Environmental reservoirs and transmission pathways of antimicrobial resistance across the pork production continuum

Abstract

Background

Antimicrobial resistance (AMR) is a major One Health challenge linking human, animal, and environmental health, yet the contribution of food production environments to resistance transmission remains poorly understood.

Results

We conducted a longitudinal shotgun metagenomic study across the pork production continuum from farm to retail to identify environmental AMR reservoirs and transmission pathways of antimicrobial resistance genes (ARGs). Assembly-based, genome-resolved, and source-tracking analyses were integrated to characterize resistomes, microbial communities, and horizontal gene transfer dynamics. ARG abundance and diversity were highest at farms, slaughterhouses, and processing plants and declined toward retail, although clinically relevant resistance determinants persisted throughout processing. Slaughterhouse environments emerged as major contributors to ARG contamination on carcasses, highlighting the importance of environmental exposure at intermediate stages. Resistome structure was closely linked to microbial community composition, with persistent taxa such as Acinetobacter and Pseudomonas serving as key ARG carriers, including genes conferring resistance to tetracycline, aminoglycosides, macrolide-lincosamide-streptogramin, and β-lactams, multidrug efflux. Co-localization of ARGs with mobile genetic elements demonstrated ongoing potential for horizontal transfer across production stages, and genome-resolved metagenome-assembled genome analyses revealed overlapping resistance and virulence profiles between slaughterhouse- and processing plant-associated bacteria, indicating adaptive persistence within pork production environments.

Conclusions

Resistome composition across the pork production chain is largely shaped by stage-specific environmental sources, highlighting potential intervention points to mitigate AMR transmission.

Download video file ^{(42.2MB, mp4)}

Video Abstract

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s40168-026-02444-3.

Introduction

Antimicrobial resistance (AMR) has emerged as one of the most critical threats to human health, contributing to an estimated 4.95 million deaths in 2019 [38]. While the improper and excessive use of antimicrobials in human and animal sectors accelerates the selection and dissemination of antimicrobial resistant bacteria (ARB) and antimicrobial resistance genes (ARGs), environmental reservoirs play an equally important, yet far less understood, role in shaping exposure pathways. Human healthcare environments are well-recognized hotspots where pathogenic ARB can persist, colonize surfaces, and facilitate healthcare-associated infections [3, 5, 16, 41]. In contrast, the contribution of food production environments to ARB emergence remains comparatively underexplored [1, 24, 45]. Despite growing concerns about the public health impact of pathogenic foodborne ARB, the diversity, distribution, and ecological drivers of ARB across food-animal processing systems are still poorly characterized.

Among food-animal systems, the pork production chain represents a critical model for examining the environmental contribution to AMR emergence. Pork is one of the most widely consumed animal proteins globally, exceeding 124 million tons in 2024 (OECD [40]), and intensive swine production practices often rely heavily on routine antimicrobial administration for disease prevention and treatment. Consequently, pork production systems have become major global hotspots for the emergence and amplification of AMR [63]. In 2017, antimicrobial use in swine was estimated at 193 mg per population correction unit (PCU), the highest use among food-animal species, compared with approximately 110 mg/PCU in chickens, and swine alone are projected to contribute nearly half of the global increase in antimicrobial consumption by 2030 [52]. Such antimicrobial practices promote the selection, persistence, and spread of ARB and ARGs within swine farm environments, many of which can persist through slaughter, processing, and distribution and ultimately reach consumers [14, 44, 47, 53]. Furthermore, the pork production chain functions as a complex microbial ecosystem where ARB originating from animals, workers, equipment, and the surrounding environment repeatedly interact and accelerate AMR emergence [49, 55, 64]. Despite rigorous hygiene programs, contamination and cross-contamination remain pervasive challenges, underscoring critical gaps in understanding how ARBs survive and spread within pork processing environment.

Given the complexity of microbial interactions and the continuous movement of ARB and ARGs along the pork production continuum, a comprehensive, high-resolution approach is required to characterize environmental AMR reservoirs and ARB transmission pathways. Metagenomic surveillance offers a powerful framework for simultaneously detecting ARBs, ARGs, and mobile genetic elements (MGEs) directly from environmental samples, enabling the assessment of resistomes across diverse niches [43, 56]. Unlike culture-based or targeted molecular assays, shotgun metagenomics captures both culturable and unculturable microorganisms and allows for the reconstruction of ecological and genetic networks that shape AMR occurrence and dissemination. However, despite its capacity to reveal transmission routes and high-risk hotspots, metagenomic surveillance has been underutilized in food production systems, particularly within commercial pork processing environments [7, 61]. To address this critical gap, we conducted a longitudinal, system-wide metagenomic investigation spanning farm, slaughter, and processing facilities, and retails to map environmental reservoirs of AMR and track the flow of ARGs and ARBs. This study provides a comprehensive understanding of how pork production environments contribute to AMR persistence and spread, offering essential insights for risk mitigation and safer food production.

Materials and methods

Sample collection

For shotgun metagenomic analysis, samples were collected from four key stages of the pork production continuum: pig farm, slaughterhouse, processing plant, and retail stores. Sample collection was approved and supported by facility managers at each site. Triplicate samples were collected from environmental sources (fan, floors, gloves, drains, feed, and water), pig-associated sources (feces, belly skin, and carcasses surfaces), and pork products (Fig. 1; Supplementary Table 1). In total, 96 samples were obtained between April 18 and April 21, 2021, spanning one domestic pig farm, slaughterhouse, processing plant, and three retail stores. All pig-derived samples were rigorously tracked from the farm through retail using the Korean animal product tracking system (animal product traceability, https://mtrace.go.kr/), ensuring precise linkage of samples throughout the production chain.

Fig. 1.

