Poultry farm and slaughterhouse environments contain a diverse array of both non-pathogenic and pathogenic multidrug resistant E. coli with phylogenetic intermixing among livestock, humans, meat and environment

We collected and sequenced a total of 154 E. coli isolates from two sites, a commercial poultry farm in Shandong province, China, and the contracted slaughterhouse in the same province. Isolates were collected from three hosts: chickens (n = 82; faecal, cloacal, caecal and carcass), humans (n = 58; faecal, hand and nasal) and environment (n = 14; soil, water, feed) (Fig 1A and S1 Table). Samples were taken over a two independent 6-weeks broiler production cycles (first cycle denoted by t, second cycle denoted by l), with samples taken at three time points: t₁/l₁ –mid-cycle (3 weeks old chickens), t₂/l₂ –fully grown (6-week-old chickens) and t₃/l₃ –end-of-life (6 weeks + 1day, chickens after slaughter). In total 114 isolates were taken from the farm location (67 chicken, 11 environment and 36 human) and 40 isolates were taken from the slaughterhouse (15 chicken, 3 environment and 22 human). To compare the genetic relatedness of the cohort, a maximum likelihood phylogenetic tree for the 154 genomes based on the core genome (2788 genes, present in ≥ 99% of isolates) was constructed. The tree coupled with the in silico typing demonstrated high genetic diversity overall with phylogenetic intermixing between isolates from human and livestock (Fig 1B). The cohort spanned six of the seven common phylogroups (A, B1, C, D, E and F) as well as the more recently identified G and clade I groups (S1 Table). Phylogroups were not significantly associated with a host (chi-squared test with p-value_Bonferonni < 0.05) or farm/slaughterhouse environment (chi-squared p-value_Bonferonni < 0.05). Multi-locus sequence typing revealed diversity with 52 different sequence types across the cohort (S1 Table). 60% of the isolates could be further grouped into clonal complexes with CC155 (n = 17), CC205 (n = 13) and CC206 (n = 11) the most prevalent. CC155 [31,32] as well as other less prevalent CCs in our cohort have been previously associated with multidrug resistance (CC10 [n = 6] (52), CC469 [n = 6] [33,34], CC648 [n = 4] [3,34] and CC86 [n = 3] [35,36]. Analysis of pathotypes and serotypes (S1 Table) showed that the majority of isolates in our cohort were non-pathogenic bacteria, however, 39 isolates (25.3%) were found to be EAEC strains and 1 was EPEC (Fig 1B). The EAEC isolates were found in all phylogroups but were significantly associated with phylogroup G (chi-squared p-value_Bonferonni < 0.0001), sequence types ST117 (chi-squared p-value_Bonferonni < 0.0001) and ST542 (chi-squared p-value_Bonferonni < 0.015). EAEC isolates were found across all hosts and sources (Fig 1B), however, farmworker nose swabs were significantly associated with these pathogens (chi-squared p-value_Bonferonni < 0.0198).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. A wide array of non-pathogenic and pathogenic E. coli with phylogenetic intermixing, and high prevalence of shared multidrug resistances amongst livestock, humans, meat and environment are present in a poultry farm and the contracted slaughterhouse.

(A) Overview of the study design and cohort information. 154 human, animal, meat and environmental samples were collected at each of three time points (denoted t1-3 and l1-3) over two independent 6-weeks broiler production cycles over one farm as well as the abattoir linked to this farm in China. The total number and type of isolates cultured and sequenced by host and host body site are shown. (B) Maximum likelihood phylogenetic tree of the whole cohort based on core genome of the 154 isolates with recombination correction, cultured from the human, animal and environmental and meat samples collected from the farm and slaughterhouse. Phylogroups (inner ring), sequence types, clonal complex, collection time points, host type and resistance profile (outer ring) are shown around the outside of tree. EAEC and EPEC isolates indicated by ⁕ and ⁕⁕ respectively. (C) Numbers of susceptible and resistant isolates, based on antimicrobial susceptibility testing by broth microdilution, for each of a panel of 26 antimicrobials encompassing nine different antimicrobial groups.

https://doi.org/10.1371/journal.pcbi.1010018.g001

All isolates underwent laboratory testing for resistance to a panel of 26 antimicrobials encompassing 9 classes: beta-lactams, aminoglycosides, chloramphenicol, quinolones, glycopeptides, tetracyclines, sulphonamides, polymyxins and trimethoprim. The proportion of resistance to each antimicrobial ranged from 3% to 89% (Fig 1C, Table 1). Resistance to antibiotic of last resort colistin/polymyxin was found in only 5% of isolates (n = 8) sourced from chicken, human and environmental hosts. The median number of resistances per isolate was 13 and 96.7% of isolates showed resistance to at least one antimicrobial, with 92.2% having resistance to at least three. Notably, 12.3% of the cohort were resistant to at least 20 antimicrobials, with 23 being the maximum number of resistances for a single isolate indicating extremely high levels of antimicrobial resistance among the E. coli present in this farming environment. The most prevalent resistance profile (n = 47) in the cohort was penicillins, aminoglycosides, mono-beta-lactams, cephalosporins, chloramphenicol, quinolones, tetracyclines, sulphonamides and polymyxins. This profile was found present in different ST type clusters in the tree (ST205, ST155, ST101, ST117, ST206, ST2690) with three of these containing chicken and human isolates (ST205, ST155, ST206) as well as one containing human and an environmental isolate (ST117), see Fig 1B.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Number of resistant and susceptible isolates included in this study.

https://doi.org/10.1371/journal.pcbi.1010018.t001

Network analysis based on pairwise single nucleotide polymorphisms (SNPs) alignment highlights complex pattern of highly related isolates crossing host boundaries

To assess the relatedness of isolates across our cohort we compared the number of core gene SNPs per isolate in a pairwise manner. Across the cohort the median SNP difference was 18386 (range 0–52377, IQR 22240). Setting a maximum threshold of 15 SNPs as previously done in a similar study [37], we assessed the relatedness of the isolates in our cohort (S1 Fig). The resulting network showed several host-specific clusters of chicken- or human-only samples. The human-only clusters were non-pathogenic whilst two chicken-only clusters contained pathogenic isolates only, and a further two showed pathogenic and non-pathogenic isolates intermixed. Interestingly, the network also showed inter-host clusters. Specifically, three clusters of chicken and human samples (maximum SNP differences of 2, 5 and 8 SNPs; two non-pathogenic and one mixed); two clusters of human and environmental samples (0 core SNP differences in each cluster; one pathogenic and one non-pathogenic); three clusters of chicken and environment (maximum SNP differences of 0, 2 and 5, two pathogenic and one mixed) and finally two clusters with all three hosts intermixed (maximum SNP differences of 1 SNP per cluster; one mixed and one non-pathogenic). This close genetic relatedness (0 to 8 SNPs) among livestock, human and environmental isolates highlights the potential for transmission of pathogenic and non-pathogenic E. coli bacteria within this environment.

