Merging Metagenomics and Spatial Epidemiology To Understand the Distribution of Antimicrobial Resistance Genes from Enterobacteriaceae in Wild Owls

Antimicrobial resistance (AMR) is a multifaceted problem that poses a worldwide threat to human and animal health. Recent reports suggest that wildlife may play an important role in the emergence, dissemination, and persistence of AMR. As such, there have been calls for better integration of wildlife into current research on AMR, including the use of wild animals as biosentinels of AMR contamination in the environment. A One Health approach can be used to gain a better understanding of all AMR sources and pathways, particularly those at the human-animal-environment interface. Our study focuses on this interface in order to assess the effect of human-impacted landscapes on AMR in a wild animal. This work highlights the value of wildlife rehabilitation centers for environmental AMR surveillance and demonstrates how metagenomic sequencing within a spatial epidemiology framework can be used to address questions surrounding AMR complexity in natural ecosystems.

ABSTRACT

Antimicrobial resistance (AMR) is a well-documented phenomenon in bacteria from many natural ecosystems, including wild animals. However, the specific determinants and spatial distribution of resistant bacteria and antimicrobial resistance genes (ARGs) in the environment remain incompletely understood. In particular, information regarding the importance of anthropogenic sources of AMR relative to that of other biological and ecological influences is lacking. We conducted a cross-sectional study of AMR in great horned owls (Bubo virginianus) and barred owls (Strix varia) admitted to a rehabilitation center in the midwestern United States. A combination of selective culture enrichment and shotgun metagenomic sequencing was used to identify ARGs from Enterobacteriaceae. Overall, the prevalence of AMR was comparable to that in past studies of resistant Enterobacteriaceae in raptors, with acquired ARGs being identified in 23% of samples. Multimodel regression analyses identified seasonality and owl age to be important predictors of the likelihood of the presence of ARGs, with birds sampled during warmer months being more likely to harbor ARGs than those sampled during cooler months and with birds in their hatch year being more likely to harbor β-lactam ARGs than adults. Beyond host-specific determinants, ARG-positive owls were also more likely to be recovered from areas of high agricultural land cover. Spatial clustering analyses identified a significant high-risk cluster of tetracycline resistance gene-positive owls in the southern sampling range, but this could not be explained by any predictor variables. Taken together, these results highlight the complex distribution of AMR in natural environments and suggest that both biological and anthropogenic factors play important roles in determining the emergence and persistence of AMR in wildlife.

IMPORTANCE Antimicrobial resistance (AMR) is a multifaceted problem that poses a worldwide threat to human and animal health. Recent reports suggest that wildlife may play an important role in the emergence, dissemination, and persistence of AMR. As such, there have been calls for better integration of wildlife into current research on AMR, including the use of wild animals as biosentinels of AMR contamination in the environment. A One Health approach can be used to gain a better understanding of all AMR sources and pathways, particularly those at the human-animal-environment interface. Our study focuses on this interface in order to assess the effect of human-impacted landscapes on AMR in a wild animal. This work highlights the value of wildlife rehabilitation centers for environmental AMR surveillance and demonstrates how metagenomic sequencing within a spatial epidemiology framework can be used to address questions surrounding AMR complexity in natural ecosystems.

INTRODUCTION

Antimicrobial resistance (AMR) is a complex global problem that poses a serious threat to human and animal health. While antimicrobial use primarily occurs in health care and veterinary settings and food animal production, there is substantial evidence that residual antibiotics and associated antimicrobial-resistant bacteria (ARB) and antimicrobial resistance genes (ARGs) can also spread into natural ecosystems (1–3). To date, many studies have confirmed the widespread nature of AMR in environmental resources and wildlife, including resistance of clinical relevance to treating infections in humans and domestic animals (4, 5). This is particularly true of wild animals living in close proximity to anthropogenic sources of AMR, such as wastewater effluent, landfills, livestock facilities, and agricultural fields fertilized with manure (6–11). Thus, wildlife may play an important role in the maintenance and dissemination of AMR (12, 13). Specifically, wild animals may act as secondary or amplifying reservoirs of ARB and ARGs that may be reintroduced into humans and domestic animals at human-environment interfaces, such as through contamination of animal feed or drinking water reservoirs (5). Wildlife may also act as vectors by disseminating AMR to new geographical regions (this is particularly the case with migratory birds) or provide novel opportunities for the horizontal transfer of ARGs to zoonotic pathogens (13, 14). However, a more detailed understanding of how and when wildlife impacts AMR can be achieved only via an integrated One Health approach, where all potential sources and pathways, including those at the human-animal-environment interface, are considered.

Given the potential for wild animals to affect AMR persistence, spread, and evolution and the predicted subsequent risks for public health, there have been calls for the better integration of wildlife into current research on AMR (12), including the use of wild animals as indicators or biosentinels of AMR contamination in the environment (4, 15). In particular, shared ARB, ARGs, and plasmids illustrate potential exchanges between humans, domestic animals, and environmental compartments (4, 16, 17). However, the spatial distribution of AMR in the environment is heterogeneous, and the importance of anthropogenic sources of AMR relative to that of other biological and ecological determinants remains poorly understood (13, 18, 19). This is due in part to logistical and ethical concerns with sampling wild animals, which can limit sample sizes and the collection of relevant metadata. Wildlife rehabilitation centers may represent a convenient resource of animals for AMR surveillance in natural ecosystems. Animals may be recovered from broad geographical areas, allowing for more comprehensive surveys, and biological samples and biometrics are relatively easy to collect upon admission. Indeed, a number of studies on ARB have utilized wildlife rehabilitation data, particularly from raptors, with promising results (20–22).

