Adjacent Terrestrial Landscapes Impact the Biogeographical Pattern of Soil Escherichia coli Strains in Produce Fields by Modifying the Importance of Environmental Selection and Dispersal

Understanding the ecology of enteric bacteria in extrahost environments is important for the development and implementation of strategies to minimize preharvest contamination of produce with enteric pathogens. Our findings suggest that watershed landscape is an important factor influencing the importance of ecological drivers and dispersal patterns of E. coli.

ABSTRACT

High-quality habitats for wildlife (e.g., forest) provide essential ecosystem services while increasing species diversity and habitat connectivity. Unfortunately, the presence of such habitats adjacent to produce fields may increase risk for contamination of fruits and vegetables by enteric bacteria, including Escherichia coli. E. coli survives in extrahost environments (e.g., soil) and could be dispersed across landscapes by wildlife. Understanding how terrestrial landscapes impact the distribution of soil E. coli strains is of importance in assessing the contamination risk of agricultural products. Here, using multilocus sequence typing, we characterized 938 E. coli soil isolates collected from two watersheds with different landscape patterns in New York State, USA, and compared the distribution of E. coli and the influence that environmental selection and dispersal have on the distribution between these two watersheds. Results showed that for the watershed with widespread produce fields, sparse forests, and limited interaction between the two land use types, E. coli composition was significantly different between produce field sites and forest sites; this distribution appears to be shaped by relatively strong environmental selection, likely from soil phosphorus, and slight dispersal limitation. For the watershed with more forested areas and stronger interaction between produce field sites and forest sites, E. coli composition between these two land use types was relatively homogeneous; this distribution appeared to be a consequence of wildlife-driven dispersal, inferred by competing models. Collectively, our results suggest that terrestrial landscape attributes could impact the biogeographic pattern of enteric bacteria by adjusting the importance of environmental selection and dispersal.

IMPORTANCE Understanding the ecology of enteric bacteria in extrahost environments is important for the development and implementation of strategies to minimize preharvest contamination of produce with enteric pathogens. Our findings suggest that watershed landscape is an important factor influencing the importance of ecological drivers and dispersal patterns of E. coli. Agricultural areas in such watersheds may have a higher risk of produce contamination due to fewer environmental constraints and higher potential of dispersal of enteric bacteria between locations. Thus, there is a perceived trade-off between priorities of environmental conservation and public health in on-farm food safety, with limited ecological data supporting or refuting the role of wildlife in dispersing pathogens under normal operating conditions. By combining field sampling and spatial modeling, we explored ecological principles underlying the biogeographic pattern of enteric bacteria at the regional level, which can benefit agricultural, environmental, and public health scientists who aim to reduce the risk of food contamination by enteric bacteria while minimizing negative impacts on wildlife habitats.

INTRODUCTION

Forests and other riparian buffers can provide ecological and agricultural benefits (e.g., reducing soil erosion and leaching of chemical and fecal waste into surface water sources, providing habitat and connective pathways for wildlife) as well as esthetic benefits (1–3). While high-quality habitats (e.g., forest) offer conservation services, they may also bring unintended consequences and may increase the risk of preharvest contamination of produce crops. For one, extrahost environments, such as soil in high-quality habitats, can be critical reservoirs for enteric bacteria, leading to potential dispersal of enteric bacteria to adjacent agricultural fields (4). Direct fecal deposition onto produce by wildlife is also a potential pathway for introducing enteric bacteria onto food crops (5). By providing wildlife movement pathways, high-quality habitats may facilitate the wildlife-driven dispersal of enteric bacteria through riparian corridors to agricultural regions, possibly resulting in contamination of food crops (6–9). Subsequent persistence and/or regrowth of pathogenic enteric bacteria introduced into fresh produce fields would then further increase food safety risks (10). Since the survival of enteric bacteria and movement of wildlife vary by land use type (11, 12), it is reasonable to hypothesize that watershed landscape impacts the distribution of enteric bacteria.

Environmental selection and dispersal are two fundamental ecological forces that drive the distribution of bacteria (13–16). The essential roles of environmental selection via abiotic (e.g., pH and salinity) and biotic selective pressures on bacteria have been well documented in many local and even global habitats (17–20). Environmental selection facilitates the genetic divergence of some ecophysiological traits owing to their contribution to fitness benefits for adaptation of bacteria to diverse habitats, such as those with different land cover types (21, 22). With the influence of environmental selection, a high level of bacterial dissimilarity between locations (beta diversity) can be maintained in a wide range of environments (15, 23, 24).The role of dispersal in driving the distribution of bacteria at a local as well as a regional scale is evident, since dispersal provides a mechanism for bacteria to colonize new habitats (25, 26). The relative importance of dispersal in shaping bacterial distribution varies among microbial taxa due to diversity in the capacity of bacteria to disperse via wind, water, and wildlife. For example, bacteria with a long-range dispersal capacity (e.g., Polaromonas) tend to exhibit a more global distribution (27), while bacteria with a limited dispersal range (e.g., Rhizobiaceae, Bradyrhizobiaceae, and Xanthomonadaceae) tend to show more ecological specialization (15). Importantly, wildlife presence and movement are fundamentally affected by the physical elements and features of land (15). Thus, wildlife-driven dispersal of bacteria can be quantitatively predicted by landscape ecological models based on the relationship of wildlife behaviors and landscape characteristics (e.g., patchiness, land-use interspersion, patch connectivity, patch diversity, and land use interactions) (8, 28–30). Based on these principles, investigating environmental selection and dispersal of enteric bacteria from and within habitats with distinct landscape patterns has the power to elucidate the effect of terrestrial landscapes on the distribution of enteric bacteria and to assess the associated risk of preharvest contamination of food by pathogenic enteric bacteria.

