This study was conducted in the Khon Kaen province of North-East Thailand during September-December 2018 in 166 pig farms. Farms were separated on production scale, all farms in this study were considered either small- or medium-scale farms. In this study small-scale farms were defined as having ≤50 sows (often family owned), whilst medium-scale farms were defined as having ≥100 and ≤500 sows and (company owned). In total there was 115 small-scale and 51 medium-scale farms. At each farm we tried to obtain a fecal samples from a contact human and a non-contact human and up-to 10 pigs (depending on the number of pigs on the farm) that were pooled together. Overall we collect 143 pig, 90 contact human and 54 non-contact human fecal sample and stored in accordance to the methods described by Lunha et al. (2020) at the Faculty of veterinary medicine, Khon Kaen University.

Pig rectal swabs from the same farm (n ≤ 10) were pooled in 25 mL of buffered peptone water for enrichment, information regarding numbers of pooled pig rectal swabs from specific farms is provided in Lunha et al. (2020). One gram of human faces was enriched in 5 mL of buffered peptone water. After incubation at 37°C overnight, 10 μL of suspension was plated onto MacConkey agar (Oxoid, Hampshire, England) containing cefotaxime 1 μg/mL (Sigma-Aldrich), CHROMagar^TM mSuperCARBA^TM, and CHROMagar^TM COL-APSE (CHROMagar, Paris, France). Different morphologic types of colonies were collected and identified by matrix-assisted laser desorption ionization time of flight (MALDI-TOF, Microflex, Bruker Daltonik GmbH). All isolates were re-streaked from frozen stocks onto blood-agar plates, incubated at 37°C 18 h and checked for potential contamination before AST. All AST were performed using the Sensititre EUVSEC AST microdilution plates using Sensititre Mueller Hinton Broth tubes and the Sensititre AIM pipetting robot (Thermo Fisher Scientific, Fair Lawn, NJ, United States). The AST was done in accordance to the Sensititre EU Surveillance Salmonella/E. coli EUVSEC Plate workflow (Thermo Scientific Sensititre Plate Guide for Antimicrobial Susceptibility Testing, 2018). Quality control was performed with E. coli ATCC 25922 isolate and produced the expected results. All plates were manually read with the sensititre manual viewbox and photographed for documentation. Isolates were catogorized in accordance to definitions of Enterobacteriaceae in Magiorakos et al. (2012) where multi-drug resistance (MDR) was defined in as when isolate is non-susceptible to at least 1 agent in ≥ 3 antimicrobial categories, extensively multi-drug resistance (XDR) when the isolate is non-susceptible to at least 1 agent in all but 2 or fewer antimicrobial drug catergories and pan mutli-drug resistance when the isolate is non-susceptible all antimicrobial drug catergories.

All isolates were re-streaked from frozen stocks onto blood-agar plates, incubated at 37°C 18 h and checked for potential contamination before DNA extraction. DNA was extracted using 10–15 μL of a fresh bacterial colonies and processed using the Qiagen DNeasy blood and tissue kit in accordance to the manufacturer’s instruction (Qiagen, Hilden, Germany). All processed samples had a minimum DNA concentration of 30 ng/μL verified using the dsDNA HS Assay kits on a Qubit Fluorometer (Thermo Fisher Scientific, Fair Lawn, NJ, United States) and OD260/280 in the range of 1.8–2.0 verified by the NanoDrop spectrophotometer (Thermo Fisher Scientific, Fair Lawn (NJ), United States). Samples were shipped on dry ice to the Novogene sequencing facility (Novogene, Hong Kong, China) and sequenced on the Novoseq Illumina platform (Illumina, San Diego, CA, United States) and produced ∼5 GB of 150 bp pair-end sequencing reads per isolate.