Sampling strategy and distribution for shotgun metagenomic sequencing. A farm-to-retail sampling strategy was used for shotgun metagenomic analysis across the pork production chain

At a pig farm in Pyeongtaek, Korea, three clinically healthy Landrace pigs, one of the predominant breeds in Korean pig production, were randomly selected from three different barns. Approximately 400 cm² of abdominal skin was swabbed from each pig. Environmental samples were collected by swabbing fans and floors in each barn. Additional samples included workers’ gloves, drain water, pig feces, feed consumed by the pigs, and drinking water provided to the animals. Fecal samples were collected from freshly excreted droppings, immediately after defecation. To minimize environmental contamination, only the inner portion of freshly excreted feces that did not contact with the floor was aseptically collected. Following transport of the pigs to a slaughterhouse in Hwaseong, sampling was conducted immediately after slaughter. Approximately 400 cm² of skin was swabbed from both unwashed and washed carcasses. Environmental samples were collected from the slaughterhouse floor and splitting saws. Fluid samples included drain water from three different locations, workers’ gloves, and groundwater used for washing carcasses, equipment, and the floors. After slaughter, pigs were transported to a processing plant in Hwaseong, where sampling was conducted following meat processing, including deboning and packaging. Unprocessed carcasses and processed products, including belly and neck meats, from three different pigs were randomly selected, and approximately 400 cm² of surface area was swabbed. Environmental samples included swabs from the floor, knives used by three workers, and designated areas on worktables used during carcass portioning. Drain water from three sites and tap water used for cleaning equipment and floors were also collected. Final pork products, pork belly and pork neck, were purchased from three different retail stores in Hwaseong and Pyeongtaek. Approximately 100 cm² of surface area from each product was swabbed.

All swab samples, including pork and environmental samples, including fan, floor, gloves, saws, knives, and worktables, were collected by 3 M Dry-Sponge swabs (3 M, St. Paul, MN, USA) and placed into Whirl–Pak Speci-Sponge sampling bags (Nasco, Fort Atkinson, WI, USA). Feces, feed, and gloves were directly placed into Whirl–Pak Speci-Sponge sampling bags (Nasco, Fort Atkinson, WI, USA). Drinking water and wash water were collected in 2 L sterile sampling bottles (Mediland, Seoul, South Korea), while drain water was collected in sterile 50 mL conical centrifuge tubes (SPL Life Sciences, Pocheon, South Korea). At the pig farm, drain water was collected after sample collection followed by routine washing procedures, representing wastewater contaminated with fecal material from animal housing areas. In contrast, at the slaughterhouse and processing plant, drain water samples were collected from drainage systems after sample collection had been completed, representing wastewater present in facility drains during operation. All samples were placed in insulated containers on ice immediately after collection and transported to the laboratory within 4 h. Upon arrival, samples were immediately processed for downstream analyses without delay.

Sample preparation and DNA extraction

Swab samples were rinsed with 225 mL of buffered peptone water (BPW; Oxoid, Hampshire, UK). Feces and feed samples (25 g each) were suspended in 225 mL of BPW and homogenized using a Circulator 400 (Seward, Worthington, UK) at 230 rpm for 2 min. All homogenates and rinsates were transferred to 3MTM filter sample bags (3 M, St. Paul, MN, USA) to remove insoluble residues. The filtered suspensions were centrifuged at 7193 × g for 10 min, using an Eppendorf 5430R centrifuge (Eppendorf, Hamburg, Germany), and the resulting pellets were suspended in 5 mL of phosphate-buffered saline (PBS, Biosesang, Gyeonggi-do, South Korea). These concentrated samples were stored at −80 °C until further processing.

Genomic DNA was extracted from all samples using the DNeasy PowerSoil Pro Kit (QIAGEN, Düsseldorf, Germany) according to the manufacturer’s protocol. DNA concentration and quality were assessed using a fluorescence-based assay with a Tecan Spark microplate reader (Tecan Trading AG, Zurich, Switzerland) and by gel electrophoresis. Purified genomic DNA was stored at − 20 ℃ prior to shipment to Macrogen Inc. (Seoul, South Korea) for shotgun metagenomic sequencing.

Shotgun metagenomic sequencing

Because DNA yield and quantity varied across samples, 26 representative samples were selected for shotgun metagenomic sequencing (Supplementary Table 1). Sequencing libraries were prepared using the TruSeq Nano DNA High Throughput Library Prep Kit (Illumina, San Diego, CA, USA) according to the manufacturer’s protocol. Briefly, 100 ng of genomic DNA was fragmented using the adaptive focused acoustic method (Covaris), generating blunt-ended 5′ phosphorylated double-stranded DNA fragments. DNA fragments were size-selected using a bead-based method, followed by A-tailing and ligation of TruSeq DNA UD Indexing adapters (Illumina, San Diego, CA, USA). DNA libraries were then purified and enriched through PCR for sequencing using 8 amplification cycles, applied uniformly across all samples. Library concentrations were quantified using qPCR using the KAPA Library Quantification Kits for Illumina Sequencing platforms; library quality was assessed with the 4200 TapeStation System using D1000 ScreenTape (Agilent technologies, Santa Clara, CA, USA). Paired-end (2 × 150 bp) sequencing was conducted on the NovaSeq platform using an S4 flow cell, with all samples sequenced in a single multiplexed lane. Overall sequencing quality was high, with 96.89% of base calls having an average Phred quality score of Q30 or greater (Supplementary Table 2).

Microbial taxonomy profiling

Raw metagenomic reads were processed with fastp (v0.23.4) [18] to remove adapter sequences and trim low-quality bases. Bases with Phred scores below 15 were considered low quality and reads containing more than 40% such bases were discarded under default fastp settings. Reads shorter than 50 bp were also removed. Host-associated sequences were removed by aligning reads to the pig reference genome Sscrofa11.1 using Bowtie2 [31] with the --very-sensitive parameter, and only unmapped reads were retained for downstream analysis. Taxonomic profiling was performed using Kraken2 (v2.1.3) [36, 57] with the k2_standard_20240904 database, setting the minimum-hit-groups parameter to 3. Species-level abundance estimation was refined with Bracken (v2.9) [35], using a read length of 150 bp and requiring a minimum of 10 reads per taxon. To reduce noise, taxa with fewer than 10 total reads across all samples or those appearing in only a single sample were excluded. The resulting count matrix was normalized using Cumulative Sum Scaling (CSS) via the metagenomeSeq package, using a percentile threshold of 0.75, to account for differences in sequencing depth. Alpha diversity indices, including Shannon diversity and Observed Richness, were calculated from the normalized data using the vegan package. Differences in alpha diversity between processing stages were assessed using the Kruskal–Wallis test followed by Dunn’s post hoc test with Benjamini-Hochberg (BH) p-value adjustment. Beta diversity was visualized via Principal Coordinates Analysis (PCoA) based on Bray–Curtis distances, and statistical significance of microbial community shifts was determined using Permutational Multivariate Analysis of Variance (PERMANOVA) with 999 permutations.