As a commensal or pathogenic enteric bacterium that is widespread in diverse habitats, Escherichia coli primarily resides in the intestines of warm-blooded animals, and survives in extrahost environments such as water, soil, and sediments as well (31, 32). Soil is a habitat of particular interest for E. coli, since the high chemical and physical heterogeneity of soil across different environments could impose multifarious environmental selection pressures on E. coli (22, 23, 32). Observations that the prevalence of E. coli varies by land cover type (e.g., deciduous forest, cropland, or pasture) (22) also suggest that different land uses could stimulate different types and intensities of selective pressures that act on E. coli. The key soil variables influencing the growth of E. coli are commonly recognized as pH and moisture (23, 33), while some other variables, such as organic matter and texture, could also play a role (10). However, these properties are coarse indicators of many aspects of the soil environment. A more detailed assessment of specific chemical and physical factors affecting growth or survival is needed. In addition, wildlife, such as avian species and ruminant animals, could act as vehicles of dispersal of E. coli (34, 35). E. coli can also be transmitted between wildlife hosts through contact and can be deposited in new locations (e.g., produce fields) by defecation, which often happens when wildlife forages for food (36). Given the intensive interaction with both extrahost and host habitats, E. coli may be a useful model to build predictive capabilities surrounding interactions between bacteria and agricultural landscapes on a meter to kilometer scale. Such an understanding is particularly important for land where fresh fruits and vegetables are cultivated, as it could be used to develop better strategies for minimizing pathogen introduction into preharvest environments.

We hypothesized that the importance of environmental selection and dispersal for the distribution of E. coli is dependent on landscape and specifically hypothesized that (i) E. coli distribution in watersheds with higher coverage of agricultural environments is strongly driven by environmental selection associated with agricultural activity, while (ii) E. coli distribution in regions with higher coverage of natural environment is largely influenced by wildlife-driven dispersal. To test our hypotheses, we characterized 938 generic E. coli isolates obtained from soil samples collected from two watersheds, Flint Creek and the Hoosic River, both located in New York State, using a hierarchical multilocus sequence typing (MLST) scheme. These two watersheds allow an interesting comparison between the one with widespread produce fields and limited interaction between produce fields and forest (Flint Creek; 69% produce fields and 12% forest by area; adjacency_{produce|forest} = 23%) (Fig. 1A) and the one with heavily forested areas and strong interaction between produce fields and forest (Hoosic River; 28% produce fields and 38% forest by area; adjacency_{produce|forest} = 36%) (Fig. 1B). Next, we investigated the distribution of E. coli in these two watersheds and assessed the relationship between E. coli distribution and soil variables and the distance-decay relationship among E. coli populations. Last, we developed dispersal models for four wildlife vehicle candidates (large nuisance wildlife species, small mammals, small flocking insectivores/granivores, and migratory bird flocks) to quantify the importance of wildlife-driven dispersal for the distribution of E. coli in the two watersheds.

FIG 1.

Sampling maps of (A) Flint Creek and (A) the Hoosic River. Dots indicate the sampling sites within each watershed. Map layers for land cover (National Land Cover Database [NLCD], 2006 [https://www.mrlc.gov/data/nlcd-2006-land-cover-conus]) were acquired from the U.S. Geological Survey (USGS) Earth Explorer geographical data bank (https://earthexplorer.usgs.gov/).

RESULTS

Distribution of E. coli.

Soil samples from the Flint Creek watershed (predominately produce fields) showed considerably lower prevalence of E. coli than those from the Hoosic River watershed (predominately forests); 35% and 72% of soil samples, respectively, were positive for E. coli in these two watersheds. These samples yielded 289 and 649 E. coli isolates, respectively. Based on initial 2-gene MLST (mdh and uidA), a total of 138 isolates from Flint Creek and 277 isolates from the Hoosic River were selected for characterization by full 7-gene MLST (aspC, clpX, icd, lysP, and fadD, in addition to mdh and uidA). The 7-gene MLST generated 121 unique multilocus sequence types (ST) for Flint Creek and 191 unique ST for the Hoosic River (see Tables S1 and S2 in the supplemental material). Analysis by goeBURST identified 96 and 108 E. coli clonal groups for Flint Creek and the Hoosic River, respectively, based on ST at the single-locus-variant level (see Fig. S1 and S2). For Flint Creek, forest sites had slightly higher mean richness of E. coli clonal groups than produce field sites, but the difference was not significant (P = 0.86) (Fig. S3). In contrast, in the Hoosic River watershed, the mean richness of E. coli clonal groups in produce field sites was significantly higher than that in forest sites (P < 0.05) (Fig. S3).