Owls are commonly admitted to rehabilitation centers worldwide and may be particularly good biosentinels for human-derived AMR in the environment, given their diet of small animals commonly exposed to agricultural animal feed, by-products, and waste and the potential for the trophic accumulation of resistance organisms via their prey (23–26). Further, owls are found across a broad variety of ecoregions, including wetlands, forests, grasslands, agricultural areas, and urban/suburban areas (27, 28), which provides an opportunity to explore the relationship between landscape variation and the spatial distribution of AMR.

Here, we conducted a cross-sectional study of ARGs collected from two species of wild owls—great horned owls (Bubo virginianus) and barred owls (Strix varia)—admitted to a large raptor rehabilitation facility in the midwestern United States. First, we used an enriched metagenomic sequencing approach to characterize acquired ARGs harbored by Enterobacteriaceae, a bacterial family that is thought to play an important role in the global spread of ARGs and that includes many medically important ARB, such as extended-spectrum-β-lactamase (ESBL)-producing Escherichia coli (29, 30). Shotgun metagenomic sequencing approaches have been used to characterize the collection of ARGs (referred to as the “resistome”) across a variety of microbial communities in humans, domestic animals, and environmental sources, such as soil and water (31–34). However, such sequencing often requires a significant sequencing depth to detect ARGs in low-abundance bacterial species (35). Selective culture enrichment of bacterial taxa of interest, combined with shotgun sequencing, allows for ARG profiling of multiple taxa within the selected group while avoiding the need for an excessive sequencing depth (36). We then used analytical tools from spatial epidemiology to identify potential spatial clusters of AMR and explore how anthropogenic landscapes shape the genetic structure and geographical distribution of ARGs. Since both great horned and barred owls are nonmigratory and typically have small, localized home ranges (37), we had the ability to link an ARG-positive owl to fine-scale land use data from the geographical area where they were recovered. Together, our results demonstrate how metagenomic sequencing within a spatial epidemiology framework can be used to broaden our understanding of AMR complexity in natural ecosystems.

RESULTS

A total of 78 cloacal swab samples from great horned owls and barred owls were enriched for Enterobacteriaceae, and total DNA was extracted. Shotgun sequencing generated over 66 million read pairs, with an average of 850,859 read pairs (range, 94,875 to 1,829,247) being generated per sample (see Data Set S1 in the supplemental material). On average, 85.7% of read pairs per sample (range, 70.3% to 92.4%) survived quality filtering (Data Set S1). Subsequent bioinformatic analyses of the sequences were performed to assign the taxonomy, identify the ARGs in each sample, and characterize the genetic context of each identified ARG (Fig. 1).

FIG 1.

Flowchart depicting the bioinformatic analyses performed in this study. Numbers in parentheses are reference numbers (72, 74, 76–78, 80, 83–86). Abbreviations: ARG, antimicrobial resistance gene; CDS, coding sequence; NCBI, National Center for Biotechnology Information; nr, nonredundant.

Taxonomy assignments by use of the Kraken program classified 97.4% of the reads (range, 60.0% to 99.9%) as belonging to bacteria, with the remaining 2.6% of reads being classified as either belonging to archaea or viruses or uncharacterized. As expected, the majority of bacterial reads were assigned to the family Enterobacteriaceae (mean, 96.7%; range, 47.0% to 99.4%). Escherichia coli was the most abundant species and was identified in 92.3% (72/78) of samples (Fig. 2). Plesiomonas shigelloides was dominant in 2.6% (2/78) of samples, and Escherichia albertii, Escherichia fergusonii, Klebsiella aerogenes, and K. oxytoca were each dominant in one sample. Interestingly, Salmonella enterica was identified in all 78 samples, but at a level several orders of magnitude lower than that for E. coli (mean normalized relative abundances, 0.0023 ± 0.0023 for S. enterica versus 0.59 ± 0.19 for E. coli).

FIG 2.

Heatmap displaying the relative abundance of bacterial species and the presence/absence of antibiotic resistance genes across 78 owl cloacal swab samples. Samples (rows) are clustered based on cumulative sum scaling-normalized bacterial species abundances, using the Euclidean distance with complete linkage. Only bacterial species with a relative abundance of ≥0.1 in at least one sample are included. ARGs are colored according to their major antimicrobial chemical class. Colored squares indicate the presence of a gene; white squares indicate its absence. ARGs that were identified via inclusion of multimapping reads are marked with asterisks (see “Antimicrobial resistance gene identification” in Materials and Methods for further details). Abbreviations: P. shigelloides, Plesiomonas shigelloides; S. pseudintermedius, Staphylococcus pseudintermedius.

ARGs were associated with a range of antimicrobial drug classes, but overall ARG prevalence was low.

To identify acquired ARGs, quality-filtered sequencing reads were aligned to ARG reference sequences, where a gene fraction of ≥80% was used to decrease the likelihood of false-positive ARG identification. Across all samples, ARGs were identified in only 23.1% of owls (18/78) (Fig. 2). Inclusion of multimapping reads for gene fraction calculations considerably improved ARG identification by nine additional ARGs in seven additional owls (note the asterisks in Fig. 2; see “Antimicrobial resistance gene identification” in Materials and Methods for further details). Among the ARG-positive owls, the median number of unique ARGs was 1 (range, 1 to 6 unique ARGs). A total of 17 unique ARG groups were identified, including those that encode resistance to the antimicrobial drug classes β-lactams, tetracyclines, fosfomycin, aminoglycosides, phenicols, trimethoprim, and sulfonamides. ARGs that confer resistance to tetracyclines and β-lactams were the most prevalent across samples (n = 8 samples), with the specific gene groups tet(A) and bla_TEM being identified in five and six samples, respectively. Of particular note, bla_CMY, which encodes an AmpC β-lactamase conferring resistance to cephalosporins, was identified in one sample. Two samples also carried the ARG pair oqxAB, which confers resistance to multiple antimicrobial drug classes, including fluoroquinolones.