All AST data was converted into either a susceptible, intermediate or resistant classification for each isolate and all the individual tested antibiotic drugs in accordance to EUCAST clinical breakpoints (European Committee on Antimicrobial Susceptibility Testing, 2013). Then further processed in Python (3.8.3rc1 Documentation, 2020) using the matplotlib, pandas and seaborn packages. Genomic data processing was performed on the high computing capacity provided by SNIC through Uppsala Multidisciplinary Centre for Advance Computational Science (UPPMAX). All processing of the whole genome sequences was done with open software with an in-house bioinformatics pipeline. Our pipeline consists of four main modules: the first module performed a quality control (QC) assessment of the raw sequence files and trimming the sequence reads with FastQC (Wingett and Andrews, 2018), MultiQC (Ewels et al., 2016), and Trim Galore (Babraham Bioinformatics, 2020); the second module performed de novo assembly with Unicycler with minimum contigs of 400 bp produced (Wick et al., 2017) and assembly QC was performed by QUAST and MultiQC (Gurevich et al., 2013; Ewels et al., 2016); the third module compiled a molecular output excel report from running our genomic data through KmerFinder (Larsen et al., 2014), ARIBA (Hunt et al., 2017), ResFinder (Zankari et al., 2012), and PointFinder (Zankari et al., 2017); and the fourth module produced a core genome maximum likelihood phylogenetic tree by running Prokka (Seemann, 2014) to annotate all the isolate genomes, Roary (Page et al., 2015) to assess the pangenome and generate the core genome alignment, and IQ-Tree2 (Quang et al., 2020) to generate maximum likelihood phylogenetic tree (Figure 1). Downstream genomic data processing was done in Python (3.8.3rc1 Documentation, 2020) using the matplotlib, pandas and seaborn packages. Visualization of the maximum likelihood phylogenetic tree was done using iTOL webtool (Letunic and Bork, 2019). Additional SNP detection of matching ST isolates between groups on the same farm was done using snippy (Seemann, 2015), detection of virulence genes was done using the virulencefinder pipeline (Joensen et al., 2014) and pangenome query difference tables were produced between human and pig groups using Roary query pangenome function (Page et al., 2015).

This study was performed in accordance to the Helsinki declaration for the human subjects and the EU Directive 2010/63/EU for animal experiments; the protocol involving human participants and animals was approved by the Khon Kaen University Ethics Committee (Project ID: HE612268 and 0514.1.75/66, respectively). Informed consent for each human subject was obtained after an explanation of the experimental procedures. Pig samples were collected at the farm with the farm owners permission of the pig herd.

Raw sequence data can be obtained from the European Nucleotide Archive (ENA) under the project accession number PRJEB38313. All sequence data from computation workflow is compiled in Supplementary File 1, virulencefinder unique results in Supplementary File 2 and genomic difference output is provided in Supplementary File 3. All scripts and Python code and metadata can be provided on request.

To explore zoonotic AMR patterns in E. coli isolates we examined 156 farms in Khon Kaen province of North-East Thailand, utilizing our livestock and human participant One Health study model. A total of 143 pooled pig samples, 90 C and 54 NC human fecal samples and were collected and processed. We generated an isolate collection of 492 E. coli isolates derived from culturing fecal material on agar plates selective for ESBL-producing, carbapenemase-producing, or colistin-resistant gram-negative bacteria. From the 107 small-scale farms (SSF) 174 isolates were from livestock pigs (P), 99 isolates from livestock producers (C), and 57 isolates were from and non-livestock contact individuals (NC). From the 49 medium-scale farms (MSF), 94 isolates were from Ps, 44 isolates from Cs, and 24 isolates were from NCs. Several studies in farms in South-East Asia have looked at E. coli isolates from different livestock and farm workers (Nguyen et al., 2016; Lugsomya et al., 2018; Prakobsrikul et al., 2019); to build upon these studies we wanted to additionally assess farm production size in relation to P, C, and NC and look for differences between the six groups. To record significant differences between the isolate groups we used AST and WGS with in silico methods (Figure 1) to explore the genomic diversity (Read and Massey, 2014; Abdelgader et al., 2018; Boehmer et al., 2018).