E. coli in the two watersheds displayed distinct distribution patterns. For Flint Creek, principal-coordinate analysis (PCoA) based on the dissimilarity of E. coli clonal groups clustered sampling sites by land use (Fig. 2a). Both permutational multivariate analysis of variance (PERMANOVA) test and analysis of similarities (ANOSIM) showed that this clustering by land use was significant (P < 0.05) (Tables S3 and S4). In contrast, for the Hoosic River, the sampling sites were not significantly clustered by land use in PCoA (Fig. 2b) of E. coli clonal groups (PERMANOVA P = 0.86, ANOSIM P = 0.58) (Tables S3 and S4), indicating a more homogeneous composition of E. coli between produce field sites and forest sites in this watershed. To test the robustness of these results and to assess the effects of potential sampling bias on the E. coli diversity, we repeated PCoA, PERMANOVA, and ANOSIM after excluding sites with a low number of E. coli clonal groups detected (≤3); sites excluded were field 6, field 8, and forest F9 from Flint Creek. Results showed that sampling sites from this subset were significantly clustered by land use in the PCoA plot (Fig. S4) for E. coli clonal groups from Flint Creek at least at the 0.1 level (PERMANOVA P = 0.05, ANOSIM P = 0.08) (Table S5). Slightly higher P values in PERMANOVA and ANOSIM for a subset of sites (compared to P values for the whole set of sites) are likely due to the reduced sample size in these analyses.

FIG 2.

PCoA plots of E. coli clonal groups of sites in (a) Flint Creek (FC) and (b) the Hoosic River (HR). Green dots indicate forest sites; orange dots indicate produce field sites; green circles indicate the 95% confidence ellipse of forest sites; orange circles indicated the 95% confidence ellipse of produce field sites. P values of permutational multivariate analysis of variance (PERMANOVA) and analysis of similarities (ANOSIM) are shown. An asterisk indicates that the clustering of sampling sites by land use is significant at the 0.05 level. (a) PCo axes 1 and 2 explained 18.7% and 14.9%, respectively, of the variation of E. coli clonal groups for Flint Creek. (b) PCo axes 1 and 2 explained 14.6% and 12.0%, respectively, of the variation of E. coli clonal groups for the Hoosic River.

Importance of soil variables driving the distribution of E. coli.

We first employed variance partitioning analysis (VPA) to quantify the contribution of soil variables and spatial variables to the distribution of E. coli from Flint Creek and the Hoosic River. After screening for high levels of covariation, moisture, pH, sodium, phosphorus, barium, manganese, and antimony were included in VPA (Table S6). VPA revealed that the soil variables individually explained 11.7% of the biological variation of dissimilarity of E. coli clonal groups from Flint Creek, while spatial variables individually explained 5.5% of this variation (Fig. 3a). The variation explained by spatially structured soil variables was relatively low (0.1%), and about 83% of the variation was unexplained by selected variables (Fig. 3a). In contrast, for E. coli clonal groups from Hoosic River, about 99% of the variation of dissimilarity could not be explained by selected variables (Fig. 3b). Individually, selected soil variables did not explain any of the observed variation and spatial variables only explained 1% of the variation for the Hoosic River watershed (Fig. 3b). These results indicate that environmental selection outweighs spatial factors with regard to the importance for E. coli distribution in the Flint Creek watershed; E. coli isolates from this watershed appear to undergo much stronger environmental selection or a much lower rate of dispersal than E. coli isolates in the Hoosic River watershed.

FIG 3.

Variance partitioning analysis (VPA) showing relative contributions of spatial factors (Geo.), soil variables (Env.), and spatially structured environmental variables to the variations of the dissimilarity of E. coli clonal groups based on Bray-Curtis distance for (a) Flint Creek (FC) and (b) the Hoosic River (HR). (c) Partial Mantel correlation between the dissimilarity of E. coli clonal groups and geographic distance-correlated dissimilarity of soil variables for FC and HR. r is the partial Mantel correlation coefficient. Significant correlations at the 0.05 and 0.01 levels are indicated by single and double asterisks, respectively.