Machine learning identifies with high accuracy known and potentially novel genetic determinants of resistance to twenty-one different antimicrobials, discriminating drug-resistant and susceptible strains in the cohort

To test whether the antibiotic resistant/susceptible phenotypes could be explained by known AMR-associated genes (including genes that are known to confer antimicrobial resistance based on their annotations and well-known AMR genes) present in the cohort, the isolates were scanned for sequence matches to genes in AMR gene databases (CARD [38], ResFinder [39], ARG Annot [40], NCBI AMRfinder [41]). The Jaccard/Tanimoto coefficient of similarity between resistance phenotypes for each antibiotic and the known AMR-associated genes was calculated in a pairwise manner. Values spanned the range -0.13 to 0.53, where 1.0 would indicate a perfect positive association and -1 would indicate a perfect negative association. These results indicated that no single known AMR-associated gene could be found that explained the resistance phenotype observed for each antibiotic. The strongest association (0.53) was found between the gene rmtB and resistance to amikacin, an association that has been observed before in China [42]. The lack of a simple association between known AMR-associated genes and resistance together with the genetic relatedness found in our cohort motivated further analysis to identify genetic associations with our observed phenotypes.

To this end we applied a machine learning approach to specifically deduce the relationship between genome content and AMR profiles of each isolate to each one of the antimicrobials. To avoid any bias in the supervised learning approach we first tested for the presence of clonality in the population. To do this we calculated the standard association index (I^S_A) [43]. The calculated I^S_A value was 0.2497 (p-value < 0.00001) at whole cohort level and 0.1111 (p-value < 0.00001) as ST type level (one isolate used to represent each of the 52 STs) meaning no linkage disequilibrium was detected indicating the absence of clonality in the E. coli population [44].

Ten supervised learning classifiers, logistic regression, linear support vector machine and radial basis function kernel support vector machine (RBF-SVM), extra tree classifier, random forest, adaboost, xgboost, naïve bayes, linear discriminant analysis and quadratic discriminant analysis, were used to predict susceptible and resistant strains for each of the 26 antimicrobials. All genome sequences were processed to give overlapping 13 base pair k-mer features related to each sequence. Five antibiotics (amikacin, polymyxin B, polymyxin E, imipenem and meropenem) did not have enough samples in one class to allow cross validation and SMOTE and so were not taken further. To reduce the number of features analysed by the classifiers a two-step feature selection method was applied: firstly, k-mers with a p-value greater than 0.05 according to the chi-square test were discarded [45,46] and then those remaining were used as input to an Extra Tree Classifier. K-mers with a Gini feature importance above the overall mean were selected [47,48].

To ensure robustness all models were trained and tested in 30 independent runs, with the performance metrics given as the mean of all runs. To verify which classifier performs better out of the 10 classifiers studied we applied a Friedman F-test with Nemenyi post hoc test. With 10 classifiers and 21 antibiotics, the Friedman test is distributed according to the F distribution with 21−1 = 20 and (10−1)×(21−1) = 180 degrees of freedom. The critical value of F(9,180) for α = 0.0001 is 4.0366. For all the metrics studies (AUC, accuracy, sensitivity, specificity, precision and Cohen’s Kappa), the F_F statistics null hypothesis was rejected with a confidence level of 99.99%. The null hypothesis states that there are no differences between the average ranks of each classifier over the 21 antibiotics. After, the Nemenyi post-hoc test was performed, and the critical difference diagram was set at 3.407 with a confidence level of 99%. According to S2 Fig, the classifier RBF-SVM had the best rank for the metrics AUC, accuracy, sensitivity, Cohen’s kappa, second best for precision and third best for specificity. Nonetheless, according to the Nemenyi test, for precision and specificity, RBF-SVM is statistically equivalent when compared to the best ranked classifier in these two metrics. Therefore, the RBF-SVM classifier was selected for this study. Twenty-one antimicrobial models achieved high performance scores (Fig 2 and S2 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Supervised machine learning prediction of antimicrobial resistance signature profiles to 21 antimicrobials in the E. coli cohort.

(A) Flow diagram showing the machine learning pipeline including feature selection (yellow), classification (blue), post-processing (green). (B) Prediction performance results of the RBF SVM classifier that achieved the best performance amongst the three investigated is shown. Five performance indicators have been used to evaluate the classification: i) Area under the curve AUC, ii) Accuracy, iii) Sensitivity, iv) Precision, v) Specificity, and vi) Cohen’s Kappa value from 30 training runs for each antimicrobial. The scores for each performance metric are indicated in the Y axis. Predictive models were generated to classify the resistance vs. susceptibility profiles of twenty-one different antimicrobials (X axis): amoxycillin/clavulanic acid (AMC), ampicillin (AMP), ampicillin/sulbactam (AMS), aztreonam (AZM), ceftazidime (CAZ), cefoxitin (CFX), cefazolin (CFZ), chloramphenicol (CHL), ciprofloxacin (CIP), cefotaxime (CTX), doxycycline (DOX), cefepime (FEP), gentamicin (GEN), kanamycin (KAN), levofloxacin (LEV), minocycline (MIN), nalidixic acid (NAL), streptomycin (STR), trimethoprim/sulfamethoxazole (SXT), sulfisoxazole (SUL) and tetracycline (TET).

https://doi.org/10.1371/journal.pcbi.1010018.g002

Ampicillin, cefazolin, and tetracycline achieved an AUC score higher than 0.99, with 17 antimicrobials achieving an AUC between 0.9 and 0.99. Only streptomycin achieved an AUC score below 0.9. The antimicrobial with the highest AUC score (0.998 ± 0.002) was cefazolin, it achieved also an accuracy of 0.977 ± 0.009, sensitivity of 0.982 ± 0.01, specificity of 0.958 ± 0.037 and a Cohen’s Kappa score of 0.929 ± 0.029. Similarly, the precision was also high with values ranging from 0.706 ± 0.107 to 0.991 ± 0.03 and 11 models having a mean precision over 0.94.

To verify if the number of samples for each antibiotic model was large enough for the test set to be representative, we employed a wrapper backward selection (WBS) in terms of the samples using two different analyses: (i) firstly, we applied the WBS (see S3A Fig), to the original pipeline used in this work (i.e. nested cross validation with an RBF-SVM as the classifier and a SMOTE approach during the training phase to balance the classes for the training of the classifier (as used in our ML pipeline), 5 iterations of the nested cross validation were used; (ii) while, for the second analysis, we first performed a SMOTE approach as a pre-processing step to oversample the minority class and then applied the WBS method on the oversampled data, and analysed using 5 iterations of a nested cross validation with a RBF-SVM as the classifier (S3B Fig). S3A Fig shows that for 4 out of 21 antibiotics (AMP, TET, CFX and MIN) the WBS finished after a few iterations due to the small number of samples in the minority class. Nonetheless, according to the learning curves of the other 17 out of 21 antibiotics on S3A Fig the training and test datasets are representative for the initial number of samples in each antibiotic; with a reduced number of samples the majority of the antibiotics have a decrease in the test fold performance, indicating that a reduced number of samples in this case can cause an overfitting of the classifier, since the test fold is not representative anymore. The curve of the test performance for all models is reaching an asymptote indicating that increasing sample sizes further would be unlikely to significantly improve model performance. Moreover, on S3B Fig, which creates synthetic data for the minority class during the pre-processing, reveals a similar result to the first approach in S3A Fig. The shape of the learning curves is kept similar between the two studies and it confirms for all the antibiotics that an increase in the number of samples would not have a big impact in the performance when compared to the original number of samples.