Resistome profiling

Antimicrobial resistance genes (ARGs) were identified directly from quality-filtered metagenomic reads using DeepARG in short-read pipeline [4]. Paired-end reads were aligned to the DeepARG database using default parameters, including a minimum alignment identity of 80%, minimum prediction probability of 0.8, and an e-value threshold of 1 × 10^–10. Gene names were extracted from the best-hit annotations that were standardized to gene-family levels. To account for differences in sequencing depth and sparsity in metagenomic count data, ARG count tables were normalized using CSS (p = 0.75).

ARG alpha diversity was calculated at the gene level using richness, defined as the number of unique ARG genes detected per sample, and Shannon diversity index, accounting for both ARG richness and evenness, calculated from the relative abundance distribution of ARGs. Differences in alpha diversity among production stages were evaluated using the Kruskal–Wallis test, followed by pairwise Wilcoxon rank-sum exact tests with Benjamini–Hochberg correction for multiple comparisons. The resistome β-diversity was assessed using Bray–Curtis dissimilarity based on the sample-by-ARG relative abundance matrix. Principal coordinates analysis (PCoA) was performed to visualize differences in resistome structure among samples. Statistical significance of group-level differences was evaluated using PERMANOVA (adonis2, 999 permutations), followed by pairwise comparisons. The proportion of variance explained (R²) and corresponding p-values were reported on the plots.

ARG abundance was calculated as the number of reads assigned to each ARG, normalized by the total number of ARG-assigned reads within that sample. This normalization accounts for differences in sequencing depth and enables compositional comparisons across samples. ARGs were further grouped into functional classes (e.g., β-lactam, tetracycline, aminoglycoside, multidrug) based on DeepARG classifications. For visualization of dominant ARGs, the top 20 most abundant genes or ARG classes were selected based on total abundance across all samples.

Identification of ARG-associated bacterial hosts and mobile genetic elements (MGEs)

High-quality reads were assembled into contigs using MEGAHIT (v1.2.9) [33], applying a minimum contig length threshold of 500 bp. ARGs were identified with DeepARG LS mode with default parameters and a minimum probability cutoff of 0.8 [4]. To link ARGs with their potential bacterial hosts, contigs were taxonomically classified using Kraken2, and only contigs with both ARGs and taxa annotations were selected. These associations were interpreted as putative ARG-host linkages. To investigate the associations between MGEs and ARGs, MobileElementFinder (v1.1.2) was applied to screen other MGEs [27], including non-autonomous mobile genetic elements (MITEs), insertion sequences (ISs), composite transposons (ComTns), transposons (Tns), integrative and conjugative elements (ICEs), integrative mobilizable elements (IMEs), and cis-mobilizable elements. Parameters were set to --min-coverage 0.8, and only contigs > 1 kb in length were included in the analysis. An ARG was considered carried by a MGE if both were located on the same contig, separated by ≤ 500 bp, and without overlapping.

Resistome and microbiome associations

To assess correspondence between microbial community composition and resistome structure across samples, Procrustes analysis was conducted using beta-diversity ordination results derived from microbiome and resistome profiles. Microbiome composition was based on Bray–Curtis dissimilarity of taxonomic relative abundance, while resistome composition was based on Bray–Curtis dissimilarity of normalized ARG relative abundance. Procrustes analysis was applied to align the two ordination configurations using the PCoA site scores (first two axes) by minimizing the sum of squared distances between matched samples. Concordance between the microbiome and resistome ordinations was quantified using the Procrustes goodness-of-fit statistic (M²) and correlation coefficient (R). Statistical significance was evaluated using a permutation-based Procrustes test (999 permutations). Procrustes plots were used to visualize correspondence, with line segments connecting paired sample positions in the microbiome and resistome ordination spaces. All analyses were performed in R using the vegan package.

Resistome source tracking

To investigate potential source attribution patterns of ARGs across the pork production chain, FEAST source tracking analysis [51] was performed to estimate the relative contribution of upstream stages and stage-specific environmental reservoirs to downstream samples. In FEAST analysis, the ARG relative abundance tables were converted to integer pseudocount matrices by rounding to the nearest integer, generating FEAST-compatible input while preserving relative resistome structure. For each downstream stage following the pig farm (i.e., slaughterhouse, processing plant, and retail store), pig carcass or pork product samples were designated as sink samples, whereas all samples from upstream stages together with environmental sample types from the current stage were defined as potential sources. FEAST was applied independently for each downstream stage. Source contributions were summarized as proportional abundances for each sink group and visualized as pie charts using R. ARGs that could not be confidently assigned to any defined source were categorized as unknown sources. All FEAST analyses were performed using default parameters.

Metagenome-assembled genomes (MAGs) analysis

The depth and coverage of assembled contigs were estimated using SAMtools (v1.20) [21] and MetaBAT2 (v2.13) [28]. Metagenome-assembled genomes (MAGs) were reconstructed with MetaBAT2, applying a minimum contig length threshold of 2500 bp. Genome quality, defined by completeness and contamination, was assessed with CheckM2 (v1.0.2) [19], and only MAGs with higher than 50% completion and low contamination (< 1%) were retained for downstream analyses [11]. Species-level classification of over medium-quality MAGs and phylogenetic placement were performed using the GTDB-Tk (v2.3.2) classify workflow [15]. The resulting midpoint-rooted phylogenetic tree was visualized with iTOL (v7) (https://itol.embl.de). ARGs and virulence factors (VFs) of assembled genomes were identified with ABRicate (v1.0.1) [50], using the CARD [2] and VFDB [17] databases, respectively.

