produce figures is available on a Zenodo repository (https://doi.org/10.5281/zenodo.21812578).

Acknowledgments

This work was partially funded by a gift from a non-profit Achieving Cures Together (A.J.S. and A.K.), VA Merit Review I01CX002472 and I01CX002683 (J.S.B.) and Richmond Institute for Veterans Research (J.S.B.). Sample processing and DNA sequencing at Phase Genomics were funded in part by the National Institute of Allergy and Infectious Diseases [R44AI172703] and the Gates Family Foundation.

Author contributions

A.K. and J.S.B. designed the study collection; A.J.S., A.K., and J.S.B. contributed to the design of study analysis; A.J.S. performed the primary analysis; A.J.S. wrote the original article draft; and all authors contributed to draft revisions and approved the submitted version of the manuscript.

Declaration of interests

B.A. is an employee of Phase Genomics.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Biological samples
Healthy Fecal Samples (UMN Cohort)	This study	HC01-HC04
Healthy Fecal Samples (VCU Cohort)	This study	HC05-HC13
Cirrhosis Patient Fecal Samples	This study	LC01-LC19
rCDI Patient Fecal Samples	This study	CD01-CD34
Chemicals, peptides, and recombinant proteins
DNA/RNA Preservative	Puritan	SKU#VL1201SAFE
Phase Genomics ProxiMeta Kit	Phase Genomics	v4.0
ZymoBIOMICS DNA Miniprep Kit	Zymo Research	Cat #D4300T
Watchmaker DNA Library Prep Kit	Watchmaker	N/A
Deposited data
Metagenomic Sequence Data	This Study	NCBI BioProject: PRJNA1438274
Code for Figure Generation	This Study	https://doi.org/10.5281/zenodo.21812578
Software and algorithms
R version 4.4.2	N/A	https://www.r-project.org/
fastp (v0.20.1)	Chen et al.52	https://github.com/opengene/fastp; RRID: SCR_016962
MEGAHIT (v1.2.9)	Li et al.53	https://github.com/voutcn/MEGAHIT; RRID: SCR_018551
BWA-MEM (v0.7.17)	Li & Durbin54	https://github.com/lh3/BWA; RRID: SCR_010910
SAMBLASTER (v0.1.24)	Faust & Hall55	https://github.com/GregoryFaust/samblaster; RRID: SCR_000468
SAMtools (v1.9)	Li et al.56	https://github.com/samtools/samtools; RRID: SCR_002105
ProxiMeta platform	Press et al.57	https://explore.phasegenomics.com/
CheckM (v1.1.3)	Parks et al.58	https://github.com/ecogenomics/checkm; RRID: SCR_016646
Mash (v 2.1)	Ondov et al.59	https://github.com/marbl/mash; RRID: SCR_019135
GTDB-Tk (v2.4.0)	Chaumeil et al.60	https://github.com/ecogenomics/gtdbtk; RRID: SCR_019136
VIBRANT (v1.2.1)	Kieft et al.61	https://github.com/AnantharamanLab/VIBRANT
AMRFinderPlus (v3.10.5)	Feldgarden et al.62	https://github.com/ncbi/amr
CheckV (v1.0.1)	Nayfach et al.63	https://bitbucket.org/berkeleylab/checkv/src/master/
ANCOM-BC (v2.8.1)	Lin & Peddada64	https://github.com/FrederickHuangLin/ANCOMBC
ggplot2 (v4.0.1)	Villanueva et al.65	https://github.com/tidyverse/ggplot2; RRID: SCR_014601
Rstatix (v0.7.2)	Kassambara et al.66	https://github.com/kassambara/rstatix; RRID: SCR_021240
Inkskape (v1.4.3)	N/A	https://gitlab.com/inkscape/inkscape
Other
Illumina NovaSeq X platform	N/A	N/A
Puritan HydraFlock Swab	Puritan	SKU#25-3706-H

Experimental model and study participant details

Study participants

After IRB approval (details below) and informed consent, we enrolled adult participants at 2 sites: Richmond VA Medical Center and University of Minnesota Medical Center. We included outpatients with cirrhosis (16 females and 3 males) and healthy controls (4 men) from Richmond VA Medical Center. Healthy controls did not have any chronic disease and were not on any prescription medications. Cirrhosis was diagnosed using liver biopsy, or history of or current decompensation (variceal bleeding, hepatic encephalopathy, ascites), or presence of varices in the setting of chronic liver disease, or transient elastography >12.5KPa. Patients unable to consent or provide stool samples, or with current or recent (<4 weeks) antibiotics (apart from rifaximin), were excluded. Patients were considered compensated if they had not experienced any prior decompensation, and decompensated patients had prior variceal bleeding, HE, or ascites. Half of the patients with cirrhosis had a recent (<1 month) outpatient procedure (6 upper endoscopy, 4 paracenteses, and 2 colonoscopies).