To identify the key soil variables correlated with the dissimilarity of E. coli clonal groups, we performed partial Mantel tests by correcting the correlation of geographic distance and biological dissimilarity. For Flint Creek, partial Mantel tests showed that geographic distance-corrected phosphorus and antimony concentrations in soil were significantly correlated with the dissimilarity of E. coli clonal groups (correlation coefficient r = 0.18 and P < 0.05 and r = 0.41 and P < 0.05, respectively) (Fig. 3c). In addition, for sites in Flint Creek, we observed that the mean concentration of phosphorus was significantly higher in soil samples from produce field sites than those from forest sites (Mann-Whitney test P < 0.05) (Fig. S5a). Also, in Flint Creek, the mean concentration of antimony was higher in soil samples from produce field sites than those from forest sites, but the difference was not significant (Mann-Whitney test P = 0.45) (Fig. S5b). Of note, in our initial analyses, soil phosphorus and antimony concentrations showed high correlation with potassium (Pearson correlation coefficient r = 0.72 and P < 0.05) and lead (r = 0.79 and P < 0.05), respectively; these covariations are consistent with the facts that (i) phosphate is often applied to land as potassium phosphate and (ii) lead and antimony often co-occur in artificial alloys. In contrast to the findings for Flint Creek, none of the soil variables were found to be significantly correlated with the dissimilarity of E. coli clonal groups from the Hoosic River (P > 0.05) (Fig. 3c). This result was consistent with our observation, based on VPA, that soil variables minimally contributed to the distribution of E. coli clonal groups in the Hoosic River watershed.

Relationship between E. coli dissimilarity and geographic distance.

We first assessed the relationship between dissimilarity of E. coli clonal groups and geographic distance using Mantel tests. Results showed a weak correlation between the dissimilarity of E. coli clonal groups and geographic distance for Flint Creek at the 0.1 level (r = 0.16; P = 0.08) (Table S7). In contrast, no significant correlation at the 0.1 level was observed between the dissimilarity of E. coli clonal groups and geographic distance for the Hoosic River (r = 0.11; P = 0.12) (Table S7). Linear regression analysis further showed that the slope of the regression line of the linear relationship between the dissimilarity of E. coli clonal groups and geographic distance for Flint Creek (slope = 3.4 × 10⁻³; R² = 0.027) (Fig. 4a) was about 3 times steeper than that for the Hoosic River (slope = 9.7 × 10⁻⁴; R² = 0.011) (Fig. 4b), indicating a stronger distance-decay relationship in E. coli in Flint Creek than the Hoosic River. Overall, results of the Mantel test and linear regression analysis along with the VPA suggest that spatial factors play a more important role in driving the distribution of E. coli clonal groups in Flint Creek than the Hoosic River, and the dispersal of E. coli in Flint Creek was slightly limited, while E. coli in the Hoosic River was more likely not constrained by dispersal limitation.

FIG 4.

Linear relationship between biological dissimilarity of E. coli clonal groups and geographical distance for (a) Flint Creek (FC) and (b) the Hoosic River (HR). The biological dissimilarity of E. coli clonal groups was calculated as Bray-Curtis distance. Geographical distance was calculated as the actual physical distance. The linear regression line is in gray; the shaded area is the 95% confidence region. R² indicates the variability explained by the fitted linear regression model, and the formula of the linear relationship is shown.

To evaluate the impact of land use on the relationship between E. coli and geographic distance, we conducted Mantel tests and assessed the distance-decay relationship for produce field sites and forest sites separately within each watershed. For Flint Creek, Mantel tests showed that the correlation between the dissimilarity of E. coli clonal groups and geographic distance was not significant at the 0.1 level; the correlation coefficient for forest sites (r = 0.27) was slightly larger than that for produce field sites (r = 0.22) (Table S8). For the Hoosic River, the dissimilarity of E. coli clonal groups and geographic distance for forest sites was significantly and highly correlated (P < 0.05; correlation coefficient = 0.49), while there was no significant correlation for produce field sites (P = 0.56) (Table S8). Consistent with this, linear regression analysis assessing the distance-decay relationship showed that the slope of the linear regression line for forest sites in both Flint Creek and the Hoosic River (slope = 7.0 × 10⁻³ and 4.3 × 10⁻³, respectively) was steeper than that for produce field sites (slope = 5.1 × 10⁻³ and −3.0 × 10⁻⁴, respectively) (Fig. S6a and b). These results suggest that dispersal limitation for E. coli tends to be weaker in produce field sites than in forest sites, which implies that dispersal of E. coli may be more efficient in the produce fields.

Wildlife-driven dispersal of E. coli.

Four common classes of wildlife vehicles (large nuisance wildlife species, small mammals, small flocking insectivores/granivores, and migratory bird flocks) were selected for identifying potential dispersal vehicles of E. coli (characteristics of these dispersal vehicles are detailed in Table 1). By adjusting distances among sampled sites to account for movement preferences of these four types of wildlife vehicles (i.e., cost-distance or landscape resistance modeling), we sought to assess whether dispersal associated with wildlife behavior explains the E. coli distribution better than distance alone. As shown in Table 2, the predicted dispersal model was developed based on the most likely cost and attraction models selected for each wildlife vehicle according to their characteristics. The predicted dispersal model for small mammals was defined to have a biological riparian corridor effect, no proximity effect, absolute dispersal barrier effect, and no attraction coefficient. The predicted dispersal model for large nuisance wildlife species was defined to have a biological riparian corridor effect, strong proximity effect, porous dispersal barrier effect, and habitat quality coefficient. The predicted dispersal model for migratory bird flocks was defined to have a biological riparian corridor effect, weak proximity effect, no dispersal barrier effect, and an area-independent coefficient. The predicted dispersal model for small flocking insectivores/granivores was defined to have a biological riparian corridor effect, weak proximity effect, absolute dispersal barrier effect, and an area-independent coefficient. Definitions of these effects can be found in Table 1.