Utilizing the E. coli isolates to assess AMR, we could establish AMR patterns to demonstrate single antibiotic drug efficacy. Several antibiotic drugs had an average group MIC concentration that exceeded the clinical breakpoint: ampicillin, chloramphenicol, ciprofloxacin, cefotaxime, gentamicin, and trimethoprim (Figure 2). This was not surprising for ampicillin or cefotaxime due to the selective culturing method, enriching, e.g., for ESBL E. coli. Whilst for other drugs we observed clustering around the clinical breakpoints, e.g., colistin, ceftazidime, and tigecycline. Yet again these results were not surprising for colistin or ceftazidime due to the selection for ESBL and colistin-resistant E. coli. Our results demonstrate that ceftazidime and tigecycline both had a higher phenotypic AMR bias in the human isolates especially those from MSFs. This was not the case for colistin where there was a resistance bias toward SSFs, with the highest results occurring in the SSF-P isolates with an average MIC concentration of 4.2 mg/L, second highest in SSF-C isolates with an average MIC concentration of 3.3 mg/L, and third was the SSF-NC isolates with an average MIC concentration of 2.8 mg/L. These results suggest SSF pigs as a colistin AMR reservoir and zoonotic transfer to the human population. Colistin resistance in Thailand has been frequently reported (Nguyen et al., 2016; McEwen and Collignon, 2018; Ström et al., 2018; Wangchinda et al., 2018; Collignon and McEwen, 2019) but here we report an accumulation in SSFs and signs of zoonotic transmission to humans. However, for chloramphenicol and cefotaxime we saw a preference for human-associated host AMR with a bias toward MSFs rather than SSFs. The average MIC value for chloramphenicol had the highest average MIC value in MSFs in NCs 53.7 mg/L, followed by 44.4 mg/L in Cs and lowest in Ps 31.7 mg/L. Cefotaxime had the highest average MIC value in MSF in C 3.9 mg/L, followed 3.5 mg/L in NCs and lowest in Ps 3.1 mg/L. The only drug that had wide-scale average MIC concentration that indicated susceptibility in all groups was meropenem. In our isolate collection only 6 isolates were cultivated from our carbapenem-selective plates, out of these isolates only 2 were confirmed by AST testing to be meropenem resistant (Supplementary Figure 1). Both isolates were obtained from humans on independent farms; one from a NC and the other from a C. To ensure our findings were statistically relevant we performed one-way ANOVA statistical tests with a Bartlett’s test for each antibiotic based on the group MIC values to see significant differences both in the mean and the standard deviation. Four antibiotic drugs had a significant difference between groups: chloramphenicol, colistin, cefotaxime, and meropenem. In addition we also converted our group isolates into the susceptible (S), intermediate (I) or resistant (R) categories using the EUCAST breakpoints (Supplementary Figure 2). In this analysis we observed the same trends for the four antibiotic drugs that had significant differences in the MIC-value comparison. For instance we still observed with colisin the zoonotic spill-over effect in the SSFs where overall resistant isolates in the groups were 55, 38, and 29% in P, C and NC, respectively, compared to in MSFs 5.3, 13.6, and 20.8% in the same respective order.

We also wanted to examine the collateral damage of AMR of multiple agents in different antibiotic classes in our isolate data. As others have stated that this is a prolific AMR problem in South-East Asia, from our data we can also confirm this problem both in pigs and humans (Lai et al., 2014; Chereau et al., 2017). Our results are summarized in Table 1 showing the distribution of drug resistance from 0 up to 8 different classes of antibiotics. Following the definitions provided by Magiorakos et al. (2012) we found the most extensive multi-drug resistant (XDR) isolates in the NC cohort. We also found more XDR in the MSFs compared to the SSFs in all groups, with the highest occurrence difference being in P and C. Further, we observed significantly more multi-drug resistant (MDR) in the MSF-C group compared to the SSF-C group.

Following our phenotypic data, we continued with genotypic characterization. Unlike other preceding studies we used WGS of our 492 isolates to see the full extent of the antibiotic resistance genes (ARGs) and known chromosomal mutations present. Within our data we saw higher diversity of ARGs in the SSFs compared to the MSFs (Figure 3), consistent with a less regulated and uniform AMU in family owned SSFs as compared to company owned MSFs. The majority of these ARGs were in the aminoglycoside, beta-lactam, colistin, fosfomycin, rifampicin, quinolone and trimethoprim antibiotic drug classes (Figure 4A). A total of 118 different ARGs were discovered in the SSFs compared to 92 ARGs in the MSFs, with 31 unique to SSFs and 6 unique to MSFs (Supplementary Table 1). The most striking difference was observed in the SSF-P group where resistance genes to beta-lactam and colistin antibiotic drug classes accounted for 39 and 15% of total ARGs, respectively. The abundance of colistin ARGs in the SSF also correlated with our previously discussed phenotypic results where the highest abundance of colistin resistance genes was observed in Ps followed by Cs and lastly by NCs. We suspect this is due to the fact that it is common for colistin to be used in pig feed and treatments given to SSF-Ps, and consistent with reports of high colistin resistance prevalence in Thai E. coli isolates from livestock (Nguyen et al., 2016). We did not detect the same correlation with the beta-lactam ARGs where pigs in SSFs had the highest abundance whilst the other groups appeared to have similar values.

Looking in our WGS dataset we were also interested in detecting chromosomal mutations that confer known antibiotic resistance (Figure 4B). From our literature-supported results, most of the chromosomal mutations conferred resistance to the quinolone antibiotic drug class with the exception of one associated with kasugamycin resistance and the other to ampicillin and clavulanic acid resistance and cephalosporin drugs. We found three point mutations that were present in all sample groups ≥10%, these were: gyrA S83L (Everett et al., 1996), parC S80I (Kumagai et al., 1996), and gyrA D87N (Sáenz et al., 2003). All of these mutations are well-known to confer quinolone resistance and are frequently seen both in human and farm animal samples. All the detected chromosomal mutations in our dataset had a human rather than animal bias and were generally biased toward MSFs compared to SSFs.