De novo assemblies revealed that multiple ARGs were associated with commonly found MGEs.

To gain insight into the genetic environment of the identified ARGs, reads were de novo assembled and ARG-positive contigs were annotated, with a particular focus being placed on ARG colocalization and association with mobile genetic elements (MGEs). The average number of assembly contigs of ≥500 bp per sample was 925 ± 1,035.4, with a read mapping rate of over 99% (mean, 99.4% ± 0.6%) (Data Set S1). Twenty-one ARGs across 14 samples were identified in assembly contigs, including one ARG (dfrA17) that was not identified via read mapping due to a gene fraction of <80% and another two ARGs that were colocalized on the same contig [bla_TEM-1b and tet(A)] (Table 1). The majority of ARG contigs could be classified as either plasmid (8/21) or chromosomal (5/21), based on top BLAST hits and the presence of plasmid replicons or species-specific housekeeping genes. The remaining eight contigs matched equally well to both plasmid and chromosomal NCBI reference sequences and therefore were classified as “unknown.” Predicted plasmid types of contigs included IncFIA/FII in two samples and IncI1 and Inc1/FII in one sample each. Among the chromosomal ARG contigs, predicted taxonomic origins included E. coli, K. aerogenes, and K. oxytoca, a finding which was further supported by the presence of typically chromosomally encoded ARGs (i.e., fosA7 in Salmonella Heidelberg [38], fosA in K. aerogenes [39], and bla_OXY-2-6 in K. oxytoca [40]).

TABLE 1.

Genetic environment of antimicrobial resistance genes identified in assembly contigs

Sample	Resistance genea (GenBank accession no.)	Contig identifier	Size (bp)	DNA molecule type	Predicted origin	Associated mobile genetic element(s)b
MB95	ant(2″)-Ia (X04555)	NODE_2877	573	Unknown	Enterobacteriaceae	Unknown
MB132	aph(6)-Id (M28829)	NODE_452	1,177	Unknown	Enterobacteriaceae	Unknown
MB128	blaOXY-2-6 (AY077485)	NODE_8	273,128	Chromosome	K. oxytoca	None
MB132	blaTEM-1b (AY458016)	NODE_356	2,026	Unknown	Enterobacteriaceae	Tn2 transposon
MB134	blaTEM-1b (AY458016)	NODE_1997	958	Unknown	Enterobacteriaceae	Unknown
MB143	blaTEM-1b (AY458016)	NODE_2	77,719	Plasmid, IncFIA/FII	Enterobacteriaceae	Tn2 transposon
MB158	blaTEM-1b (AY458016)	NODE_13	140,727	Plasmid, IncFIA/FII	Enterobacteriaceae	Tn2 transposon
MB79	blaTEM-1b (AY458016)	NODE_14	35,371	Plasmid, IncI1	Enterobacteriaceae	Tn2 transposon flanked by IS200/IS605-like elements
MB81	dfrA17 (FJ460238)	NODE_2242	501	Unknown	Enterobacteriaceae	Unknown
MB132	dfrA5 (X12868)	NODE_137	11,916	Plasmid	Enterobacteriaceae	Class 1 integron downstream of TnAs1/Tn21 transposon
MB74	fosA (ACZD01000244)	NODE_10	159,439	Chromosome	K. aerogenes	None
MB103	fosA7 (LAPJ01000014)	NODE_5	272,244	Chromosome	E. coli	ISL3 family-like transposon
MB130	fosA7 (LAPJ01000014)	NODE_3	305,867	Chromosome	E. coli	ISL3 family-like transposon
MB150	fosA7 (LAPJ01000014)	NODE_11	137,767	Chromosome	E. coli	ISL3 family-like transposon
MB134	hugA (AF324468)	NODE_2049	928	Unknown	P. penneri	Unknown
MB132	sul2 (HQ840942)	NODE_391	1,708	Unknown	Enterobacteriaceae	Unknown
MB132	tet(A) (AF534183)	NODE_269	3,812	Plasmid	Enterobacteriaceae	TnAs1-like transposon
MB143	tet(A) (AF534183)	NODE_224	8,342	Plasmid	Enterobacteriaceae	TnAs1-like transposon
MB158	tet(A) (AF534183)	NODE_13	140,727	Plasmid, IncFII/FIA	Enterobacteriaceae	TnAs1-like transposon
MB78	tet(B) (AP000342)	NODE_60	6,905	Unknown	Enterobacteriaceae	Tn10/IS10 transposon
MB141	tet(B) (AF326777)	NODE_22	89,530	Plasmid, IncI1/FII	Enterobacteriaceae	Tn10/IS10 transposon missing IS10L element

^aAntimicrobial resistance gene identification was based on reference sequences from the ResFinder database (77).

^bSchematic representations of mobile genetic elements associated with ARGs are shown in Fig. S1 in the supplemental material.