Since our goal was not only to predict resistance with high accuracy, but to also extract key insights from the data, we asked whether the uncovered genetic features are either true determinants of AMR or possible artefacts of the statistical learning algorithm. Therefore, we mapped the k-mers back to the sequences of the whole cohort to identify the genes to which the k-mers from each model were related to. Across all models the k-mers mapped back to 2949 unique genes of which 476 were known AMR-associated genes (held in the CARD [38], ResFinder [39], mutationDB [49], NCBI AMRFinder [41] or ARG-Annot [40] databases). To further identify the most significant genes we identified those with the greatest difference between the percentage of resistant isolates identified as having a k-mer hit to the gene (Pr) and the percentage of susceptible isolates identified as having a k-mer hit to the gene (Ps). Genes with a difference |Pr–Ps| in frequency equal or higher than 30% over the whole cohort, were selected and identified as important to differentiate resistant and susceptible isolates (S3 Table). Altogether, 361 genes were significantly differentiated between resistant and susceptible isolates. Of these 361 genes, 35 were present in three or more antibiotic classes in the predictive antibiotic models (S3 Table). These 361 genes were analysed for their distribution and frequency among the different hosts in both the farm and slaughterhouse environments (S3 Table).

We discuss the results obtained for three antibiotics representing two important classes—beta-lactams (penicillins and cephalosporins) and aminoglycosides—selected because of their clinical relevance in China. Results for the other 18 antibiotics are given in S4 Fig.

The ampicillin/sulbactam (penicillins) model identified 20 genes as significantly separating resistant and susceptible isolates, and of these 18 were mostly present in the resistant isolates (S3 Table). These included six known AMR-associated genes (lysU [49], lptD [49], aadA2 [38], aph(4)-Ia [38], floR [38], bla_OXA-10 [38]). The floR gene was present in > 92% of resistant farm samples but in less than 32% of resistant slaughterhouse samples (chi-square p value < 0.0001). Of the known AMR-associated genes aadA2, floR and bla_OXA-10 had multiple k-mers associated with each gene (16, 7 and 4 respectively), statistically unlikely to occur in each gene by chance (binomial test p values < 0.0001) whilst the others has only 1 association (mean p value = 0.02). Among the 14 genes associated with either a resistant or susceptible phenotype but that are not known to confer antimicrobial resistance based on their annotations (alx, hcaB, hdfR, hha, nadB, tnpR, traC₁, traC₂, traV, yciC, iroB, iroN, iucC and fes): four had functions associated with HGT (tnpR, traC₁, traC₂ and traV [50,51]); and 4 were virulence genes involved in siderophore-mediated iron transport (iroB, iroN, iucC and fes [52]).

For cefepime (cephalosporins) resistance/susceptible model (S3 Table) 18 genes (aadA2, floR, bla_CTX-M-55, cai, cusF, ecpD, exuT₁, exuT₂, finO, aph(4)-Ia, mdtB, nagE, aph(3”)-Ib, papC, tdh, traI, traQ, traS and yfeZ) were significantly associated with the resistance phenotype and a further three with the susceptible phenotype (mutL, yehL and yehA). Seven genes (mutL [49], aadA2 [38], aph(3”)-Ib [38], aph(4)-Ia [38], bla _CTX-M-55 [38], floR [38] and mdtB [38]) are known AMR-associated genes. Interestingly, floR was significantly more present in the resistant chicken isolates, 77%, compared to the resistant human isolates (from farm and slaughterhouse), 28%, (chi-square p value < 0.001). For the ESBL resistance gene bla_CTX-M-55 however no statistically significant difference was observed and no co-occurrence of these two resistance genes in the same contig was observed (Jaccard coefficient 0.0). The gene mutL, a DNA repair gene, which was the only known resistance gene to be more prevalent in resistant human isolates (67%) compared to chicken isolates (33%), chi-square p value = 0.13. The known resistance gene bla _CTX-M-55 had 17 unique k-mers differentiating the susceptible and resistance phenotypes predicted to be associated with it, unlikely to arise by chance (binomial p value < 0.0001), whilst two genes (aph(3”)-Ib and floR) had between 3 and 6 k-mers (binomial p values > 0.0001) and four genes (mutL, ecpD, aadA2, aph(4)-Ia) had only a single k-mer related to them. A pattern of higher prevalence in chicken isolates was observed in several of the genes found by the machine learning but not annotated to have AMR function: nagE, cai, exuT₁, tdh and papC (chi-square p values <0.01). One gene was a known virulence-related pilus gene (ecpD [53]). Not surprisingly, none of these genes were found to significantly co-occur with each other or the known AMR-associated genes.

The kanamycin (aminoglycosides) resistance/susceptible model (S3 Table) showed that two genes tetA and rihA were significantly associated with susceptible isolates whilst 26 were associated with resistant isolates (aph(3’)-Ia, aph(3”)-lb, aph(4)-Ia, bla_CTX-M-55,floRr, bla_TEM-1, bla_OXA-10, der, hdfR, mntB₁, mntB₂, mntB₃, dcm, tfaE, traC₁, traC₂, sopB, traD, ylpA, hcaB, traQ, cai, cmi, higB-1, dsbC, and repB). Eight of these were known AMR-associated genes, present in the CARD database (aph(3’)-Ia, aph(3”)-Ib₂, aph(4)-Ia, bla_CTX-M-55, floR, bla_TEM, bla_OXA-10 and tetA [38]). Of these aph(4)-Ia frequently co-occurred with the aph(3’)-Ib and aph(3”)-Ib (Jaccard coefficients of 0.51 and 0.20) but no other significant co-occurrence was observed. The genes were generally present in similar proportions of chicken and human isolates. The known AMR genes had significantly more unique k-mers associated with them compared to the genes not known to confer AMR with a mean of 17 compared to a mean of 4 (t-test p value < 0.01). All the known AMR genes has a statistically significant number of k-mers, unlikely to occur by chance (binomial p value <0.0001 for all except aph(4)-Ia with p value < 0.01). In addition, several of these genes, were also unlikely to be spurious hits: der (p value < 0.0001), hdfR (p value < 0.0001), mntB₁ (p value < 0.0001), mntB₂ (p value < 0.0001), mntB₃ (p value < 0.0001), dcm (p value < 0.0001), tfaE (p value < 0.0001), tetA (p value < 0.0001), traC₁ (p value < 0.0001), traC₂ (p value < 0.0001), sopB (p value < 0.0001), traD (p value < 0.001), ylpA (p value < 0.0001), hcaB (p value < 0.0001)).