The rCDI patient cohort consisted of 34 patients (28 women and 6 men) from the University of Minnesota Medical Center. rCDI was defined as those with ≥2 CDI within 6 month without a new antibiotic provocation following the initial CDI episode. Healthy controls included nine Minnesota subjects (4 females and 5 males), including seven individuals participating in the FMT donor program. Details of rCDI and cirrhosis patient demographics are available in the supplemental materials (Tables S1 and S2).

Ultimately healthy controls were recruited from both sites and the ones from University of Minesota included healthy FMT donors. The influence of gender was not evaluated given the small number of study participants. Subjects were not allocated prospectively to any specific experimental group but rather evaluated on the basis of their disease and clinical history.

Ethical approval

The study protocol and activities was approved by the University of Minnesota Institutional Review Board (STUDY00003519) and Richmond VA Institutional Review Board (BAJAJ0033).

Method details

Sample collection and DNA extraction

Fecal samples from all patients were collected using Puritan HydraFlock swabs placed into 1 mL of the DNA/RNA preservative (Puritan Medical Products, Avantor, Radnor, PA). The samples were stored at -80°C until extraction at local laboratories before being sent to the Phase genomics laboratory.

Hi-C and shotgun library prep

Hi-C libraries were prepared using the Phase Genomics ProxiMeta Kit v4.0 according to the manufacturer protocol (Phase Genomics, Seattle, WA).⁶⁷ Briefly, ∼0.5 g of fecal material was crosslinked (15 min), quenched (20 min), and rinsed. The sample underwent bead-beating lysis for 20 minutes, followed by centrifugation to isolate and wash the nuclear pellet. After a 15-minute incubation at 65°C in lysis buffer, chromatin was recovered on magnetic beads, fragmented (1 hr at 37°C), and proximity-ligated (4 hr at room temperature). Following reverse-crosslinking at 65°C, Hi-C junctions were purified via streptavidin beads and prepared for PE150 sequencing on an Illumina NovaSeq X platform. For shotgun libraries, non-crosslinked DNA was extracted using the ZymoBIOMICS DNA Miniprep Kit and prepared with the Watchmaker DNA Library Prep Kit.

Metagenome assembly and deconvolution

Shotgun reads were trimmed and filtered using fastp (v0.20.1) using defaults for read quality filtering and adapter trimming before assembly with MEGAHIT (v1.2.9).^52,53 Hi-C reads were then aligned to the resulting contigs using BWA-MEM (v0.7.17) with the -5SP option.⁵⁴ PCR duplicates were removed via SAMBLASTER (v0.1.24) and non-primary alignments were filtered out using SAMtools (v1.9, command -F 2304); contigs under 1 kb were excluded.^55,56 The final assembly was deconvoluted into high-quality metagenome-assembled genomes (MAGs) using the ProxiMeta platform,⁵⁷ which clusters contigs based on both shotgun and Hi-C data. Finally, MAG quality was verified with CheckM (v1.1.3),⁵⁸ and taxonomy was assigned using Mash (v2.1) and GTDB-Tk (v2.4.0).^59,60

Viral annotation and binning

Viral contigs were identified using VIBRANT (v1.2.1).⁶¹ Contigs were classified as prophages if viral annotations exceeded 50% of the sequence length; otherwise, they were labeled microbial. To reconstruct viral metagenome-assembled genomes (vMAGs), ProxiPhage utilizes a hybrid strategy combining proximity ligation signal clustering with traditional binning. An overlap network was constructed to identify shared contigs, followed by a proprietary greedy network collapse algorithm to refine the clusters. Finally, each vMAG was adjusted based on Hi-C linkage strength to ensure high-quality, low-contamination clusters.⁶⁸