TABLE 1.

Basic dispersal model characteristics

Class and abbreviation	Description
Dispersal vehiclesa
Large nuisance wildlife species (e.g., deer or feral swine) (LT)	Forests, scrublands, grasslands, wooded wetlands, and cultivated croplands impose low movement costs; roads impose increased movement costs based on census category
Small mammals (e.g., shrews, voles, and mice) (ST)	Light urban development, grassland, and scrublands impose low movement costs; roads impose increased movement costs based on census category
Small flocking insectivores/granivores (e.g., starlings) (LB)	Open water and heavy urban development are higher cost; all other movement costs are low
Migratory bird flocks (e.g., Canada goose) (MB)	Heavy urban development imposes higher cost; all other movement costs are low.
Riparian corridor (movement costs are reduced by half)
Adjacency	All land parcels that overlap a 100-m zone around the main river/creek
Distance	Land within 100 m of the main river/creek
Biological	Land below the 50-yr flood height for the main river/creek and adjacent wetlands
None	No riparian corridor effect
Dispersal barriers
Absolute	Major roads and waterways are absolute barriers (movement cost, 40,000 per pixel)
Porous	Major roads and waterways are porous barriers (movement cost, 200 per pixel)
None	No barrier effect
Proximity effects (specifics vary by vehicle)
Strong	Nearness to high-quality habitat substantially reduces movement cost
Weak	Nearness to high-quality habitat somewhat reduces movement cost
None	No benefit of proximity to good cover
Attraction (gravity) coefficients (specifics vary by vehicle)
Habitat quality	Proximity, interspersion, and area of high-quality habitat increase the chances that E. coli will be deposited
Reduced habitat effect	Effect of high-quality habitat is reduced by half
Area independent	As habitat quality model, but area of high-quality habitat does not impact the result
None	Attraction does not influence dispersal
Load (source) coefficient
E. coli load estimation	Areas close to forest- or pasture-class landcover are higher load

^aConceptual models are based on published literature (75–86).

TABLE 2.

The predicted dispersal model for each wildlife vehicle

Vehicles	Most likely cost modela			Most likely attraction model (coefficient)b
	Riparian corridor	Proximity effects	Dispersal barriers
ST (small mammals)	Biological	None	Absolute	None
LT (large nuisance wildlife species)	Biological	Strong	Porous	Habitat quality
MB (migratory bird flocks)	Biological	Weak	None	Area independent
LB (small flocking insectivores/granivores)	Biological	Weak	Absolute	Area independent

^aCost model was developed based on three factors: riparian corridor, proximity effects, and dispersal barriers. Definitions of the three factors can be found in Table 1. The most likely cost model was selected based on the characteristics of each wildlife vehicle (Table 1).

^bDefinitions of classes of attraction model can be found in Table 1. The most likely attraction model was selected based on the characteristics of each wildlife vehicle (Table 1).

Mantel tests showed that none of these dispersal models significantly predicted the composition of E. coli clonal groups in Flint Creek (P > 0.05), while two wildlife-driven dispersal models—dispersal via migratory bird flocks (M459) and via small flocking insectivores/granivores (M556)—were found to be significantly correlated with the dissimilarity of E. coli clonal groups in the Hoosic River (r = 0.17 and r = 0.16, respectively; P < 0.05) (Fig. 5). In addition, the model for dispersal via large nuisance wildlife species (M377) was marginally significantly correlated with the dissimilarity of E. coli clonal groups (r = 0.14, P = 0.056), so the role of dispersal of E. coli by large nuisance wildlife species cannot be absolutely excluded. Thus, migratory bird flocks, small flocking insectivores/granivores, and large nuisance wildlife species were identified as potential dispersal vehicles which were associated with the distribution of E. coli in the Hoosic River. The observation that cost-distance model correlated with the dissimilarity of E. coli clonal groups in the Hoosic River better than geographic distance alone (r = 0.11; P = 0.12) (Table S7) suggests some dispersal among sites by the action of wildlife. Our results also suggest that wildlife may play a more important role in affecting the distribution of E. coli in produce fields in the Hoosic River watershed (predominated by forests) than in produce fields in the Flint Creek watershed.

FIG 5.

Mantel test result of wildlife-driven dispersal models and composition of E. coli clonal groups for Flint Creek and the Hoosic River. M106, M377, M459, and M556 on the x axis are the identification numbers of the predicted dispersal models for small mammals, large nuisance wildlife, migratory bird flocks, and small flocking insectivores/granivores, respectively. The description of the predicted dispersal model for each wildlife vehicle is detailed in Table 2. Models that are significant or marginally significant at the 0.05 level in Mantel tests are indicated with asterisks.