From our genomic analysis we wanted to observe the prevalence of ESBL and colistin resistance genes within our isolate groups (Table 2). We saw a high prevalence of ESBL genes in all groups with the lowest value being 96.6% in MSF-C group, this was much higher than expected due to previous reports in South-East Asia (Nguyen et al., 2016; Atterby et al., 2019). Despite the higher prevalence in our study compared to other reports, the observation of ESBL genes matches our phenotypic results as the average MIC for all groups as was above the clinical breakpoint for cefotaxime, a third-generation cephalosporin (Figure 2). We also saw the same trend for colistin resistance as seen in phenotypic and ARG diversity results. There was a diminishing resistance trend in SSF from P, C to NC, whilst the trend was in the opposite direction on MSF. It should be noted that exact resistance prevalence per farm cannot be directly compared between human and pig samples, as human samples were obtained from single individuals, and pig samples from pools of up to 10 swabs from different pigs. However, the difference in trends is not dependent on exact prevalence, and thus also these results suggest the highest colistin AMU on SSFs, that could be a hot-spot for colistin resistance development and zoonotic spread, unless colistin in pigs is prohibited.

To assess the clonality of the isolates from the different study groups we did in silico multi-locus sequence typing analysis to identify the sequence types (ST). Then we looked for the most frequent ST in each group (Table 3) and plotted overlapping STs in the different groups (Figure 5). There was a large variation of STs among our isolates and the data suggests some ST types having a high propensity toward a given host (i.e., a large ST overlap between contact and non-contact humans in small-scale farms in Figure 5, and a large number of STs only present in pigs) (Sheppard et al., 2018). However, there were also signs of clonal transfer or shared STs between pigs and humans: (i) ST48 was the most common ST in small-scale farm pigs and contact humans but in no other groups; (ii) the ST overlaps between contact humans and pigs were larger than the overlaps between pigs and non-contact humans in both farm types. To further investigate the clonality of the different isolates within our collection we generated a maximum-likelihood phylogenetic tree (Supplementary Figure 2). From our maximum-likelihood phylogenetic tree we did not observe clonality from farm isolates, we only saw some clustering based on MLST type and sample group. ST48 has been widely reported for harboring ESBLs mainly in LMICs or from imports (Seenama et al., 2019; Tian et al., 2020) but also from human contamination in drinking water in a high-income country (Madec et al., 2016). However, we concluded that several sequence types were circulating within the farm environment with different host preferences that all may play a key role in AMR.

To further investigate clonality we examined total variant differences between matching ST isolates in different groups found on the same farm (Table 4). We saw 6 instances where ≤25 variants were detected between group isolates, the majority of these were between C and NC isolates where 3 were from SSFs and 2 from MSFs and one case between a C and P on an MSF. We assumed this bias was due to gene differences required between pigs and humans, such as virulence genes. To confirm this we did virulence gene assessment using virulencefinder (Joensen et al., 2014) but only found four different genes between the human isolates and P isolates, these were cma, air and eilA in the human sourced isolates and stb in Ps (Supplementary File 2). We also use the query pangenome function of the Roary pipeline (Sitto and Battistuzzi, 2020), however we did not see any conclusive differences (Supplementary File 3). We tended to observe higher clonality amongst isolates from the same host and similarity of ARG carriage depending on the farm type, this was most clearly observed between small-scale farms between pigs and contact humans with colistin resistance. From further analysis we were able to detect the origin of replication of multiple plasmids (Supplementary Figure 3), thus supporting that transient zoonotic transmission could allow for acquisition of ARGs by horizontal gene transfer to the host’s endogenous E. coli population, as we could observe ST differentiation, but no significant difference in ARG carriage depending on the antibiotic drug. In our data set we only saw one shared strain with 11 SNP or INDEL genomic variants difference between a contact human and a pig on farm 120, a medium-scale farm. We could not see extensive zoonotic sharing of clones between human and pigs. The exception in farm 120 is likely the result of a transient zoonotic transmission event, where the transient bacterial strain could disseminate any of its ARGs to the host’s endogenous Gram-negative bacterial community thus giving rise to new antibiotic resistant strains. The frequency of these events and the amount of antibiotics used both in humans and pigs as a selective pressure for ARGs maintenance will affect the impact.