Statistical analysis and data visualization

All data analyses and visualizations were performed in R (v4.4.3). The packages tidyr (v1.3.1), vegan (v2.6-10), dplyr (v1.3.1), forcats (v1.0.0), tidyverse (v2.0.0), pairwiseAdonis (v0.4.1), and FSA (v0.10.0) were used for data handling and statistical analyses. Microbiome compositional dissimilarity across pork processing stages was assessed using pairwise PERMANOVA tests. Differences in Shannon diversity between processing stages and sample types were evaluated using the Kruskal–Wallis test, followed by Dunn’s pairwise post hoc comparisons. An adjusted p-value < 0.05 was considered statistically significant. For data visualization, the packages ggplot2 (v3.5.1), paletteer (v1.6.0), ggpubr (v0.6.0), ComplexHeatmap (v2.22.0), circlize (v0.4.16), networkD3 (v0.4.1), ggalluvial (v.0.12.5), gggenes (v.0.5.1), and pheatmap (v. 1.0.12) were used.

Results

Pork farm and slaughterhouse showed highest ARG abundance

To evaluate potential antimicrobial resistance (AMR) transmission along the pork production continuum, we profiled the resistome of all samples collected across the pork production continuum at four sequential stages, swine farm, slaughterhouse, processing plant, and retail stores. Figure 1 summarizes the systemic, longitudinal sampling strategies designed to characterize the dynamics of ARB and ARGs (Fig. 1). Samples were analyzed using shotgun metagenomics to capture dynamics in microbial and resistome profiles along the farm-to-retail pathway. The sample distribution and the number of samples used for shotgun metagenomic sequencing are shown in Supplementary Table 1. Sequencing depth per sample ranged from approximately 87 to 320 million reads, with a total of approximately 5.8 billion raw reads generated across all samples (Supplementary Table 2).

To characterize resistome structure across the pork production continuum, ARG profiles were quantified using a read-based approach. ARG richness, defined as the number of unique ARGs detected per sample, differed significantly across production stages (Kruskal–Wallis, p = 0.0031; Fig. 2A). Pairwise comparisons revealed a significant reduction in resistome complexity from pig farm to retail samples (adjusted p = 0.033; Supplementary Data 1), indicating a progressive decline toward the final production stage. In contrast, Shannon diversity did not differ across production stages (Kruskal–Wallis, p = 0.26; Fig. 2B), indicating that ARG diversity remained stable. However, overall resistome structure differed across stages, as shown by clear stage-associate separation in Principal coordinates analysis (PCoA) based on Bray–Curtis dissimilarity of ARG relative abundance profiles (PERMANOVA, R² = 0.473, p = 0.001; Fig. 2C), with pairwise comparisons confirming significant differences between stages (Supplementary Data 1). ARG richness and Shannon diversity were further evaluated across sample types within each production stage (Fig. 2D). Environmental samples, including drain water, fans, gloves, and floor, showed relatively consistent ARG richness from the pig farm to the slaughterhouse and processing plant, but a marked decrease was observed in pork products at the processing plant and retail stages. In contrast, Shannon diversity showed remained stable across sample types and production stages (Fig. 2D). These findings suggest that environments of pig farm, slaughterhouse, and processing plant may act as critical reservoirs of ARGs, with potential transmission to carcasses during slaughter and processing. ARG co-occurrence pattern highlighted additional stage-specific features of ARG transmission. Upstream samples, particularly from the slaughterhouse and processing plant, exhibited greater correlation networks, indicating the presence of interconnected ARG clusters that may facilitate ARG persistence and transfer (Supplementary Fig. 1). In contrast, these correlation structures weakened markedly and were largely absent in retail samples, suggesting a reduced potential ARG transmission at later stages of the production chain.

Fig. 2.

Resistome profiling across the pork processing chain. A Antimicrobial resistance gene (ARG) richness measured by unique ARG counts across stages of the pork processing chain. B ARG diversity measured by the Shannon diversity index across processing stages. A, B Points represent individual samples; boxplots summarize distributions within each stage. Statistical significance was assessed using the Kruskal–Wallis test, with p-values indicated on the plots. Differences in Shannon diversity were assessed using the Kruskal–Wallis rank-sum test followed by pairwise Wilcoxon rank-sum tests. C Bray–Curtis dissimilarity of ARG profiles across sample types along the processing stages. Statistical differences were evaluated using PERMANOVA. Colors represent different sample types, and shapes represent different processing stages. D ARG richness and Shannon index across different sample types within the pork processing chain. Bars represent the mean values of ARG richness and Shannon diversity calculated across samples within each sample type. Error bars indicate the standard deviation among samples. The relative abundance of ARG-associated drug classes across processing stages (E) and sample types (F). Unclassified refers to ARGs that could not be assigned to a specific resistance class within the DeepARG database. The relative abundance of individual ARGs across processing stages (G) and sample types (H), showing the top 20 most abundant ARGs. UC, unwashed carcass; WC, washed carcass; PC, processed carcass

ARG class profiles were broadly similar across the pork production chain, with multidrug resistance genes representing the dominant class at all stages (Fig. 2E, F; Supplementary data 1). Tetracycline, aminoglycoside and macrolide–lincosamide–streptogramin (MLS) classes were consistently detected at high proportions across sample types. Pig farm samples showed distinct ARG class profiles compared to other stages (Fig. 2E), with greater heterogenicity at the farm, whereas slaughterhouse and processing plant samples showed more similar ARG class profiles (Fig. 2F), indicating shifts in resistome composition beyond the farm stage.

Analysis of the top 20 individual ARGs revealed stage- and sample type-associated patterns (Fig. 2G, H; Supplementary Data 1). Slaughterhouse and processing plant samples showed similar ARG profiles, whereas pig farm samples harbored distinct ARG profiles. Aminoglycoside resistance genes aad(9) and aadE were enriched in pig farm environments (Fig. 2G), particularly in drain water, floor, and fecal-associated samples (Fig. 2H). The MLS resistance gene lnuA showed relatively high abundance in glove and pig-associated samples, suggesting potential handling-related transmission (Fig. 2H). In contrast, multidrug-associated genes such as rpoB2 were prominent in processing plant environments (e.g., floor and knives) and remained abundant in retail pork products, indicating persistence of specific ARG signatures along the production chain (Fig. 2H). Notably, differences in ARG composition were observed between pork belly and pork neck samples at retail (Fig. 2F, H), suggesting that distinct processing, handling, or packaging steps may contribute to differential ARG profiles in final products.