Genetic features associated with both transposons and integrons were identified in 13 samples within close proximity to ARGs, including bla_TEM-1b, dfrA5, fosA7, tet(A), and tet(B) (Table 1; Fig. S1). Specifically, the bla_TEM-1b gene, conferring resistance to β-lactams, was located within the Tn2 transposon, the dominant bla_TEM-containing transposon in commensal E. coli isolates (41, 42). Among genes conferring resistance to tetracyclines, tet(A) and its repressor, tet(R), were associated with the Tn3-like TnAs1 transposon, whereas tet(B) and tet(R) were flanked by regulation genes tet(C) and tet(D) and Tn10/IS10 transposon elements. The fosfomycin resistance gene, fosA7, was found adjacent to an ISL3 family-like transposase within the E. coli chromosome. Finally, a trimethoprim resistance gene, dfrA5, was adjacent to the clinical class 1 integron integrase, intI1, which was directly upstream of the TnAs1 transposase and resolvase and the Tn21 modulator gene, tnpM.

Both host characteristics and landscape features were important predictors of the ARG spatial distribution.

To better understand the geographical distribution of identified ARGs from owls, we conducted several spatial analyses utilizing information about anthropogenic features surrounding owl recovery locations. First, to determine whether there was any spatial clustering of ARG-positive owls, we performed a purely spatial cluster analysis with SaTScan software (http://www.satscan.org) for each ARG and antimicrobial chemical class. A single high-risk cluster of owls harboring at least one tetracycline resistance gene [i.e., tet(A), tet(B), or tet(J)] was identified in southern Minnesota (relative risk = 21.74, P = 0.014; Fig. 3). No other high- or low-risk clusters were identified for any other antimicrobial chemical class or ARG.

FIG 3.

Map displaying owl recovery locations and the results of spatial-only cluster analyses performed with SaTScan software. Points indicate recovery locations, with red points indicating tetracycline resistance gene-positive owls.

Second, to investigate the role of host characteristics and anthropogenic landscape features on the spatial distribution of ARGs, we used multimodel inference with model averaging to identify important independent variables. Spatial metrics associated with each owl were based on the recovery location and were calculated for both the owl’s potential home range area and the local watershed. Overall, only a few landscape features and host-specific characteristics were identified to be important predictors of ARG-positive owls (Fig. 4; Table S1). The sequencing read count was an important predictor in all models, with the probability of identifying ARGs increasing with the sequencing depth (35). Seasonality and agricultural land cover were identified to be additional important predictors of the likelihood of an owl with at least one ARG conferring resistance to any antimicrobial chemical class, with ARG-positive owls more likely to be recovered during warmer periods of the year and in watershed areas with higher agricultural land use (Fig. 4A). For the likelihood of identifying an owl with at least one β-lactam resistance gene (primarily bla_TEM-1b), owl age and domestic cattle density were the most important predictors. Specifically, birds in their hatch year were more likely to possess a β-lactam resistance gene than adult birds, and, somewhat surprisingly, β-lactam resistance gene-positive owls were more common in watershed areas with a lower cattle density than in watershed areas with a higher cattle density (Fig. 4B). The remaining predictors included in the final models (host species and body condition score, poultry density, and point source density) had confidence intervals overlapping zero, which suggests that they had little to no importance for predicting ARG-positive owls. All other antimicrobial chemical classes tested, including tetracyclines, included the null (intercept-only) model in the top model set. This indicated that the observed spatial variation of ARG-positive owls could not be adequately explained by any of the proposed covariates. Using Moran’s I to test for a residual spatial correlation in each regression analysis, only the tetracycline top models suggested an unexplained spatial correlation (Table S2).

FIG 4.

Zero-averaged and standardized coefficients from multimodel inference of the presence of any antimicrobial resistance gene (A) and β-lactam resistance genes (B) identified in owl cloacal swabs. Separate model sets were generated for spatial predictors based on the water catchment (blue points) and estimated home range (orange points) of owl recovery locations. Only the predictors included in the final model sets (ΔAICc < 2) are shown. Points with black outlines indicate predictors where the 95% confidence intervals do not overlap zero. Abbreviations: GH, great horned owl; BD, barred owl; HY, hatch year; A, adult; temp., temperature.

DISCUSSION

The presence of AMR in natural ecosystems, including wild animals, is a widespread and well-documented phenomenon (1, 2). However, to date, we still lack a complete picture of the forces shaping the geographical distribution of AMR in the environment, including the importance of anthropogenic sources of AMR compared to that of other landscape- and host-specific determinants (13, 18, 19). Here we demonstrate how analytical tools from spatial epidemiology, in combination with selective shotgun metagenomic sequencing, can be used to investigate how landscape features and host ecology shape the spatial distribution of ARGs from wild owls.

Overall, the ARG prevalence identified in this study was comparable to that identified in other studies in wildlife (6, 7, 9), with acquired ARGs being identified in 23% of samples. Past studies of resistant Enterobacteriaceae in raptors, including owls, have reported a wide range of resistance levels for various tested antibiotics and identified ARGs (24, 43–46). In part, these differences could be due to variability in sampling, culturing, and AMR detection methods between studies, with some studies reporting the phenotypic resistance profiles of only individual bacterial isolates (24, 43–45) and others including additional information on the specific ARGs identified (21, 46). In the present study, a sequence-based approach was used to identify ARGs, the presence of which, it should be noted, does not necessarily reflect phenotypic resistance. Further, AMR prevalence may be underestimated if resistance is conferred by novel ARGs that are undocumented in reference databases. Additionally, our study focused on antimicrobial resistance in Gram-negative Enterobacteriaceae because they are thought to play an important role in the global spread of ARGs and included a number of medically important ARB (29, 30). However, prior studies of AMR in raptors have also focused on Gram-positive enterococci, in which the levels of resistance detected were similar to or higher than those detected in the Enterobacteriaceae (23, 46, 47), suggesting that we may have missed an important component of the raptor resistome by focusing only on Enterobacteriaceae.