DISCUSSION

E. coli has widely been used as an indicator of fecal contamination (37), which may suggest the potential presence of other pathogenic enteric bacteria in water (38). E. coli comprises a wide spectrum of phenotypes, including harmless commensal variants as well as distinct pathotypes with the capacity to cause either intestinal or extraintestinal infections in humans and many animals (39). The fecal-oral transmission route of E. coli often involves transient presence in extrahost habitats (e.g., surface water, soil, and plant surfaces), possibly including produce fields (23). Therefore, understanding the ecology of E. coli in extrahost habitats not only will provide an improved understanding of E. coli interaction with such environments but will also benefit public health by providing knowledge that can be used to minimize introduction of E. coli and possibly other enteric pathogens into preharvest produce-growing environments.

Environmental stressors, such as limited availability of nutrients and water, presence of toxic molecules, and large alterations in temperature and moisture, can impose a fitness cost on E. coli and other microbes (40). Fragmented landscapes with smaller forest and grassland patches expose surface soil to sunlight and greatly increase daily perturbations in soil conditions. Reduced forest and grassland cover could also hinder the movement of wildlife, increasing genetic isolation, drift, and extinction of bacterial subpopulations (41). In this scenario, landscape structure imposes constraints on environmental selection and dispersal, which is particularly essential for the dispersal of E. coli among different extrahost habitats.

To quantitatively probe the importance of environmental selection and dispersal in driving the distribution and composition of E. coli in soil under the impact of landscape, we compared the biogeographic patterns of E. coli isolated from two watersheds with distinct landscape patterns (i.e., Flint Creek, an area with widespread produce fields and limited interaction between produce fields and forest, and the Hoosic River, a heavily forested area with strong interaction between produce fields and forest). Our data specifically suggest that in the watershed with widespread produce fields and sparse forest coverage, environmental selection, possibly caused by soil phosphorus, and slightly limited dispersal may result in relative heterogeneous composition of E. coli between produce field and forest sites and potential local adaptation in E. coli. In contrast, in the watershed with heavily forested areas, no evidence of environmental selection was observed, and wildlife such as migratory bird flocks and small flocking insectivores/granivores may enhance the dispersal of E. coli populations across environments, resulting in relative homogeneous composition between produce field sites and forest sites in this watershed. This higher level of homogeneity is consistent with greater interaction between produce fields and forests in the Hoosic River watershed than in the Flint Creek watershed.

Agricultural practice involving input of phosphorus into soil may enhance the selective pressure on E. coli.

Agricultural activities normally involve cultivation and soil amendments, which can dramatically change soil organic matter and nutrient pools in comparison to undisturbed systems (e.g., forest) (42). Consequently, long-term organic and chemical amendments could dramatically impact the abundance, diversity, and composition of bacterial communities in soil of agricultural land (43). This is because such alterations of soil properties can trigger selective pressures on bacteria, selecting for the individuals or traits that better cope with modified soil condition, which has been termed “local adaptation” (23). For example, copper amendment in agricultural soil has been found to significantly increase the frequency of copper‐resistant Gram-negative bacteria (44). Based on the results of the study reported here, agricultural practices may have caused selective pressure on soil E. coli, partially resulting in the distinct E. coli composition between produce fields sites and forest sites in the Flint Creek watershed. Consistent with our findings, Dusek et al. (22) observed E. coli population structures that differed between cropland and forest, with much lower prevalence of E. coli in cropland than forest. The diverse lifestyles and phenotypes of E. coli strains have been hypothesized to be caused by population expansion paired with differential niche adaptation under specific selective pressures in the last 5 million years (39). The strong association between soil property and biological dissimilarity of E. coli observed in the Hoosic River watershed suggests that agriculture-stimulated selective pressures may contribute to E. coli diversification.

Besides directly yielding selective pressures on E. coli, soil property alterations caused by agricultural activities may also indirectly impact the adaptation of E. coli by changing the interaction with other microbial taxa in the community. For example, E. coli has been reported to exhibit bacteriocin-mediated competitive interactions (45) and cooperative interactions using cross-feeding of metabolic products with other taxa in the microbial community (46). Changes in microbial community structure and composition caused by agricultural activities thus may also generate selective pressures on E. coli by altering the competition and cooperation behaviors with other microbial taxa.

The high correlation between phosphorus and antimony and the dissimilarity of E. coli composition observed in the watershed with widespread produce fields (Flint Creek) suggests that the alteration in these two soil variables (or covariant soil variables, i.e., potassium and lead, respectively) may change the structure of E. coli populations in Flint Creek. Phosphorus is one of the soil variables that are well documented to dramatically change after the conversion of undisturbed systems to agriculture (42, 47). The input of phosphorus in fertilizer (which typically contains potassium phosphate) and manure to agricultural systems has been reported to often exceed the output in harvested crops (48). Phosphorus availability could act as an important selective force driving the structure and function of bacteria. For example, Bergkemper et al. (49) reported that phosphorus richness and depletion in forest soils could drive bacterial communities toward a higher potential for inorganic phosphorus solubilization and efficient phosphate uptake, respectively. Antimony is a toxic metalloid present widely at trace concentrations in natural soil (50, 51). Its concentration could be elevated or even reach a contamination threshold in agricultural lands due to human activities (52). A previous study showed that increased antimony could prevent the growth of E. coli, Bacillus subtilis, and Staphylococcus aureus and may affect the nitrogen cycle in soil by changing urease activity under neutral pH (53). However, the concentration of antimony detected in this study was relatively low across all sites, and for some sites the concentration was below the detection limit; thus, the presence of antimony may not necessarily inhibit the growth of E. coli in the soil studied here. Therefore, it is also possible that the observed associations with phosphorus and antinomy represent proxies for agriculture- or human activity-related selective pressures on E. coli in which other physical soil parameters may have played a role in the E. coli composition observed in the watershed with widespread produce fields.