In addition to the k-mer mapping to the annotated genes of the cohort, many hits were found to map to intergenic regions of the DNA sequences. We analysed these hits to identify any patterns and found that the k-mers mapped to 14 different transcription factor binding sites (arcA, argR2, crp, fis, fur, ilvY, lexA, lrp, nagC, phoB, purR, rpoH2, rpoN and tus) where there was greater than 30% difference between hits to resistant and susceptible isolates, S5 Fig. Of particular interest, the transcription factors rpoH2 and crp were significant in 5 different antibiotics: ampicillin, ampicillin/clavulanic acid, ampicillin/sulbactam, levofloxacin and sulfisoxazole; ciprofloxacin, chloramphenicol, ampicillin/sulbactam, doxycycline and cefoxitin, respectively. Both these transcription factors are global transcriptional regulators binding regulating many genes. The well-studied regulator Crp (cyclic-AMP receptor protein) regulates over 180 different genes, primarily in energy-metabolism pathways [53]. Whilst the alternative sigma factor RpoH is involved in regulating the cellular response to heat shock, increasing bacterial survival at high temperatures [54]. The gene purR, found in 4 antibiotic resistance/susceptible predictive models (cefepime, aztreonam, cefotaxime and cefoxitin) is involved in the regulation of genes in the purine metabolic pathway. Importantly, purine metabolism and synthesis pathways have been previously found to be enriched in multidrug resistant bacteria [22]. The transcription factor fur present in the cefotaxime, ampicillin, ampicillin/clavulanic acid and doxycycline models, represses genes involved in high affinity iron sequestration important in virulence [55] and is important in conferring protection from oxidative stress [56]. The gene tus, involved in the termination of DNA replication, was significant in doxycycline, ampicillin/sulbactam, minocycline and streptomycin. All other transcription factors were significant in 3 or fewer models.

Network analysis powered by the machine learning uncovers communities of similar isolates and reveals ARGs sharing between hosts

To map the reservoirs of possible antibiotic resistance related genes that pose the highest risks of horizontal gene transmissibility across humans, animals and their environment, we built gene sharing networks and a hierarchically clustered heatmap (clustermap) for each one of the 21 selected antimicrobials (Figs 3, 4 and S4). The networks were analysed to assess modularity and betweenness centrality. In gene sharing networks, nodes with higher betweenness centrality are potentially more interesting as they are involved in a larger number of associations.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Networks of AMR-associated genes shared across host species (animals and workers) and environments (farm and slaughterhouse) for three antibiotics.

The AMR-associated gene sharing networks for: (A) ampicillin/sulbactam, (B) cefepime, and (C) kanamycin are shown. For each network: (i) Panel on the left, indicates the communities and their respective numbers found using Louvain heuristics. Each community is indicated with a distinct colour. Each community is a set of nodes or clusters, which are densely and connected with statistical significance and identically coloured. For each network the node represents a sample and is shown in a distinct colour and shape depending on the metadata of the sample (e.g., source, AMR profile, location). In all panels the nodes are separated according to the Euclidean distance between the isolates; the central panel shows the AMR phenotype (resistant or susceptible) as the shape of the node and the source of the samples (human, chicken or environmental) as the colour of the node. An edge represents the Euclidean distance between two samples, and it is coloured according to the associated statistical value. The panel on the right indicates the location of the sample (farm or slaughterhouse) as the shape of the node and the type of sample as the colour of the node.

https://doi.org/10.1371/journal.pcbi.1010018.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Hierarchically clustered heatmaps of the AMR-associated genes across the cohort.

Clustered heatmaps showing the genes associated with the significant k-mers from the machine learning and used to build the networks. The antibiotic models shown are: (A) ampicillin/sulbactam, (B) cefepime and (C) kanamycin. The columns on the left show the metadata (Class, Location, Source, Type, Phylogroup, Community number). The presence of the genes (in terms of its related k-mers) is indicated in black, while the absence is indicated in white). Genes known to confer antimicrobial resistance based on their annotations from public databases are highlighted in red.

https://doi.org/10.1371/journal.pcbi.1010018.g004

In our dataset, we found an extensive network of genes correlated with AMR phenotypes shared between E. coli communities of humans, animals, meat and the environment, that was not seen in the SNP network analysis.

The ampicillin/sulbactam gene sharing network (Fig 3A) showed four different communities. Communities C0, C2 and C3 contained most resistant isolates (86%) whilst community C1 contained most of the susceptible isolates (83%), almost all of which were from human hosts (18/20). The three resistant communities found were clearly separated from the susceptible community. Samples from resistant community C0 were primarily chicken and environmental samples. Six resistant human isolates (nose and hand samples) were present in C2 and of these, all except one were identified as pathogenic EPEC or EAEC strains. These were intermixed with chicken isolates that were either pathogenic or non-pathogenic. Only 40% (16/40) of human isolates were resistant to ampicillin/sulbactam compared to 95.1% (58/61) chicken isolates, and of the human isolates hand and nose swabs were more often resistant than faecal isolates (56% compared to 27%). The presence of the genes lysU and lptD and the absence of the genes floR and hdtR were important in differentiating susceptible from resistant isolates (Fig 4A). The floR gene is known to be involved in sulphonamide (sulfathiazole) and bicyclomycin resistance [57,58], and was present in all resistant communities. The analysis of the communities C0 and C2 indicates the co-presence of aadA2 (streptomycin resistance), bla_OXA-10 (beta-lactamase) and aph(4)-Ia (phosphorylates the antibiotic hygromycin B) which are all known AMR-associated genes, however these are not co-occurring in the same sequence contigs. In communities C0 and C3, among the genes not annotated as conferring AMR we observed the co-presence of the following virulence genes: iroB, iucC, iroN and iroD, of which all except iucC frequently co-occur on the same contig. The five nodes with highest betweenness centrality were found to be three chicken isolate (one caecal, one faeces and one anal swab) and two human isolates (hand swabs from the farm and slaughterhouse).

The cefepime network (Fig 3B), contains three communities (C2, C3 and C4) where most of the isolates (humans, chicken, and environment) show resistance to this antibiotic (77%), and two communities (C0 and C1) where most isolates are susceptible (76%). The two primarily susceptible communities are clustered together, but the resistant communities show greater separation. Community C4 was populated with mostly pathogenic samples (13/18), indicating the higher relatedness of these isolates. According to the clustermap analysis (Fig 4B), the AMR-associated genes (aph(3”)-Ib, floR, mdtB and bla_CTX-M-55) are clustered together; with aph(3”)-Ib and floR presenting mainly in C2 and C3 and bla_CTX-M-55 and mdtB present across communities C2, C3 and C4. Additionally, the resistant and susceptible isolates also differ by the presence of the genes yehL (ATPase) and yehA (fimbrial-like adhesin) genes in susceptible isolates. For cefepime, the five nodes with highest betweenness centrality were four chicken samples (one carcass, two faeces and one anal swab) and one environmental (slaughterhouse water) which indicates that the main hotspot is the chicken samples.

In the kanamycin network (Fig 3C) there are three close communities mostly containing susceptible isolates (C0, C2 and C3, 78% susceptible) and the other four (C1, C4, C5 and C6) contained predominantly resistant isolates (91% resistant). Community C6 is formed by only pathogens. Interestingly, there are two major clusters for this antibiotic, one mainly related to resistant samples (communities C1, C4 and C6) and the other mainly related to susceptible samples but also containing resistant community C5. The genes bla_CTX-M-55, traC, repB, higB-1, dcm, aph(4)-Ia, bla_TEM-1, dsbC, hcaB and bla_OXA-10, are absent in community C5 but present in the other resistant communities (Fig 4C), with four of them known to confer antimicrobial resistance (bla_CTX-M-55, bla_OXA-10, bla_TEM-1 and aph(4)-Ia). The other known AMR-associated genes, aph(3’)-Ia, aph(3”)-Ib and floR were found present in all the resistant communities. We hypothesise that sharing of AMR-associated genes uncovered by ML is more frequent between the chicken isolates on the farm, since most human and environmental samples are susceptible. Further, 24 out of 38 slaughterhouse isolates were susceptible to kanamycin, indicating that this antibiotic was not significant in the slaughterhouse. For kanamycin, the five nodes with highest betweenness centrality are from multiple sources (one environmental, two humans and two chicken).