Our sample processing also involved a freeze-thaw step, which may have affected cell viability, resulting in a decreased diversity of bacterial strains and/or ARGs. The two subsequent culture steps, first in nonselective brain heart infusion broth, to aid in bacterial recovery after the thaw, and then in selective lauryl tryptose broth, to enrich for Enterobacteriaceae, also likely impacted our results. Given that some bacterial species are unculturable and that there can be substantial growth variation between different strains of bacteria, the diversity and abundance of the resulting sequenced bacterial subpopulations were likely not representative of those of the true starting community (36). If certain ARG-positive strains did not recover well after freezing or did not adequately grow during either culture step, there may have been additional ARGs in our samples that we were unable to detect. That said, the relatively high cost of metagenomic sequencing of nonenriched samples necessitates a trade-off between the completeness of resistome characterization (i.e., sequencing depth) and the number of samples to be sequenced (35). By using a hybrid of culture enrichment and metagenomic sequencing, we were able to include samples from a broad range of landscape types and generate the larger sample size necessary for statistical analyses.

Heterogeneity in AMR prevalence between studies, as well as between individuals within the same study, may also be partly attributable to host biological and ecological characteristics. While owl species was not identified to be an important predictor in our multimodel regression analyses, both owl age and collection season were associated with the likelihood of detecting ARGs. Specifically, birds in their hatch year were more likely to harbor ARGs conferring resistance to β-lactams than adult birds. While this pattern has been observed in both wild and domestic mammals (48–50), to our knowledge, this is the first report of an age-related distribution of AMR in wild birds. Seasonality was also an important predictor of the presence of ARGs, with birds sampled during warmer months being more likely to harbor at least one ARG than birds sampled during cooler months. A similar association between seasonality and the presence of a resistant E. coli strain has been observed in both wild mammals and birds (51, 52). The specific mechanisms underlying the relationship between both of these variables and AMR still remain unknown, but differential exposures and/or selective pressures or underlying changes in gut microbiomes between ages or seasons may play a role (51, 53, 54). For example, one hypothesis that has been put forward to explain the age-related decline of AMR in domestic mammals proposes that certain resistant bacterial strains in young animals may have a selective advantage over susceptible strains due to a genetic linkage between ARGs and other fitness-enhancing genes (49).

Beyond host-specific determinants of AMR, we also investigated the importance of landscape features, including potential anthropogenic sources of AMR. Overall, we found a limited spatial clustering of ARG-positive owls, and few landscape features were significantly associated with the likelihood that an owl harbored ARGs. This may have been due to our relatively small sample size and low AMR prevalence, which may have impeded our ability to detect all patterns present in the data set. That said, the patterns that we were able to detect may be indicative of robust support. Specifically, agricultural land cover was identified to be an important predictor for the presence of any ARG, with ARG-positive owls more likely to be recovered from water catchments containing a high density of hay, pasture, or row crops. One potential explanation for this is the common practice of spreading animal manure or biosolids on agricultural soils as fertilizer. These materials can act as a reservoir for residual antibiotic compounds, ARB, and ARGs, which may persist in the soil long after application (55–57). Similarly, the use of pesticides on agricultural fields has been shown to coselect for ARGs (57). Together, these agricultural practices may increase the AMR exposure risk for local wild animals, adjacent water bodies, or groundwater (57). Thus, owls may have acquired ARB and ARGs via the consumption of small mammals commonly inhabiting agricultural fields (e.g., rodents) or directly via contaminated drinking water (23, 58). Further, the fact that both cattle density and agricultural land cover were significant predictors when calculated only on the basis of water catchment and not the estimated owl home ranges suggests that water systems may be particularly important drivers of AMR dissemination within natural ecosystems (59, 60).

Interestingly, we also identified a significant high-risk cluster of tetracycline resistance gene-positive owls in southern Minnesota, where the vast majority of food-animal production takes place in the state (61). While there is a lack of data on the usage of tetracyclines in Minnesota (but see reference 62), it is well established that tetracyclines are one of the most extensively used antibiotic classes in most food-producing animals, including cattle, swine, and poultry, which are all present at higher densities in southern Minnesota than in the northern half of the state (63). However, this observed cluster of tetracycline resistance gene-positive owls could not be adequately explained by any of our host or spatial variables, suggesting that our regression analyses may have been underpowered due to a small sample size and/or that there may be unknown factors that we did not account for in our models. Further, because this was a convenience sample of owls, there may have been some geographical areas from which bird submission was more or less likely (i.e., urban versus wilderness areas). While randomized or spatially stratified sampling would have been ideal, that is rarely possible in field studies of wildlife. That being said, we were able to obtain an ample sample representation across a gradient of anthropogenic landscapes and ecoregions. Finally, these results may simply reflect the pervasive nature of many of the ARGs that we identified in natural ecosystems. Specifically, the majority of identified ARGs have similarly been reported in a wide range of wild animals and environmental resources for decades (1, 2), suggesting that even if they were originally of anthropogenic origin, the signal was effectively lost as the genes spread throughout various hosts and environments. For example, in the present study, genes conferring resistance to β-lactams and tetracyclines were the most common ARGs identified. However, the most frequently identified gene groups, bla_TEM and tet(A), are both commonly found in natural environments (31, 33), indicating that we would not necessarily expect a higher prevalence of these genes in anthropogenically impacted landscapes than in more pristine environments. Indeed, we observed the reverse association, with β-lactamase ARG-positive owls more likely to be recovered from areas with a low density of domestic cattle (including dairy cattle and cattle used for beef production) than from areas with a high density of domestic cattle. We suspect that had more recently emerged ARGs of clear human origin been identified in our samples (e.g., mcr-1) (64), there may have been a stronger association with anthropogenic landscape features.