E. coli in a watershed with high forest coverage may experience very weak selective pressure and a proximity effect of forest.

In this study, environmental selection tended to be very weak on E. coli in the watershed with higher forest coverage (Hoosic River). This relatively weak environmental selection might be because, in contrast to produce fields, plant cover and shading in forest could moderate perturbations in soil moisture, nutrients, and temperature, thus providing more favorable and stable conditions with fewer environmental stressors for E. coli (4, 22). As previously proposed (4), soil in undisturbed temperate forests could act as a potential habitat for long-period persistent, even resident E. coli populations rather than acting as a transient habitat. Although E. coli may be exposed to fewer or less intense stressors in undisturbed environments, compared to disturbed ones, some factors, such as temperature, moisture, and nutrients, have been shown to be correlated with E. coli density in forest (54, 55). Due to the lack of niche differentiation caused by environmental selection, we observed more homogeneous E. coli compositions between forest and produce field in the watershed with higher forest coverage. We also observed that E. coli was much more prevalent in the watershed with higher forest coverage (72%) than in the watershed with lower forest coverage (35%), consistent with previous findings by Dusek et al. (22).

The higher prevalence of E. coli in the watershed with higher forest coverage might be caused by a proximity effect, which proposes that the likelihood of E. coli isolation from surrounding sites such as produce fields increases with the proximity to forests (22). Such a proximity effect is formed by the spread of E. coli out of forests into surrounding areas, given that forest is a vital sink for E. coli (4). In addition, the high degree of adjacency between forest and produce fields in the watershed with higher forest coverage, which indicates strong direct interactions between the two land covers, may enhance the proximity effect. Consistent with our findings, Dusek et al. (22) reported that E. coli was more prevalent in a landscape with greater forest coverage; that study specifically showed that E. coli was most prevalent in soils sampled in close proximity (0 to 38 m) to forests but was up to 90% less prevalent when forest cover in a 250-m radius was less than 7%. In addition to E. coli, such proximity effects of forest have also been reported for Listeria monocytogenes and other Listeria species. Weller et al. (56) found that with a 100-m increase in the distance of a sampling site from forests, the likelihood of isolation of L. monocytogenes and other Listeria species in croplands decreased by 14% and 16%, respectively. While the consistency between our findings and previous data, as discussed here, suggests that one may be able extrapolate from our findings to other locations, particularly those in New York State and other locations with similar climatic and ecological characteristics, it is important to emphasize that our findings are based on data collected over one season in two locations and that further studies in other locations and over longer time frames are needed.

Watershed landscape may impact the distribution of soil E. coli by influencing the movement of wildlife hosts.

Wildlife, which is thought to be an important vehicle for the dispersal of foodborne pathogens between hosts and locations (36), could enable bacteria to overcome landscape barriers and make the dispersal of bacteria, including E. coli, more active, in particular through defecation by animal hosts (11, 57). Migratory bird flocks, small flocking insectivores/granivores, and large nuisance wildlife were identified, in our study here, as potential vehicles that disperse E. coli and affect the distribution of E. coli in the watershed with high forest coverage. Migratory bird flocks (e.g., Canada goose) tend to have low movement cost in all land use types except for heavily developed urban areas (58). Small flocking insectivores/granivores (e.g., European starling) tend to have low movement cost in all land use types except for open water and heavily developed urban areas (58). Large nuisance wildlife (e.g., white-tailed deer and feral swine) tend to have low movement cost in forests, scrublands, grasslands, wooded wetlands, and cultivated croplands, while they can undergo high movement cost when crossing roads (59).

Each of these three classes of wildlife was reported previously (34, 35, 60) to serve as a dispersal vehicle of E. coli and identified as a concern in terms of agricultural contamination with foodborne pathogens. For example, a study of E. coli presence in fecal samples from Canada geese over a year in Colorado, USA, reported that the prevalence of E. coli ranged from 2% during the coldest months to 94% during the warmest time of the year (60). European starlings, which are considered an invasive species in the United States and a nuisance pest to agriculture, were also proposed to be a potential suitable reservoir and vector of E. coli O157:H7 and have been reported to carry and disseminate this human pathogen to cattle (34). In addition, deer feces were reported to contaminate fresh strawberries, being responsible for an outbreak of E. coli O157:H7 infections in Oregon (35). Forest is a relatively stable environment with less disturbance of anthropologic activities, thus serving as an ideal living habitat for wild animals (61). Forest may provide easy transport pathways for small flocking insectivores/granivores and large nuisance wildlife to move around and support high density of migratory birds, which increases the chance of colonizing E. coli being dispersed from forest to adjacent produce fields. In contrast, the observation that dispersal of soil E. coli in the watershed with widespread produce fields was relatively limited could be explained by the poor connectivity of agricultural areas, which may impede the movement of wildlife that disperses E. coli.