Stage-specific microbiota exists, but overall community structure persists across the pork processing chain

To explore if ARGs abundance and diversity were associated with microbial community in the pork production chain, we analyzed its profile based on normalized shotgun metagenomic reads. Overall bacterial diversity measured by the Shannon index at the species level, differed significantly across stages (Kruskal–Wallis, p = 0.0087; Fig. 3A). Pairwise tests confirmed that samples from the retail exhibited significantly lower diversity than those from the farm (adjusted p = 0.01; Fig. 3A). At the sample type level, environmental samples generally showed higher microbial diversity than pig- or pork-associated samples (Fig. 3B), suggesting their role as microbial reservoirs as well. An exception was unwashed carcasses, which retained the greatest diversity among slaughterhouse samples, suggesting potential contamination from animal and facility sources prior to washing. These results indicate that the processing environment harbors a broader microbial community with potential for transfer to product surfaces. Consistent with these patterns, Bray–Curtis dissimilarity analysis revealed significant differences in microbial community composition across production stages (PERMANOVA, R² = 0.306, p = 0.001; Fig. 3C). Pairwise comparisons (Supplementary data 2) further showed that all stage comparisons were significant (p ≤ 0.012), except between processing plant and the retail samples (p = 0.124), indicating partial compositional convergence at later stages.

Fig. 3.

Microbial community profiling and dynamics across the pork processing chain. A Alpha diversity of the microbiome measured by the Shannon index, illustrating changes in microbial richness and evenness across processing stages. Each dot represents an individual sample. Boxplots show the median (center line), interquartile range (box), and 1.5 × IQR (whiskers). Statistical significance was assessed using the Kruskal–Wallis test followed by Dunn’s post hoc test (Benjamini–Hochberg adjusted). B Alpha diversity across all sample types collected at each stage of the pork processing chain, highlighting stage-specific microbial community patterns. Bars represent the mean Shannon index for each sample type, error bars indicate the standard error of the mean, and overlaid points represent individual samples. C Beta diversity of the microbiome based on Bray–Curtis dissimilarity, showing compositional differences along the pork processing chain. Statistical significance was evaluated using PERMANOVA, followed by pairwise comparisons, with p values indicated in the figure. D Persistent bacterial species identified across the pork processing chain and E across different sample types, defined as taxa detected in all samples. F Bacterial species with relative abundance greater than 3% across different sample types and processing stages and G across different sample types. H Prevalence and I relative abundance of major foodborne pathogens, including members of the ESKAPE group, Salmonella enterica, and Escherichia coli, across the pork processing chain. WC, washed carcass; PC, processed carcass

Despite these compositional differences, a subset of taxa was consistently detected across stages. A total of 29 persistent bacterial species were identified out of 11,623 taxa (Fig. 3D, E; Supplementary Data 2), but the relative abundance varied substantially among stages and sample types. Among them, Acinetobacter baumannii, an opportunistic pathogen, was consistently present at all stages, with higher relative abundance in slaughterhouse samples. Lactobacillus amylovorus was particularly prevalent at the farm stage, especially in drain water, feces, floor swabs, and pig abdomens. A. johnsonii dominated the slaughterhouse environment, while Kocuria varians increased in abundance during processing and retail, particularly on floors, knives, pork necks, and unprocessed carcasses. We next defined dominant taxa at each stage, as species comprising more than 3% relative abundance in any samples, to identify stage-associated microorganisms. Based on this criterion, 54 dominant species were identified (Supplementary Data 2), showing stage- and sample-type specificity (Fig. 3F, G). At the farm, L. amylovorus, also identified as a persistent taxon, prevailed in feces, floor, and abdomen samples. In the slaughterhouse, microbial composition was homogeneous across sample types, indicating extensive microbial exchange within the facility in which A. johnsonii was the most dominant taxon. However, the processing plant displayed a more heterogeneous but interconnected pattern, suggesting the introduction of microbes from multiple sources. At the retail stage, distinct dominant microbial communities were observed, indicating product-specific microorganisms likely shaped by packaging and storage conditions. Notably, key ARB pathogens, belonging to the clinically relevant ESKAPE group (i.e., Enterococcus faecium, Staphylococcus aureus, Klebsiella pneumoniae, Acinetobacter baumannii, Pseudomonas aeruginosa, and Enterobacter spp.), Salmonella enterica, and Escherichia coli were detected across the production chain (Fig. 3H, I). Most displayed distinct persistence patterns. A. baumannii was present across the stages, peaked at the slaughterhouse and processing plant, then decreased at retail. E. faecium was prevalent at the farm and slaughterhouse, declined sharply during processing, and reappeared at moderate levels in retail samples. E. coli and S. aureus persisted earlier stages but gradually declined, whereas E. cloacae, E. hormaechei, S. enterica, and P. aeruginosa declined steadily.

Diverse taxa contribute to the resistome across the pork processing chain

As shown in Figs. 2 and 3, ARG profiles and microbial community composition showed parallel dynamics across the production continuum. To evaluate whether the resistome profiles were associated with microbiota composition, we performed a Procrustes analysis. A significant concordance was observed between ARG composition and microbial community structure (M² = 0.258, r = 0.861, p = 0.001; Fig. 4A), indicating that variation in resistome composition is strongly associated with shifts in microbial community structure across stages and sample types. To identify the major bacterial hosts carrying ARGs, we examined the co-localization of species-specific marker genes and ARGs on the same assembled contigs (Fig. 4B; Supplementary Data 3), providing evidence of putative genomic linkage between ARGs and their bacterial hosts, which are mainly chromosomally encoded. The top 20 major carriers revealed distinct stage-specific associations (Fig. 4B). High-priority aminoglycoside and tetracycline resistance determinants were concentrated in farm-associated taxa. For example, Faecalibacterium prausnitzii and Anaerobutyricum hallii were identified as primary hosts for multiple ARGs, including those associated with glycopeptide, MLS, multidrug, and peptide resistance. Specifically, F. prausnitzii was a key carrier for tetracycline resistance genes like tet(W) and MLS genes like erm(B) (Fig. 4B; Supplementary Data 3), which are common in the swine gut. In contrast, Pseudomonas and Acinetobacter species predominated as major ARG carriers in the slaughterhouse and processing plant stages. Notably, Acinetobacter johnsonii, Acinetobacter sp. TTH0-4, and Acinetobacter guillouiae were identified as resident taxa that consistently harboring beta-lactam resistance genes, including OXA-type variants and carO (Fig. 4B; Supplementary Data 3). Furthermore, Moraxella osloensis was identified as a stable carrier of multidrug efflux genes across these stages, suggesting persistence of these traits in processing environments. Overall, this taxonomic shift suggests that production environments may selectively maintain specific ARG–host associations.