Our identification of certain mobile genetic elements does provide compelling evidence that some ARGs identified in this study may be of anthropogenic origin. In particular, the presence in one sample of the trimethoprim resistance gene dfrA5 in association with the clinical class 1 integron integrase gene intI1 suggests a potential anthropogenic source, as intI1 is thought to be a good proxy of anthropogenic pollution (65, 66). Indeed, the complete tnpA-tnpR-tnpM-intI1-dfrA5 region has been identified in E. coli plasmids from both humans and domestic animals (67–69), but there is currently no published record of it in a wild animal. Similarly, all other ARG-MGE sequence regions found in our samples have also been commonly identified in Enterobacteriaceae from a broad range of human and animal hosts. However, few of these records come from wildlife or natural sources, such as water and soil. While this could indicate that novel and recent transmission events led to the presence of these sequences in our wild owl samples, a more likely explanation is that such ARGs-MGEs are present in natural ecosystems but have not yet been identified due to undersampling and a lack of sequencing techniques that provide the genetic context of ARGs. This highlights the ongoing need for studies of AMR in wildlife to go beyond simply identifying resistance phenotypes and ARGs and to start exploring the more complex details of AMR transmission pathways between anthropogenic and natural ecosystems. These necessary next steps will be greatly aided by the wide array of sequencing and analytical tools now available to researchers, with the methodology presented here being one of many possibilities.

MATERIALS AND METHODS

Study population and sample collection.

Great horned owls (Bubo virginianus) and barred owls (Strix varia) are the most common owls in North America and can be found across a broad variety of ecoregions, including wetlands, forests, grasslands, agricultural areas, and urban/suburban areas (27, 28). Great horned and barred owls are nonmigratory, and most individuals have permanent or long-term residencies in a particular region (27, 28). Home ranges are typically small (71 to 560 ha for great horned owls and 84 to 498 ha for barred owls [37]), although they may change in response to owl population density or prey availability (70, 71). Both species are generalist predators with a diet that includes small and medium-sized mammals, birds, and large invertebrates (27, 28).

Samples were opportunistically collected from all great horned owls (n = 47) and barred owls (n = 31) admitted to The Raptor Center (TRC; https://www.raptor.umn.edu) at the University of Minnesota (St. Paul, MN, USA) between December 2016 and July 2017. Samples were taken as part of clinical evaluation and routine care of wild patients undergoing treatment in the medical facility; thus, Institutional Animal Care and Use Committee approval was not required. The owls were recovered by public volunteers throughout Minnesota (n = 69), western Wisconsin (n = 8), and northeastern North Dakota (n = 1) and transported to TRC. Given the opportunistic nature of locating an injured owl, greater numbers of individuals were recovered from more densely human populated areas, particularly Minnesota’s two largest cities, Minneapolis and St. Paul. The latitude and longitude (geographic coordinate system WGS84) of each owl recovery location were determined from the nearest intersection or other prominent landmark.

Immediately upon admission to TRC, before any antimicrobials were administered, a single cloacal swab sample was collected from each owl by rubbing a sterile flocked swab (HydraFlock; Puritan Medical Products, Guilford, ME) against the cloacal mucosa until fecal material was transferred. The swabs were then immediately placed in 1 ml of brain heart infusion broth (Oxoid, ThermoFisher Scientific, Waltham, MA) with 15% glycerol (Fisher Chemical, ThermoFisher Scientific, Waltham, MA), vortexed for 10 s, and stored at −80°C for future use. Information on owl age class (adult or hatch year), sex (male, female, unknown), reason for admittance (e.g., starvation, collision), and body condition score (0.5 to 5.0) as a proxy for health status were also established at the time of admission (see Data Set S2 in the supplemental material).

Sample processing and sequencing.

To enhance the detection of antimicrobial resistance genes in low-abundance species of the Enterobacteriaceae, the samples underwent culture enrichment before sequencing (36). Sample glycerol stocks were thawed at room temperature and gently agitated to resuspend any substrate. One hundred microliters of the sample was then added to 10 ml of brain heart infusion broth (Oxoid, ThermoFisher Scientific, Waltham, MA), and the mixture was incubated at 37°C for 3 h. Lauryl tryptose broth (Becton, Dickinson and Company, Franklin Lakes, NJ), a selective growth medium for coliforms, was then added to achieve a 1× final concentration, and the culture was incubated at 37°C for 18 h. For DNA extraction, 5 ml of the culture was centrifuged at 4,000 rpm for 15 min, the supernatant was removed, and the bacterial pellet was resuspended in 300 μl of sterile 1× phosphate-buffered saline (VWR, Radnor, PA). Total DNA was extracted using a DNeasy PowerSoil kit (Qiagen, Hilden, Germany). Shotgun sequencing library preparation was performed with a Nextera XT kit (v2; Illumina, Inc., San Diego, CA, USA), and the libraries were multiplexed and sequenced on an Illumina MiSeq platform with v3 chemistry (2 × 300 bp).