Out of the 361 genes that were selected with a |Pr-Ps| > 30%, 303 genes (278 unique excluding allelic variants) were not annotated as antibiotic resistance genes in public AMR databases. Therefore, to characterize these genes without annotated AMR function GO [59] and KEGG pathways [60] enrichment analyses were conducted (S4 Table and S6 Fig). The GO terms showed enrichment in one molecular function encompassing catalytic activity, and thirteen biological processes: metabolic, cellular, cellular metabolic, organic substance metabolic, monocarboxylic acid catabolic, organic cyclic compound metabolic, cellular aromatic compound metabolic, antibiotic catabolic, 3-phenylpropionate metabolic, 3-phenylpropionate catabolic, primary metabolic, small molecule metabolic, oxoacid metabolic and cellular response to chemical stimulus. In addition, four KEGG pathways were enriched: metabolic, microbial metabolism in diverse environments, phenylalanine metabolism and degradation of aromatic compounds (S4 Table and S6 Fig).

Identification of the mobile genetic elements associated to the interconnected resistomes and their E. coli hosts that are naturally transferred among animal, human and the environment

Although the gene sharing network analysis powered by machine learning found extensive interconnected resistomes among E. coli communities of humans, animals and environments within and between the farm and slaughterhouse, it remains unclear whether the uncovered interconnections are mediated by mobile genetic elements. We analysed the MGE content of the isolates in our cohort using the mob-suite package [61], then we grouped the isolates by using hierarchical clustering by replicon type. In total, 153 of the 154 isolates in our cohort carried at least one plasmid with 147 of those carrying a mobilizable or conjugative plasmid. Across the cohort, 17.18% of genes in the pan-genome were located on plasmids, with a range of 0–14.6% per isolate. Comparing different hosts, the proportion of genes carried on plasmids in chickens was significantly higher than in human farmworkers (adjusted p-value_Bonferroni = 0.00993) and abattoir household members (adjusted p-value _Bonferoni = 0.0403) (S7 Fig). The types of plasmid within the cohort were diverse (Fig 5), and did not show any significant clustering by host, collection time point or source.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Presence/absence of different plasmid types in the cohort.

Heatmap showing the presence (green) and absence (grey) of each predicted plasmid replicon type per isolate. Phylogenetic tree (replicated from Fig 1B) is at the top of the heatmap with the phylogroup. Sequence types, hosts and isolate sources are shown below the tree.

https://doi.org/10.1371/journal.pcbi.1010018.g005

Next, we tied the above results with the information obtained with machine learning. Replicon types: IncFIA, IncFIB, IncI1/B/O, IncHI1B, IncK, IncR and IncY, previously associated with multidrug resistance [62] carried between 2.57 and 8.45% of the total 4626 genes associated with resistant phenotypes identified by the supervised learning analysis (Fig 6A). Replicon types IncX1 (1.25%) and IncN (1.64%) carried fewer genes, whilst IncI2, IncX4, Col156, ColpVC and Col(MG828), also reported in literature to carry resistance genes [62], carried very few (>1%) of the predicted genes from this study (S5 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6.

Presence of AMR-associated genes predicted by machine learning in plasmids and hierarchical clustering of isolates based on similarity of plasmid type (A) Heatmap of presence (pink) or absence (blue) of genetic determinants identified by machine learning located on different replicon types of plasmid within the cohort. (B) Hierarchical clustering of isolate plasmid sequences from replicon types with high number of AMR determinants. Clustering was based on gene presence or absence. Green underscores indicate areas of high similarity between isolates from different hosts, suggestive of transmission between hosts.

https://doi.org/10.1371/journal.pcbi.1010018.g006

In addition to considering the AMR-associated genes uncovered by machine learning, we also annotated the plasmid sequences to assess the gene content of each plasmid replicon. Many known ARGs were carried on multiple plasmids (e.g. aph(3’)-Ia [n = 6], aph(3”)-Ib [n = 6],bla_CTX-M-55 [n = 5], aadA2 [n = 5], tetR [n = 5]), S6 Table. Notably, IncI2 also carried resistance gene bla_CMX-M-97 and colistin resistance gene mcr-1, an association that have previously been noted to give a fitness advantage via carriage of this gene [63]. Hierarchical clustering of the presence/absence of all genes on each replicon type was conducted to assess the similarity of plasmids from different isolates. We hypothesise that isolates with plasmids clustering with only small differences are indicative of relatively recent transmission of those plasmids between isolates through horizontal transfer. Where those plasmids are present in different host/isolates this would be suggestive of a transmission route for these bacteria.

Although the majority of the plasmids shared predominantly different mobile elements between livestock and humans, we found 17/58 (29.3%) human isolates sharing closely related AMR-associated mobile elements with those found in livestock. Most of the clusters contained different STs which is also indicative of horizontal gene transfer of mobile genetic elements between lineages [37]. Within the IncFIA group there were several small clusters of isolates, Fig 6B. One cluster contained 2 human farmworker nose swabs and an environmental animal feed isolate. Average nucleotide values (ANI) of the WGS of these isolates showed high similarity over >99.995%, compared to a mean ANI value across the cohort of 98.08%. IncFIA is an important replicon for AMR in this cohort and contained several known resistance genes namely: aadA2, tetA, tetR, cmlA5, bla_TEM, bla_CTX-M-97 and floR. Within the IncI1/B/O replicon group, a chicken carcass sample (non-pathogenic, phylogroup A, ST206) showed high similarity to the hand swab of a household member of an abattoir worker (non-pathogenic, phylogroup G, ST174) (ANI value of WGS = 99.995%) and two chicken faeces samples (non-pathogenic, phylogroups B1 and F, ST101 and ST457 respectively) showed high similarity to a farmworker hand swab isolate (non-pathogenic, phylogroup F, ST457) (ANI values of WGS: 99.995% and 97.1423%). In IncHI1B chicken faeces isolates (non-pathogenic, phylogroup B1, ST205) have clustered with a farmworker nose swab isolate (non-pathogenic, phylogroup B1, ST205) (ANI values of WGS > 99.994%). As with IncFIA, IncHI1B carries many known AMR genes: cat, bla_OXA-1, bla_CTX-M-97, aadA2, cmlA5, aph(3”)-Ib, tetA, and tetR. In IncFIB, carrying known resistance genes tetA, tetR, cmlA5, aph(3”)-Ib, bla_TEM, and floR, a chicken carcass plasmid (non-pathogenic, phylogroup A, ST206) clustered with a hand swab of a household member of an abattoir worker (non-pathogenic, phylogroup A, ST206) (ANI value of WGS = 99.996%). All the IncR plasmids (6 isolates, all EAEC positive, phylogroup A, ST542) show high similarity and except for one isolate are all sourced from chicken carcass samples. The other isolate was taken from the water at the abattoir, suggesting a possible transmission route between the chicken carcass and abattoir water system. Other isolates with less significant AMR genes but also sharing plasmids with high pairwise ANI values > 99.996% and showing inter-source clustering, again indicate direct transmission of the bacteria (S7 Fig). We found no evidence of human-to-human bacterial transmission in the farm environment.

