Abstract
Metagenomics can unveil the genetic content of the total microbiota in different environments, such as food products and the guts of humans and livestock. It is therefore considered of great potential to investigate the transmission of foodborne hazards as part of source-attribution studies. Source-attribution of antimicrobial resistance (AMR) has traditionally relied on pathogen isolation, while metagenomics allows investigating the full span of AMR determinants. In this study, we hypothesized that the relative abundance of fecal resistome components can be associated with specific reservoirs, and that resistomes can be used for AMR source-attribution. We used shotgun-sequences from fecal samples of pigs, broilers, turkeys- and veal calves collected across Europe, and fecal samples from humans occupationally exposed to livestock in one country (pig slaughterhouse workers, pig and broiler farmers). We applied both hierarchical and flat forms of the supervised classification ensemble algorithm Random Forests to classify resistomes into corresponding reservoir classes. We identified country-specific and -independent AMR determinants, and assessed the impact of country-specific determinants when attributing AMR resistance in humans. Additionally, we performed a similarity percentage analysis with the full spectrum of AMR determinants to identify resistome signatures for the different reservoirs. We showed that the number of AMR determinants necessary to attribute a resistome into the correct reservoir increases with a larger reservoir heterogeneity, and that the impact of country-specific resistome signatures on prediction varies between countries. We predicted a higher occupational exposure to AMR determinants among workers exposed to pigs than among those exposed to broilers. Additionally, results suggested that AMR exposure on pig farms was higher than in pig slaughterhouses. Human resistomes were more similar to pig and veal calves’ resistomes than to those of broilers and turkeys, and the majority of these resistome dissimilarities can be explained by a small set of AMR determinants. We identified resistome signatures for each individual reservoir, which include AMR determinants significantly associated with on-farm antimicrobial use. We attributed human resistomes to different livestock reservoirs using Random Forests, which allowed identifying pigs as a potential source of AMR in humans. This study thus demonstrates that it is possible to apply metagenomics in AMR source-attribution.
Highlights
Due to the genetic nature of antimicrobial resistance, and the possibility of horizontal transfer of resistance genes between bacteria of different species, we propose here to assess transmission of antimicrobial resistant (AMR) by considering the abundance of genetic determinants of resistance derived from metagenomic sequencing
In order to compare the performance of HRF1 and HRF2, we investigated the percentage of concordance between the crisp class and the true class, the list of AMR determinants with positive MDA and the OOB errors at each country node, and the balanced accuracy at each terminal node
With the first hierarchical Random Forests model, HRF1, we identified a subset of AMR determinants that allows an accurate classification of a resistome into its animal reservoir, independently of the country of origin
Summary
Source-attribution estimates the proportion of human cases of a foodborne disease attributable to different reservoirs and/or vehicles of transmission, i.e., different sources (Pires et al, 2018). Microbial subtyping studies of source-attribution have successively contributed in many countries to distribute the burden of specific foodborne diseases by animal reservoirs (Hald et al, 2004; Mughini-Gras and van Pelt, 2014; Mughini-Gras et al, 2019; Pires et al, 2014; de Knegt et al, 2016; Thépault et al, 2018; Zhang et al, 2019). Attribution estimates can be obtained using different data inputs and distinct modeling approaches (Mughini-Gras et al, 2018), but traditionally many singlepathogen targeted studies have relied on frequency-matching models and phenotypic data (Hald et al, 2004; Mughini-Gras and van Pelt, 2014; de Knegt et al, 2016). However, with the greater availability of genotyping information of foodborne pathogens isolated from animals, food, the environment and clinical cases, population genetic models have become an increasingly popular choice (Thépault et al, 2018), and several model developments have been seen in order to accommodate whole genome sequencing (WGS) data (Cheng et al, 2013; Raj et al, 2014; Lees et al, 2019; Tonkin-Hill et al, 2019). At the same time, machine learning has been exploited as an alternative approach to perform source-attribution with WGS (Zhang et al, 2019; Munck et al, 2020). For example, a machine learning decision tree algorithm has been recently proposed for the attribution of Salmonella Typhimurium infections in humans using core-genome multilocus sequencing (Munck et al, 2020), with results comparable to those previously obtained when combining Multiple Locus Variable-number Tandem Repeat Analysis (MLVA) data with the Hald frequency-matching model (de Knegt et al, 2016).
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