Bioinformatics.

Figure 1 depicts a flowchart of the bioinformatic steps taken in this study. Raw paired-end reads were trimmed and quality filtered using the Trimmomatic trimmer (v0.33) with the following options: the ILLUMINACLIP command was used to remove adaptors and other Illumina-specific sequences (72); a sliding window of 4 nucleotides was used to clip reads starting at the 5′ end, once the average quality in the window fell below Q15; and reads with a length of less than 36 nucleotides were discarded. Quality assessment of raw and trimmed reads was performed with the FastQC program (v0.11.5) (73).

Taxonomic classification.

To identify the bacterial species present in each sample, taxonomic labels were assigned to each read using the program Kraken (v1.1) with default parameters and the prebuilt MiniKraken 8-GB database (v20171019) (74) (Fig. 1, flowchart section labeled “1) Taxonomic classification”). Since the samples were enriched for coliforms before sequencing, the majority of the reads were expected to be classified to the family Enterobacteriaceae. A custom script was used to calculate the number of reads assigned directly to each taxon using the output from the command kraken-translate–mpa-format. To account for the differential sequencing depth between samples, read counts for each taxon were normalized via cumulative sum scaling with the R package metagenomeSeq (v1.18.0) (75, 76), using the 0.5 default percentile.

Antimicrobial resistance gene identification.

To identify acquired ARGs, trimmed reads were aligned to ARG reference sequences from the ResFinder database (downloaded 9 August 2018) (77), using the BWA-MEM algorithm (v0.7.17) (78) (Fig. 1, flowchart section labeled “2) ARG identification”). To decrease the likelihood of the false-positive identification of resistance genes, the BBmap script (https://sourceforge.net/projects/bbmap/) pileup.sh was used to calculate the percentage of nucleotides in a given reference gene with at least one aligned read (i.e., the gene fraction). Only reference ARGs with a gene fraction of ≥80% were considered to be present in a sample (79). Given the high degree of sequence homology between some ARGs, a small subset of our reads aligned equally well to multiple ARG reference sequences (i.e., multimapping reads). In such situations, we observed that the BWA-MEM algorithm randomly assigned each multimapping read to a single reference sequence, effectively reducing the likelihood that a sufficient number of reads required to meet the ≥80% gene fraction threshold would align to any one of the homologous reference sequences. Thus, to improve the sensitivity of ARG identification, all alternative alignments were included in the BWA-MEM output SAM file’s XA:Z field. The BWA xa2multi.pl script was then used to parse alternative alignments into multimapping alignments (78), and a custom script was used to identify multimapping reads. To avoid double counting, reads that mapped to multiple reference sequences in different gene groups [e.g., bla_TEM, tet(A), sul2, etc.] were removed from the SAM file prior to calculating gene fractions. Reference ARGs with a gene fraction of ≥80% in each sample were then aggregated to the gene group level to avoid the double counting of reads mapping to multiple reference sequences within the same ARG group.

Characterization of antimicrobial resistance gene genetic context.

A drawback to using a read mapping approach to identify ARGs is that we lack information on the genetic context of the ARGs (i.e., the genetic environment). Particularly for studies of AMR transmission dynamics, it is important to know whether an ARG is colocalized with other ARGs or whether it is associated with mobile genetic elements. Trimmed paired and unpaired reads were assembled with the metaSPAdes program (v3.10.0), using the default parameters and k values of 21, 33, and 55 (80) (Fig. 1, flowchart section labeled “3) ARG genetic context”). Additional assemblies were run by consecutively adding k values of 77, 99, and 127. Contigs with a length of <500 bp were removed from the assemblies. Assembly quality was assessed with the QUAST program (v4.3) (81), and the assembly with the highest N₅₀ value for each sample was chosen for downstream analysis. To calculate the number of total reads used in the assemblies, paired and unpaired reads were mapped back to contigs with the BWA-MEM algorithm, and aligned reads were identified by use of the SAMtools program (v1.9) (82). Coding sequence (CDS) predictions for each contig were made by use of the Prokka program (v0.7.17) (83). ARGs were identified by aligning CDS features to the ResFinder database using the blastn program (84) with a percent identity of ≥90% and a reference gene coverage of ≥80%. ARG-carrying contigs were manually annotated using the blastx program (84) and the following databases: ISfinder for insertion sequences (85), INTEGRALL for integrons (86), PlasmidFinder for plasmid replicons (87), and the National Center for Biotechnology Information’s (NCBI’s) collection of nonredundant (nr) protein sequences (https://www.ncbi.nlm.nih.gov/). Particular focus was paid to determining whether each ARG-carrying contig was chromosomal or plasmid derived and identifying mobile genetic elements associated with ARGs. The predicted bacterial origin of each ARG-carrying contig was determined using the blastn program and NCBI’s nucleotide collection and noting the taxonomic origin(s) of the top hits.

Data analyses.

Given that the culture enrichment step prohibited the accurate quantification of genes in the original samples (36), only the presence/absence of different ARG groups was considered. Heatmaps displaying the normalized relative abundance of bacterial species and the presence/absence of ARG groups across samples were created with the heatmap.2 function from the R package gplots (v3.0.1) (88). Annotated contigs were viewed with SnapGene Viewer software (v4.2.11) (GSL Biotech).

Spatial cluster analysis.