Ethics statement

A key component of the programme of work was the recruitment of human faeces, hand swabs, nose swabs, chicken faeces and cloacal swabs and chicken carcasses for conventional microbiological analysis for presence of Escherichia coli.

Ethical approval has been obtained from the Research Ethic Committee in the School of Veterinary Medicine and Science at the University of Nottingham, and from Food Safety Risk Assessment, China National Center for Food Safety Risk Assessment, Beijing, China and assigned the following application IDs: 2340 180613 and 2018018.

Sample collection and bacterial isolation

A total of 284 samples were taken over two independent 6-weeks broiler production cycles (March to June 2019) from a farm and connected slaughterhouse in Shandong province, China. The farm implements self-breeding uses the closed-end management model and contains on average 12000 chickens. A total of 284 isolates, were collected and included: soil surrounding the chicken barn (n = 12), chicken drinking water (n = 10), chicken feed (n = 16), farmworker faeces (n = 21), veterinary surgeon faeces (n = 3), farmworker hand swabs (n = 20), veterinary surgeon hand swabs (n = 3), worker nasal swabs (n = 20), veterinary surgeon nasal swabs (n = 3), chicken faeces (n = 60), chicken caecal droppings (n = 12), chicken cloacal swabs (n = 10), abattoir operator faeces (n = 11), abattoir operator nasal swabs (n = 11), abattoir operator hand swabs (n = 11), abattoir operator household faeces (n = 8), abattoir operator household nasal swabs (n = 8), abattoir operator household hand swabs (n = 9), abattoir waste water (n = 4), chicken carcass (n = 32). Connectivity of the samples is explained below for the E. coli positive samples. Farmed chickens were treated with kanamycin (aminoglycoside) for the first 5 days, tylosin (macrolide) from day 6 to 9, neomycin (aminoglycoside), from day 12 to 15, and florfenicol (chloramphenicol) from day 18 to 21 in the first production cycle and with amoxicillin (penicillin) from day 5 to 8 and neomycin (aminoglycoside) from day 9 to 12 in the second production cycle. To isolate E. coli strains an approximately 10g mixed fresh sample of 2–3 chicken faeces or an approximately 2g mixed fresh sample of 2–3 chicken caecal droppings was collected from the bottom of the chicken cage using a sterilized spoon. Chicken cloacal swab samples were collected using cotton-tipped swabs (108C.USE, Copan Diagnostics, Italy) from the same cage. Chicken carcass samples were collected in abattoir using sponge swabs (SS100NB, Hygiena International, Watford, UK). About 10g soil surrounding the chicken barn was collected at depth of 1-3cm within a distance of 5m from chicken barn. Not less than 20mL each of chicken drinking water of farm and waste effluent water from the slaughterhouse were collected from the water pipe or by using pipettes. About 10g each feed sample was collected using a sterilized spoon. Sterilized sampling spoons were used to collect 8g each of faeces from humans including chicken farm workers, the poultry veterinarian, slaughterhouse workers and slaughterhouse worker households. The hand swab sample and nasal swab samples were also collected using cotton-tipped swabs. All samples were collected using aseptic techniques, and then kept in secure containers at 4°C during transportation to the laboratory and analysed within 24h.

Bacteria isolation and diarrheagenic E. coli (DEC) identification

A quantity (volume) of 1 g (mL) sample of faeces, soil, feed, drinking water and polluted water each was vortexed with 9 mL of sterile buffered peptone water tube (BPW; Luqiao Inc., Beijing, China) for 1 min; nasal swab samples, hand swab samples and cloacal swab samples were vigorously vortexed with 9 mL BPW for 1 min in a test tube; chicken carcass sponge samples were homogenised with 10 mL BPW for 1 min in a stomacher bag. Approximately 1mL dilution (of any of the above samples) was then added to 9 mL E. coli (EC) broth (Luqiao Inc.) and incubated at 37°C for 16-20h in order to enumerate presumptive E. coli populations. A loopful of these solutions was then streaked onto an eosin-methylene blue (EMB) agar and MacConkey (MAC) Agar (Luqiao Inc.) and incubated at 37°C for 18-24h. Typical E. coli colonies were screened and subsequently characterized by a Bruker MALDI Biotyper (Germany). The positive isolates identified were further confirmed by PCR using E. coli-specific primers ITS-F (5’-CAATTTTCGTGTCCCCTTCG-3’) and ITS-R (5’- GTTAATGATAGTGTGTCGAAAC-3’). Thermal amplification conditions were as follows: pre-incubation at 94°C for 5 min, followed by 30 cycles of denaturation at 94°C for 30s, annealing at 55°C for 30s, elongation at 72°C for 30s, and a final extension of 72°C for 5 min. Out of 284 isolates analysed, 154 were identified as E. coli positive and included: chicken faeces (n = 49), farmworker faeces (n = 19), chicken carcasses (n = 15), cloacal (anal) swabs (n = 9), chicken caecal dropping (n = 9), farmworker hand swabs (n = 9), farmworkers nasal swabs (n = 8), abattoir operators faeces (n = 7) and their households’ faeces (n = 7), animal feed (n = 6), abattoir polluted water (n = 3), abattoir operators’ hand swabs (n = 3), soil surrounding the chicken barn (n = 3), abattoir operator households’ nasal swabs (n = 2), abattoir operator households’ hand swabs (n = 2), chicken drinking water (n = 2), abattoir operators’ nasal swabs (n = 1). Chicken faeces samples were taken from small cages containing up to 10 chickens (cage ID numbers given in S1 Table), with up to two pooled samples taken per cage (each pooled sample containing 2–3 faeces). Sampling in this way gave a 5.2–11.4% probability of sampling the same chicken more than once at any time point (18.7–34.1% over both time points). A single pooled caecal sample containing 2–3 individual chicken samples gave rise to a 1.0–2.8% probability of sampling the same chicken more than once (5.2–11.4% over both time points). Cloacal swabs were taken from individual chickens and were all independent. Finally, humans were sampled at multiple body sites once per time point (human ID numbers given in S1 Table).

Confirmed E. coli strains were screened for five diarrheagenic pathotypes, EAEC (enteroaggregative E. coli; aggR and/or astA and/or pic), EPEC (enteropathogenic E. coli; escV), STEC (shiga toxin-producing E. coli; stx1 and/or stx2), ETEC (enterotoxigenic E. coli; lt and/or stp and/or sth), and EIEC (enteroinvasive E. coli, invE), by the DEC multiplex PCR kit (Meizheng, Inc., Beijing, China).