Fig. 4.

Association between the resistome and microbiome and major bacterial carriers of antimicrobial resistance genes (ARGs) across the pork processing chain. A Procrustes analysis showing concordance between microbiome and resistome ordination structures across processing stages. Circles indicate microbiome ordination positions and squares indicate resistome ordination positions; arrows connect matched samples. Statistical metrics are indicated in the figure. B Heatmap showing the top 20 bacterial species identified as major ARG associated species across different sample contexts. Color intensity represents the number of ARGs detected per species. Left-side bar plots indicate the total number of ARGs detected per species, while right-side stacked bar plots depict the relative composition of ARG classes carried by each species, with colors corresponding to ARG class categories

ARG composition is mainly shaped by stage-specific source contributions

Given the sequential and interconnected nature of the pork production continuum, we investigated the potential source contributions to ARG profiles across production stages using FEAST source tracking analysis. For each downstream stage, samples from all upstream stages together with stage-specific environmental samples were defined as sources, whereas pig carcass or pork product were defined as sinks (Fig. 5A). This framework enabled differentiation between ARGs from upstream carryover and those newly introduced from local environments.

Fig. 5.

Resistome source tracking across pork processing chain. A Schematic representation of the pork supply chain illustrating the framework used for antimicrobial resistance gene (ARG) source tracking from production environments to pork products. Arrows indicate the sequential movement of animals, carcasses, and products through the processing chain and highlight potential opportunities for microbial and ARG dissemination across stages. Environmental sources are denoted as Env. B–D Pie charts summarizing inferred source contributions for pork product samples collected at the slaughterhouse (B; unwashed and washed carcass), processing plant (C; processed carcass), and retail store (D; pork belly and pork neck). Pork product samples were treated as sinks, while all samples from upstream processing stages and environmental samples from the analyzed stage were treated as sources. Colors indicate the relative proportional contribution of each upstream source category. Source contributions were inferred using FEAST source-tracking analysis, providing a quantitative assessment of the relative influence of different production environments on downstream product resistome composition

At the slaughterhouse, FEAST analysis revealed that the slaughterhouse environment was the primary contributor to ARGs detected on pig carcasses, accounting for 93% of the ARG pool, particularly knives (60%) and gloves (17%) (Fig. 5B). In contrast, pig farm-associated sources contributed 3%, indicating partial carryover, while 4% of ARGs originated from unknown sources, suggesting uncharacterized reservoirs. These results indicate that ARG contamination at this stage arise predominantly from on-site environmental exposure rather than transmission from upstream stages. At the processing plant, ARG transmission reflected more complex mixture of sources (Fig. 5C). Processing plant environmental samples, particularly floors (31.1%) and worktables (22%), were the major contributors, along with gloves (6.3%) and knives (4.1%). Notably, slaughterhouse-associated sources collectively contributed 13.8%, indicating persistence of certain ARGs from earlier stages. At the retail stage, ARGs detected on pork products were derived mainly from pig-associated sources originating at earlier processing steps, with the largest contribution (50%) linked to pork products from the processing plant (Fig. 5D). Unknown sources accounted for 17%, indicating that ARG profiles at retail largely reflect downstream processing-associated inputs. Collectively, these FEAST analyses indicate that ARG composition across the pork production chain is shaped by stage-specific source contributions, with a strong influence from local environmental and handling-associated sources at each stage and limited evidence of direct upstream carryover.

MGEs accelerate ARG spread in the pork production chain

Horizontal gene transfer (HGT), primarily mediated by mobile genetic elements (MGEs) such as plasmids, transposons (Tns), insertion sequences (ISs) and integrons, plays a central role in accelerating the spread of ARGs across microbial populations and environments [20, 32, 55, 59]. To assess the mobility potential of ARGs, we examined their co-localization with MGEs across the pork processing continuum. As shown in Fig. 6, Sankey network analysis revealed associations among ARGs, their predicted bacterial hosts, and production stages, with ARG carrying bacterial hosts detected at the pig farm, slaughterhouse, and processing plant. Overall, 13 of 405 ARGs were located on contigs carrying Tns, ISs, and miniature inverted-repeat transposable elements (MITEs), indicating potential for horizontal transfer. These included genes conferring resistance to aminoglycosides (aac(6′)-Ie-aph(2′′)-Ia, aph(3′)-IIa, aph(6)-I), macrolide and lincosamide (ermB, lnuC, mipA), multidrug efflux (oqxB), sulfonamides (sul2), streptothricin (sat-4), and vancomycin (vanS) (Fig. 6A). These ARGs were associated with multiple bacterial genera, indicating broad host distribution across phylogenetically diverse taxa. Stage-specific ARG-MGE associations were also observed. The vanS gene was linked to pig farm, and the oqxB and sul2 genes were linked to slaughterhouse. Aminoglycosides- and MLS-associated genes were predominantly linked to pig farm and processing plant. The sat-4 was consistently linked to Tn5405, detected in Brachyspira pilosicoli, Lactobacillales, and unclassified hosts from both farm and slaughterhouse samples (Fig. 6B). The vanS gene was associated with ISEfa4 in E. faecium, ISBth17 in Bacillus licheniformis, and ISBth16 in Virgibacillus salarius at the farm (Fig. 6C). The lnuC gene was carried by ISag10 in Kushneria phosphatilytica at the slaughterhouse and by Ligilactobacillus animalis in the processing plant (Fig. 6D). Together, these findings indicate that MGEs substantially contribute to ARG dissemination across the pork production chain.

Fig. 6.