Plasmid identification and binning

To distinguish plasmid sequences within the assembly, all resulting contigs were subjected to a BLASTn search against a curated version of the PLSDB database (v2020_06_29). This comparative analysis focused on identifying homologous sequences by prioritizing the “best hit” for each query. Specifically, the evaluation of hit quality was restricted to the primary alignment, ensuring that noise from lower-confidence secondary matches did not influence the classification. The identification process utilized a multi-tiered filtering approach to ensure that only high-confidence plasmid candidates were retained. A minimum 90% nucleotide identity was required between the query contig and the reference sequence to establish a significant evolutionary or functional link. To eliminate false positives arising from short, conserved motifs or technical artifacts, a floor of 100 base pairs (bp) was set for all alignments. Further, a contig was definitively labeled a plasmid only if at least 50% of the query contig’s length aligned to the reference, and the alignment covered at least 50% of the total reference plasmid length. Hi-C data was also used to bin plasmid contigs into multi-contig plasmid bins (pMAGs). The same logic from the ProxiPhage algorithm was used for binning plasmid contigs.⁶⁸

AMR gene annotation

Antimicrobial resistance (AMR) genes were annotated from the metagenomic assembly using AMRFinderPlus (v3.10.5) with the --plus option enabled, and default parameters.⁶² AMR gene annotations were cross-referenced with MAG taxonomic assignments and the host-viral and host-plasmid association matrices generated by ProxiMeta to determine the origin of AMR genes (i.e., genomic, viral, or plasmid). A binary matrix was created to track the presence of AMR genes in each MAG and classify their origin accordingly.

Viral and plasmid MAG validation and host assignment

The quality of viral contigs and vMAGs was assessed using CheckV (v1.0.1).⁶³ For plasmids, completion and contamination values were estimated based on the percentage of aligned sequences and the total length of aligned contigs relative to the reference sequences. Both viral and plasmid host assignment were carried out using the ProxiPhage host attribution tool. Long-range Hi-C linkages between viral/plasmid contigs and their prokaryotic host genomes were analyzed to assign likely hosts. Viral-host and plasmid-host linkages were filtered to include hits with at least 2 Hi-C read links, a connectivity ratio > 0.1, and a copy number threshold. The copy number threshold was determined through a ROC curve of the fraction of all hits kept vs the fraction of viruses with one host, modified by the mobile element copy count per cell. The optimal cut-off was determined from the ROC curve as the value that produces the point to the top left of the plot, or the cut-off that removed the maximum number of virus-host links while still finding at least one host for the maximum number of viruses. Across ROC analyses, the mean area under the curve (AUC) was 0.83 (range 0.54-0.95) for plasmid-host linkages and 0.92 (range 0.78-0.97) for viral-host linkages. Connections with an average copy count less than 80% of the highest copy count were also removed to avoid false positives. Table S9 includes the plasmid–host assignments representing proximity-ligation links.

Analysis of identified MAGs, plasmids, and ARGs/VFs

Read counts of microbial MAGs with at least 70% completeness and less than 5% contamination were summed to estimate taxonomic composition of reads within each sample. Sample level relative abundance of genomic features was visualized using ggplot2 (v4.0.1) in R (version 4.4.2).⁶⁵ Differential abundance analysis was performed at the family level using ANCOM-BC (v2.8.1) (10% prevalence threshold with BH correction) on summed read counts of high-quality MAGs to estimate bias-corrected log-fold changes in bacterial family abundances between the healthy reference group and the rCDI and decompensated samples.⁶⁴ The log-fold changes of significant bacterial families identified by ANCOM-BC were visualized using ggplot2.⁶³

Read counts were summed for plasmid and viral MAGs with at least 50% completeness and less than 5% contamination to estimate the proportion of mobilome reads assigned to different bacterial taxa. Relative plasmid load within each sample was determined by summing the TPM of plasmid bins at the family level. Statistical differences between healthy, decompensated, and rCDI treatment groups were assessed using a Wilcoxon ranked sum test (‘Rstatix’ package, v0.7.2) with a Benjamini–Hochberg p-value correction to control the false discovery rate.⁶⁶

To identify ARG and VF load between samples, AMRFinderplus feature tables were generated from all contig assemblies and then filtered to remove general stress response genes and retain only hits with at least 90% adjusted coverage (defined as percent coverage against reference gene multiplied by the percent identity). Differences in ARG abundances between healthy, decompensated, and rCDI samples were assessed using Wilcoxon ranked sum test with a Benjamini–Hochberg p-value correction. Spearman rank correlations were used to assess relationships between VFs and ARGs.

Quantification and statistical analysis

Summary of statistical analysis

All statistical analysis is detailed in the above STAR Methods section. In brief, differential abundance testing was performed using ANCOM-BC and p-values were adjusted using Benjamini–Hochberg correction. Differences in contig abundance between treatment groups were determined using Wilcoxon ranked sum test with a Benjamini–Hochberg correction as implemented in Rstatix (v0.7.2). Compensated samples (n = 5) were not included in formal treatment level statistical comparisons.