Antimicrobial susceptibility testing

Antimicrobial susceptibility to a panel of agents was determined by broth microdilution and interpreted according to the criteria based on the Clinical & Laboratory Standards Institute (CLSI) interpretive criteria (CLSI 2009). The minimum inhibitory concentrations (MIC) of 26 antimicrobial compounds were measured for the 154 E. coli isolates, and these included ampicillin (AMP), ampicillin/sulbactam (AMS), tetracycline (TET), chloramphenicol (CHL), trimethoprim/sulfamethoxazole (SXT), cefazolin (CFZ), cefotaxime (CTX), ceftazidime (CAZ), cefoxitin (CFX), gentamicin (GEN), imipenem (IMP), nalidixic acid (NAL), sulfisoxazole (SUL), ciprofloxacin (CIP), amoxycillin/clavulanic acid (AMC), polymyxin E (PE), polymyxin B (PB), minocycline (MIN), amikacin (AMK), aztreonam (AZM), cefepime (FEP), meropenem (MEM), levofloxacin (LEV), doxycycline (DOX), kanamycin (KAN) and streptomycin (STR). E. coli ATCC 25922 was used as a control for the antimicrobial susceptibility testing. The MIC50 value (the MIC required to inhibit 50% of cells), MIC90 value (the MIC required to inhibit 90% of cells), resistance rate and MIC distribution were calculated.

DNA purification and extraction

The identified E. coli isolates were kept in brain heart infusion broth (BHI) medium with 20% glycerol at -80°C and genomic DNA (gDNA) was purified using Omega EZNA Bacterial DNA Kit (Omega Bio-tek, GA, USA). All the E. coli isolates were subjected to genomic DNA extraction in accordance with the manufacturer’s protocol of E.Z.N.A. Bacterial DNA Kit (Omega Bio-Tek, Norcross, GA, USA).

Library construction and whole genome sequencing

The template genomic DNA was fragmented by sonication to an insert size of 350 bp using NEBNext Ultra DNA Library Prep Kit for Illumina (NEB, USA) following the manufacturer’s recommendations and index codes were added to attribute sequences to each sample and sequenced using an Illumina Hiseq 2500 PE150 achieving a median genome coverage of 270x (Range 140x-380x).

Genome assembly and annotation

All sequences were pre-processed through readfq v10. To clean the data, reads containing low-quality bases (mean quality value ≤ 20) over 40% were removed. Reads with greater than 10% unidentified bases (N) were removed as well as the adapters. The whole-genome shotgun sequencing produced 154 high-quality reconstructed E. coli genomes with a N50 larger than 50,000 and less than 250 contigs (S1 Table). Cleaned data were processed for genome assembly with SPAdes v3.13, and QUAST v4.5 was used for assessing the assembly. The contigs with length shorter than 500 nucleotides were filtered out. The completeness and contamination of genomes were assessed through checkM with the lineage_wf pipeline. Genomes were annotated with Prokka v1.14.5 [90] using default parameters with—addgenes–usegenus.

Screening of annotated genes against ABR databases

The whole genome sequences were screened against the CARD [38] database with a minimum coverage and identity of 90% to identify known AMR-associated genes in the isolate cohort. In addition the annotated genes obtained from Prokka v1.14.5 [90] were screened against the CARD [38], ARG-Annot [40] and Resfinder [39] databases using Abricate [91] and the NCBI AMRfinder [41] database; all the comparative analyses have been done with a minimum coverage and identity of 90%. For some genes identified in the AMR databases, the gene name differs from the name assigned by Prokka, due to synonymous genes. In these cases, we have opted to use the more widely recognised gene name as reported in the AMR databases throughout the text, though the original name can be found in S3 Table. Additionally, gene names were screened against the MutationDB database that records genes reported in literature to have undergone mutations due to antibiotic stress [49]. Where available, gene accession numbers were added to S3 Table.

In silico subtyping identification and standard index of association (I_A) calculation

Sequence types were identified through MLST which mapped the sequences to the PubMLST E. coli MLST database. Clonal complexes were annotated from known CC types in the MLST database. Phylogroups were identified using in silico Clermont typing [92]. Serotypes were identified through the EcOH database [92] using Abricate [91].

Linkage disequilibrium of the 154 isolates in our cohort was evaluated using the standardized index of association (I^S_A) [43], which estimates the homologous recombination for the cohort by assessing the linkage disequilibrium among the seven MLST loci. The LIAN Ver. 3.7 program was used to calculate the I^S_A for all the isolates and for a subset of them (one isolate for each ST type) from the ratio of the variance of observed mismatches in the test set (V_D) to the variance expected for a state of linkage equilibrium (V_e), scaled by the number of loci used in the analysis (L) [43,93].

The significance of I^S_A was determined by a Monte Carlo simulation with 10⁵ resamplings.

Whole-genome phylogenetic analysis and network analysis based on pairwise SNP alignment

All annotated genomes were taken as input for pan-genome analysis with core gene alignments through Roary v3.13 [94]. IQTree v2.0.3 [95] was then used to construct the phylogenetic trees from the core genome alignment. Fifty-three different nucleotide replacement models were tested automatically with the GTR (+F+R5) replacement model selected as the best model. The Ultrafast bootstrap algorithm was used with 1000 replicates to assess branch support. The phylogenetic trees were subsequently visualised through iTOLv5 [96]. The core genome alignment was taken as input to produce a file of core gene SNPs present in the cohort using snp-sites [97]. A network of E. coli isolates collected from different sources in the farm (human, chicken and environmental) was created using a pairwise hamming distance comparison based on SNPs in the core genome. Each node represents an isolate while the edge represents the hamming distance between two isolates multiplied by the total number of SNPs found in our cohort (133,631 SNPs). A threshold of 15 or less SNPs difference was used to filter the edges in the network as suggested by Ludden et al (2019) [37].

Machine learning analysis

Machine learning methods were used to search for the features in the genome sequence of each isolate which could strongly correlate to resistance to each one of the of the 26 selected antimicrobials. Sample genomes were first split into overlapping 13-mers using GenomeTester [98] to produce a feature table for all samples. The AMR phenotype (resistant, susceptible or intermediate) of each sample (S1 Table) was used as the class label with intermediate phenotypes neglected. As the classes were unbalanced a synthetic minority oversampling technique [99] was used during the training phase of the classifiers to balance the proportion of classes in the data set. The Python package Scikit-learn [100] was used to make the classification and to select the most important features.

Fig 2A indicates the pipeline with the five steps used in this work. The Python package Scikit-learn [100] was used to make the classification and to select the most important features.

1. A two-step approach to overcome the disadvantages of applying a chi-square test in multiple comparisons is used: i) all the k-mers that have a p-value > 0.05 (the null hypothesis was not rejected) were discarded; ii) the ones with a p-value < 0.05 were used as input to an Extra Tree Classifier (randomized decision trees) and the k-mers with a Gini feature importance above the overall mean were selected.
2. A panel of machine learning methods (logistic regression (LR), linear support vector machine (L-SVM), radial basis function support vector machine (RBF-SVM), extra tree classifier, random forest, adaboost, xgboost, naïve bayes, linear discriminant analysis and quadratic discriminant analysis) were then run using as input the k-mers uncovered on the first step and their performances were evaluated.
3. A BLASTN approach on the whole genome sequence of each sample to identify the genes related to each k-mer selected on the first step.
4. The genes with a relative presence higher than 30% between resistant and susceptible samples are selected to be input for the clustermaps and gene sharing networks.
5. The genes selected in the previous step are evaluated in AMR public databases (CARD [38], ARG-Annot [40], ResFinder [39]) to confirm if they are known AMR genes.