ARG–MGE–host relationships across the pork processing chain. A Sankey diagram linking antimicrobial resistance genes (ARGs), associated mobile genetic elements (MGEs), insertion sequences (ISs), transposons (Tns), and miniature inverted-repeat transposable elements (MITEs), and inferred bacterial hosts across processing stages. Ribbon width is proportional to the number of supporting contigs. ARGs were classified as MGE-associated when an ARG and an MGE were co-localized on the same contig within ≤ 500 bp and without overlap. B–D Representative genetic contexts showing co-localization of ARGs and MGEs on individual contigs. Examples include sat-4 (B), vanS (C), and lnuC (D), illustrating potential MGE-mediated carriage of ARGs across diverse bacterial taxa

MAGs reveal process-associated bacterial adaptation

To achieve genome-level resolution and track potential transmission events along the pork production chain, we constructed metagenome-assembled genomes (MAGs). In total, 382 high-quality bacterial MAGs were generated, including 331 from the pig farm, 34 from the slaughterhouse, 16 from the processing plant, and one from the retail store (Fig. 7A). The MAGs were dominated by Bacillota lineages (including Bacillota, Bacillota_A, Bacillota_B, Bacillota_C, and Bacillota_D; 54.45%), followed by Bacteroidota (16.23%), Actinomycetota (13.87%), and Pseudomonadota (7.59%). These MAGs were assigned to 209 known genera, most prominently Prevotella (9.57%), Corynebacterium (5.74%), Streptococcus (3.83%), and Anaerococcus (3.83%), and to 196 known species, although 141 MAGs remained unclassified at the species level. The most represented species included Turicibacter bilis (n = 4), VUNA01 sp002299625 (n = 4), and Coprococcus_A catus_A (n = 4).

Fig. 7.

Genetic profiling of metagenome-assembled genomes (MAGs) across the pork processing chain. A Midpoint tree of metagenome-assembled genomes (MAGs) recovered from the pork processing chain. A total of 382 bacterial MAGs with high assembly quality (completeness > 50% and contamination < 1%) were generated. Bacterial genomes from different phyla are shaded in distinct colors and labeled on the outer ring of the phylogenomic tree. Colored strips outside the tree indicate the processing stage from which the MAGs were recovered. B Sankey diagram illustrating the relationships among processing stages, sample types, dominant bacterial taxa, and associated genetic elements. Nodes on the left represent processing stages, which are connected to corresponding sample types collected at each stage. Intermediate nodes denote bacterial species identified from reconstructed MAGs, while nodes on the right indicate associated genetic elements, including antimicrobial resistance genes (ARGs) and virulence factors (VFs). The width of each flow is proportional to the relative abundance of supporting MAGs. C, D Heatmaps showing the presence–absence patterns of VFs (C) and ARGs (D) and detected in MAGs across processing stages. Rows correspond to individual genes, columns represent processing stages, and colored annotations indicate gene functional classes. Hierarchical clustering highlights similarities in genetic profiles among MAGs along the pork processing chain

MAG based Sankey network analysis linking ARGs, production stage, sampling site, bacterial hosts, and MGEs revealed multiple potential reservoirs and transmission routes for ARGs and virulence factors (VFs) (Fig. 7B). Environmental contact surfaces, including worktables, gloves, knives, and drains, served as major convergence points connecting diverse bacterial taxa to ARGs, VFs, and plasmids. Notably, E. coli, S. suis, and several opportunistic pathogens were associated with ARG and VF signals, suggesting that both commensal and environmental bacteria may contribute to resistance dissemination. A substantial proportion of ARGs (n = 33) was detected in contigs associated with an E. coli-like MAG recovered from a processing plant worktable, with additional ARGs identified in MAG-associated contigs from environmental samples at the pig farm and slaughterhouse. Most VFs were similarly identified in contigs associated with an E. coli genome (n = 26) and in Pseudomonas_E palmensis (n = 32) recovered from a glove at the pig farm. Resistance and virulence profiling further showed that, as expected, fewer ARGs were detected at the MAG level than by read-based approach (Fig. 7C). Nevertheless, MAG-based ARG and VF patterns displayed clear process-specific signatures. The processing plant harbored MAGs with ARGs and VF profiles overlapping those observed in the slaughterhouse, but with no direct overlap with pig farm genomes. Although specific gene sequences varied, many identified VFs were functionally associated with bacterial secretion systems and fimbriae-related traits (Fig. 7D). Collectively, MAG-based analyses suggest stage-associated genome-resolved patterns along the pork production chain, with overlapping resistance and virulence profiles observed in slaughterhouse and processing plant genomes.

Discussion

This study provides an integrated, genome-resolved perspective on the distribution of AMR across the pork production continuum. ARG richness was higher at the pig farm, slaughterhouse, and processing plant stages but decreased significantly at retail. Source tracking analyses suggests that local environmental inputs contribute substantially to ARG diversity at each stage. In addition, environmentally persistent bacteria may serve as key reservoirs for ARGs, potentially supporting their persistence and dissemination within production chain.

We previously used 16S rRNA gene sequencing to characterize bacterial community shifts across the pork production chain, revealing stage-specific dominance of Acinetobacter in slaughterhouses, Psychrobacter in processing plants, and persistence of Acinetobacter, Pseudomonas, and Staphylococcus in retail products; however, species-level resolution was limited [29]. In the present study, shotgun metagenomics enabled species- and genome-level characterization of microbial communities in relation to ARGs, MGEs, and VFs across the pork production continuum using read-, contig-, and MAG-based analyses. Our findings demonstrate that both microbial diversity and ARG profiles are dynamic rather than static, undergoing progressive restructuring from the pig farm through slaughterhouse and processing facility to retail. Our findings are supported by recent shotgun metagenomic study reporting high ARGs in slaughterhouse environmental samples, reinforcing their role as key AMR hotspots [59]. In contrast, we observed significantly lower ARG abundance in retail pork products compared to pig farm and slaughterhouse samples, differing from earlier report in which retail pork carried higher ARG abundance than feces and environmental samples [59]. This discrepancy may reflect differences in processing conditions or sampling strategies between studies.