To detect the geographical areas in which ARG-positive owls occurred at higher or lower frequencies than random expectations, we performed a cluster detection analysis with SaTScan software (v9.6) (http://www.satscan.org), using the purely spatial scan statistic with a Bernoulli model (89). The spatial scan statistic is based on the likelihood ratio test and utilizes a moving circular window where the presence of a cluster is determined by whether the estimated risk within a window is significantly greater than the estimated risk outside of the window (89). A maximum geographical window size of 50% of the sampled population was used, and cluster P values were obtained through Monte Carlo hypothesis testing with 999 replications. Separate cluster detection analyses were performed for the presence/absence of any ARG and for the presence/absence of resistance gene groups associated with particular antimicrobial classes that were identified in at least 5% (4/78) of the owls. Owl recovery locations and statistically significant spatial clusters (P < 0.05) were mapped with ArcGIS desktop software (v10.6.1.9270) (90).

Risk factor analysis.

To assess the relative importance of host and environmental factors on the presence of ARGs recovered from Minnesota owls (n = 69), we performed logistic regression analyses within a multimodel inference framework. Multimodel inference can account for model selection uncertainty and reduces model selection bias, making it more robust and conservative than traditional stepwise model selection analysis (91). Response variables included the presence/absence of any ARG and the presence/absence of ARGs associated with particular antimicrobial classes that were identified in at least 5% (4/78) of the samples. Given the relatively small sample size and the rare number of ARG-positive samples, we used Firth’s penalized likelihood approach implemented in the R package brglm (v0.6.1) (92).

The number of raw sequencing read pairs was included as a covariate to control for the correlation between sequencing depth and the likelihood of identifying ARGs. Host covariates included owl species (great horned owl versus barred owl); age (hatch year versus adult); body condition score; and recovery month temperature, which was a proxy for the recovery season. Spatial covariates included the number of point sources per square kilometer; the number of swine, bovine, and domestic fowl units per square kilometer; the presence of an impaired water body (yes versus no); and the percentage of land cover classified as agricultural. Geospatial data on point sources, animal units, and impaired water bodies were obtained from the Minnesota Pollution Control Agency (MPCA) via the Minnesota Geospatial Commons (https://gisdata.mn.gov; accessed 10 January 2019). Point sources included spatial features associated with wastewater, storm water, hazardous waste, solid waste, and feedlots (93). An “animal unit” was defined as a measure of the relative manure production at feedlots and manure storage areas, accounting for the variation in production by animal species and size (94). Separate animal unit totals were calculated for swine (including all members of the family Suidae, primarily domestic hogs), bovine (subfamily Bovinae, primarily domestic cattle), and domestic fowl (order Galliformes, primarily turkeys and chickens). “Impaired water bodies” were classified by the MPCA as any body of water that did not meet the quality standards defined by the Environmental Protection Agency, including acceptable levels of pollutants and coliforms (95). Agricultural land cover data were obtained from the University of Minnesota’s Remote Sensing and Geospatial Analysis Laboratory (96) (https://rs.umn.edu/mndata; accessed 14 January 2019) and included all areas classified as hay, pasture, or row crops.

Spatial covariates for each owl were calculated in ArcGIS desktop software (v10.6.1.9270) (90) over two different spatial scales: potential home range and water catchment (Fig. S2). Potential home range area was based on an estimated true home range size of 5.6 km² (diameter ≈ 2.67 km) (37). However, because an owl may have been recovered at any spatial point within its true home range, including an outermost edge, we defined the potential home range area to be a circle with a diameter of 2 × 2.67 km (5.34 km; area [A] = 22.4 km²) with the owl recovery location at the circle’s center. Water catchments were based on level 08 delineations, as defined by the Minnesota Department of Natural Resources (97), and geospatial boundaries were obtained from the Minnesota Geospatial Commons (https://gisdata.mn.gov/dataset/geos-dnr-watersheds; accessed 10 January 2019). Due to the high level of correlation between potential home range and water catchment spatial covariates, separate covariate sets were created for each spatial scale (Table S3). The number of raw sequencing read pairs and all host covariates were included in both covariate sets.

Separate regression analyses and model-averaging procedures were performed for each response variable and covariate set. For each analysis, all possible combinations of covariates, including the null intercept-only model, were used to generate a candidate model set. To account for potential collinearity among covariates, variance inflation factor (VIF) values were calculated for each model using the R package car (v3.0) (98), and models exhibiting VIF values of >2.0 were removed (99). Candidate models were then compared using Akaike’s information criterion corrected for small sample size (AICc), and the top model set included all models with a change in the AICc (ΔAICc) of 2.0 from the lowest AICc, as recommended by Burnham and Anderson (100). Parameter estimates and standard errors were obtained by averaging over all models in the top model set using the zero (or full) average method (100, 101) with the model.avg function in the R package MuMIn (v1.40.4) (102). To make the comparisons between covariates more interpretable, coefficient estimates were standardized using standard deviations of the covariate and response variable (103). Confidence intervals that did not span zero were used to identify important covariates (104). If the null (intercept-only) model was included in the top model set, we concluded that the variation in the response variable could not be explained by any of the proposed covariates.

To test for a spatial autocorrelation in the residuals from each regression analysis, we computed Moran’s I autocorrelation coefficient for the model with the lowest AICc in each model set using the Moran.I function from the R package ape (v5.2) (105, 106). Distances between sample recovery locations were calculated using the distm function from the geosphere package, where the shortest distance between two points was based on a geographic coordinate system WGS84 reference ellipsoid (107).

Data availability.

All Illumina MiSeq sequencing data generated for the current study have been deposited in NCBI’s Sequence Read Archive under BioProject accession number PRJNA605836.