Nested Cross-validation (NCV) [101] was employed to assess the performance and select the hyper-parameters of the proposed classifiers. In NCV, there was an outer loop split of the data set into test and training sets. For each training set, a grid search (inner loop) was run, to find the best hyper-parameters of the classifier using accuracy as a performance metric. Then, the test set was used to score the best classifier found in the inner loop. Thirty iterations were carried out, wherein each iteration an NCV was employed. The inner loop of the NCV found the best hyper-parameters of each classifier using stratified 3-fold cross-validation; the outer loop measured the ROC-AUC (receiver operating characteristic area under the curve) accuracy, sensitivity, specificity and Cohen’s kappa of the test data set (unseen in the inner loop for the training) using 5-fold stratified cross-validation, to compare all the classifiers [102]. A synthetic minority oversampling technique (SMOTE) was used to reduce the impact of unbalanced classes in the antimicrobial label groups. The number of splits in the nested cross validation and the number of k-nearest neighbours for SMOTE necessitated at least 12 samples in each class. The prediction metrics accuracy (TP+TN/(P+N)), sensitivity (true positive rate: TP/P), specificity (true negative rate: TN/N), Area Under the ROC Curve (AUC) and Cohen’s Kappa were used, as typically used to measure ML performance, see [22,30,103–105]. Violin plots from the Seaborn package [106] were used to show the final prediction metrics.

To compare the results obtained by the 10 different classifiers used, a Friedman Statistical F-test was employed. The Friedman test F_F with Iman-Davenport correction [107] is employed for statistical comparison of multiple classifiers as suggested by [108]. First, we rank the algorithms for each dataset separately, i.e., the best algorithm gets ranking 1, the second best ranking 2, and so on. In case of ties, average ranks are assigned. Next, we apply the F_F and verify if the null hypothesis is rejected. After, the post-hoc Nemenyi test [108] was used to find if there is a single classifier or a group of classifiers that differ in terms of their average rank after the F_F test has rejected the null hypothesis that the performance of the comparisons on the groups of data is similar.

To analyse if the antibiotics studied had enough samples to make the test representative, a wrapper backward selection (WBS) approach in terms of samples was used, i.e. if initially we have 100 samples we are first going to test the model using all 99 possible sample combinations (leaving one out), the worst performance is identified and the sample that was left out on this performance is removed for the next iteration. This is done until the minimum number of samples (12) is reached for one of the classes. This minimum number of samples is a requirement for the SMOTE approach used in the classification framework.

Annotation and abundance of the ARGs in the farm and slaughterhouse

Where the machine learning was able to predict the antimicrobial class based on k-mers, these were then used to search the genome for genes that contained the k-mers. The k-mers were mapped to the pangenome of the 154 isolates using a BLASTN [109] query with the parameters: e-value = 1000, word-size = 13, gap-open = 0, gap-extend = 0, out-fmt = 5. Genes with an identity of >70% and coverage of >70% were considered to be variants of the same gene and hence were discounted as duplicates, as done in previous literature [110], however a more stringent threshold was used to ensure all gene variants were accounted for. The k-mer hit count (how many k-mers mapped to each identified gene) of the genes identified was then assessed for statistical significance at a significance level of 0.05 using a binomial exact test, with the probability of a gene hit based on the length of the gene and number of k-mer combinations possible per gene. To analyse the co-presence of the aadA2 gene, a BLASTN search was used to find contigs in the WGS isolates with the aadA2 sequence as the query and a 95% identity threshold. Genes present in contigs matching to the aadA2 sequence were selected from the annotated whole genome sequences and summarised by isolate.

To select the most relevant genes we analysed their presence in the resistant (Pr) and susceptible (Ps) isolates. Where a gene was considered as present in a sample if at least one k-mer recognised as discriminant for the specific AMR profile by the classifiers could be mapped to that specific gene. A gene was considered not present in a given sample if all the discriminant k-mers associated to that gene were absent in that specific isolate. If the difference |Pr–Ps| was higher than δ the gene was selected and identified as important to differentiate resistant and susceptible isolates; in our case δ was equal to 30%. To analyse the co-occurrence of the genes that considered to be the most relevant the Jaccard coefficient was calculated in a pairwise manner between the two vectors of gene presence/absence (binary data) in each assembled contig within the cohort.

Functional enrichment analysis

The functional characterisation of the genes not annotated in AMR public databases that presented a |Pr-Ps| > 30% was done using StringDB v11.5 [59], that scans against both the Gene Ontology (GO) [60] terms (biological process and molecular function) and Kyoto Encyclopaedia of Genes and Genomes (KEGG) pathways [111]. All strains of the Escherichia genus present in StringDB were queried and E coli K12 MG655 gave the best match with 211 out of 278 matched genes (75.9%) while E coli O157:H7 gave a match to 200 out of 278 genes (71.9%). The functional enrichment analysis was run on both E. coli strains with default settings. The hypergeometric test with false discovery rate correction were used to analyse statistically the GO and KEGG pathways analysis. To visualize the results, a chord diagram was produced using the library nxviz v0.6.3 in python.

AMR gene sharing network analysis for the chicken farm and slaughterhouse environments

To identify the hot-spots of resistant bacteria and AMR-conferring genes and investigate their transmissibility across humans, animals and their environment, within and between farm and slaughterhouse environments a novel graph network was constructed based on the work of Bernard et al. (2016) [112]. The AMR gene sharing network for each one of the 21 selected antimicrobials was built using the ARGs recognised as discriminant by the machine learning approach and fulfilling the |Pr–Ps|>30% selection criteria. To data mine the graph network, a community detection algorithm was applied using the NetworkX [113] and the community python-louvain [114] libraries. A community is a set of nodes (cluster) that are densely connected. It is typical to identify social behaviours [115]. In order to find the communities in a graph, the algorithm partitions the nodes that maximise the modularity using the Louvain heuristics [116]. In the gene sharing network, each node represents a sample (colour and shape indicates the metadata associated with each sample), and an edge connecting two nodes represents the Euclidean distance between two samples based on the presence/absence of the discriminant k-mers related to the genes. The networks were analysed to assess modularity and betweenness centrality. Modularity measures the strength of division of a network into communities. It is usually expressed in the 0–1 range, and higher modularity values indicate a network with tightly connected samples within the communities, but sparse connections between communities. Betweenness centrality is computed for all the nodes in the network, and -for each node- quantifies how many paths connecting two other nodes pass through that node.

Plasmid reconstruction and hierarchical clustering

Plasmids were reconstructed using the MOB-recon algorithm in the MOB-suite package [61], using default settings. The presence/absence of each plasmid in the cohort was visualised in iTOLv5. Reconstructed plasmid sequences were searched using BLASTN to identify which genes from the WGS pangenome were plasmid located (search with the pangenome as a query with a 95% identity threshold, default settings). The 1113 genetic determinants of resistance were then checked to see if they were present on any plasmid in the isolate and assumed to be chromosome based if not present. Each reconstructed plasmid was annotated using Prokka v1.4.5 [90]. For each identified replicon type with more than 5 reconstructed plasmid sequences, gene presence/absence was analysed using Roary v3.13 [94]. The gene presence/absence output from Roary was used as input for hierarchical clustering of the reconstructed plasmids in R [117]. A binary distance matrix (scaled between 0 and 1) of the reconstructed plasmids was created then the plasmids were clustered hierarchically using the hclust function in R (with method = ‘Jaccard’). The heights of the clusters represent the Jaccard distance between plasmids.