Source-tracking analyses indicated that stage-specific ARG profiles may be shaped primarily by local environmental and handling-associated sources rather than direct carryover from earlier stages (Fig. 5). The persistence of ARGs is closely linked to the microbial ecology of food-processing facilities, which is shaped by dynamic interactions between environmental conditions, including temperature, moisture, and surface type, and operational practices such as sanitation frequency, disinfectant use, and production downtime (Koutsoumanis et al. [30]). These factors impose strong selective pressures that favor microbial survival, persistence, and recolonization. Consistent with recent metagenomic and culture-independent studies, our results indicate that processing environments harbor diverse bacterial communities that are not fully eliminated by sanitation, but instead subset of bacteria that establish stable niches, often associated with biofilms on equipment and facility surfaces [9, 10, 23, 38, 58, 60]. The composition of these resident populations varies across facilities, reflecting site-specific ecological and management pressures, and they may serve as long-term reservoirs of ARGs that facilitate persistence [13, 23].

Notably, the opportunistic human pathogen A. baumannii persisted across all production stages (Fig. 3), suggesting their persistence likely reflects adaptation to pork the production environment through stress-tolerance, biofilm formation, and surface adherence, enabling recolonization after cleaning. Despite microbial compositional differences, substantial overlap between persistent and stage-specific dominant taxa indicates that many microorganisms become dominant at specific production stages, reflecting ecological stability and adaptive persistence within pork production environments (Fig. 3). This pattern supports the role of both animals and environments as reservoirs capable of recolonization, particularly when sanitation or hygiene practices are insufficient or inconsistently applied. Although the overall abundance of AMR-associated pathogens, including ESKAPE groups (Fig. 3H, I), decreased toward the retail stage, their continued detection across production stages underscores a residual risk for resistance persistence and post-processing contamination. Similar observations have been reported in other meat-processing systems, where Acinetobacter and Pseudomonas remain among the most stable genera despite frequent disinfection, and long-term persistence of Listeria and Aerobacter has been attributed to spatially stable resident microorganisms [10, 23, 60]. Together, these findings indicate that microbial persistence in processing environment is not stochastic but represents a reproducible ecological feature shaped by environmental selection and facility-specific practices.

The ARG composition varied significantly across production stages. Our analysis revealed that environmentally persistent bacteria served as major reservoirs for a diverse suite of high-priority resistance determinants (Fig. 4; Supplementary data 3). These included clinically significant acquired ARGs such as aminoglycoside-modifying enzymes (aac(6') and aac(3) families) [25, 46], beta-lactamases (carO, OXA-72, PBP-2X, ADC-16, ADC-21) [6, 8, 39, 42, 54], and MLS resistance genes (ermB, lnuA-D) [48, 62]. In contrast, a substantial proportion of the remaining resistome was composed of multidrug efflux-associated genes (e.g., mexB, adeB) [22, 37]. These likely represent intrinsic or adaptive resistance mechanisms rather than mobile, acquired determinants of immediate clinical concern. The stage-specific association of distinct ARG and MGE combinations (Fig. 6), such as vancomycin resistance at the farm and sulfonamide and multidrug resistance genes at the slaughterhouse, suggests that local selective pressures shape ARG mobilization dynamics. However, detection of ARG-MGE co-localization (e.g., lnuC) should be interpreted as indicative of potential mobility rather than evidence of active horizontal gene transfer, particularly given the limitations of short-read assemblies in resolving multiple genomic contexts.

The enrichment of secretion- and fimbriae-related genes suggests that these bacteria possess adhesion and biofilm-forming capabilities that promote surface persistence (Fig. 7C). Such traits align with findings by in a UK ready-to-eat meat factory [23], where P. fluorescens dominated stable microbiota over ten weeks and exhibited genetic adaptations to cold and surface colonization. Furthermore, these findings are supported by previous report that non-pathogenic and opportunistic bacteria play key roles in maintaining resistance and mobility elements [26, 34]. Many of these organisms thrive within biofilms and display enhanced tolerance to disinfectants, temperature shifts, and desiccation [12, 45]. Notably, overlapping resistance and virulence profiles between slaughterhouse- and processing plant–associated MAGs, despite limited genome-level continuity from the farm, indicate adaptive persistence and reassembly of functionally similar resistomes downstream. Studies by Quigada [45] across European food systems also found that food-associated resistome are primarily structured by the processing environment rather than product type, underscoring that antimicrobial resistance ecology in food production is largely determined by environmental conditions. The similarity in ARG class distributions between slaughterhouse and processing plant samples may reflect shared selective pressures or environmental conditions.

While this longitudinal design provides valuable insights into stage-specific dynamics, several limitations should be considered. The limited number of replicates may constrain the capture of full ARG and microbial diversity, warranting validation with expanded and balanced sampling. In addition, assembly- and MAG-based approaches have inherent limitations in resolving ARG genomic context, as short-read assemblies may capture only dominant sequences while masking micro-diversity. Long-read or hybrid approaches will be important for resolving ARG context and host-range dynamics.

Conclusions

This study identifies pork processing environment as critical ecological hubs for AMR persistence and dissemination. Environmental reservoirs and MGE-mediated gene exchange, rather than animal carriage alone, play a central role in shaping resistome dynamics along the pork production chain. Targeting environmental conditions and operational practices that favor microbial persistence and ARG mobility may provide effective strategies to mitigate antimicrobial resistance risks in food production systems.

Authors’ contributions

DB, KCJ, and SAK acquired funding. DB, KCJ, and SAK conceptualized. DB and SK collected samples. YZ and YK analyzed, collected, and visualized data. DB, SAK, and KCJ provided resources. DB, KCJ, and SAK supervised and validated. YK, GB, YK, and SAK wrote the first draft. All authors wrote and reviewed the main manuscript.

Funding

This research was supported by the National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (grant number RS-2023-00212117 and RS-2024-00406525) to SAK, the Ministry of Food and Drug Safety (20162MFDS011) to DB and SAK, and the Bennink Foundation (F028417) to KCJ.

Data availability

Raw reads are available on the Sequence Read Archive of the National Center of Biotechnology Information (NCBI) under the BioProject number PRJNA1395050.

Declarations

Ethics approval and consent to participate

In this study, fecal samples were collected non-invasively following natural defecation, without any direct handling or intervention on the animals by the researchers. Therefore, according to institutional guidelines, this type of sample collection is considered exempt from formal animal ethics approval. In addition, all sampling procedures were conducted in collaboration with farm personnel and facility operators, following standard practices and with their consent. The consent forms were originally prepared in Korean, and an English translation has been provided as a supplementary document.

Competing interests

The authors declare no competing interests.