Besides wildlife, it is also possible that the dispersal of E. coli was directly influenced by the landscape elements of the two watersheds. Forest and most produce fields in the Hoosic River region, which is heavily forested, are both located in a floodplain. In contrast, forest in Flint Creek is in a floodplain, but produce fields are not. Since during periods of high discharge, a floodplain normally experiences flooding, such events may facilitate the dispersal of E. coli between forest and produce field, particularly in the Hoosic River watershed. This hypothesis is supported by a number of modeling studies, showing that the peak fecal bacterial levels during flooding can be more than 20 or even 50 times higher than prior to flooding (62–64). Future studies comparing the distribution of E. coli before, during, and after flooding and assessing the correlation between flooding-associated landscape factors (e.g., elevation and patchiness) and distribution of E. coli are needed for an improved understanding of the impact of landscape on the microbial biogeography.

Conclusion.

By comparing the biogeographic patterns of E. coli isolated from two watersheds with distinct landscape characteristics in New York state, we showed that terrestrial landscape could impact the distribution of E. coli by adjusting the importance of environmental selection and dispersal. Environmental stress was identified as a possible strong contributor to local adaptation of E. coli in the watershed with widespread produce fields. On the other hand, wildlife-driven dispersal, which could facilitate genetic exchange, was identified as a major force in shaping E. coli populations in the watershed with high forest coverage. Past studies have focused predominantly on detection of reservoirs of E. coli and adaptations of E. coli in agricultural landscapes without attempting explicit measurement of dispersal due to various sources (65–68). Our findings not only highlight the critical role of landscape in driving the biogeographic pattern of E. coli in perspective of ecology but also open the possibility that the evolutionary forces (e.g., positive selection, genetic drift, and gene flow) driving the diversification of E. coli vary by watershed landscape as well. In addition, our study suggests that due to the less intense environmental stress, frequent wildlife-facilitated dispersal, and proximity effect of forest on E. coli, produce fields in watersheds with high forest coverage may have a higher risk of E. coli contamination. This information can inform spatial modeling of food contamination risk associated with produce fields in different watersheds, which can be used to modify preharvest product sampling strategies and produce harvest methods to account for the spatial structure in contamination risk in a produce field.

Despite these contributions, our study has some nonnegligible limitations. First, we did not differentiate commensal and pathogenic E. coli, whose biogeographic patterns may differ. Second, we assessed the coverage of forest and produce field as only one landscape attribute using a relatively small set of sampling sites, while many other landscape attributes, such as the size, diversity, and richness of patches, may also be important for the biogeographic pattern and adaptation of E. coli. Third, wildlife population/community structure, which could be strongly affected by land use features (69), was not included in the competing models. To better understand the impact of the landscape on the biographic pattern and adaptation of E. coli, future studies warrant more intensive sampling efforts on a broader scale, sequencing techniques with higher discriminatory power (e.g., whole-genome sequencing), and comprehensive assessment of a wider range of factors, including landscape attributes, wildlife characteristics, and seasonal effects. Such future efforts, building on our data reported here, could considerably improve prediction of produce contamination risk based on the potential influence of landscape on the dispersal of E. coli to produce field, benefit the development of trade-off risk assessments of food contamination, and eventually help to decrease human exposure to pathogenic enteric bacteria.

MATERIALS AND METHODS

Study sites and soil collection.

Two watersheds with different landscape patterns, Flint Creek and the Hoosic River, located in New York State, were selected for this study based on topography and land cover composition. The Flint Creek watershed is an area with widespread vegetable and livestock production that is sparsely forested (69% produce field and 12% forest by area), whereas the Hoosic River watershed is a heavily forested area with interspersed produce production (28% produce field and 38% forest by area). Soil sampling was carried out between 4 September and 10 October 2012 on 7 farms comprising 19 produce field sites and in 16 forest sites (Fig. 1). For produce fields, two parallel 200-m transects were laid in each field, perpendicular to the forest boundary. Along each transect, five soil samples (at a depth of approximately 5 cm) were collected at 50-m intervals using sterile scoops (Fisher Scientific, Hampton, NH) and sterile Whirl-Pak bags (Nasco, Fort Atkinson, WI). Latex gloves and disposable plastic boot covers (Nasco, Fort Atkinson, WI) were worn for sample collection. Gloves and boot covers were changed between sites, and gloves were disinfected with 70% ethanol prior to sample collection. A total of 278 soil samples were collected, 142 and 136 from the Flint Creek and the Hoosic River watersheds, respectively. All samples were transported to the Food Safety Lab at Cornell University in an icebox. Samples were stored at 4 ± 2°C in the dark and processed within 24 h of collection.

Isolation of E. coli.

E. coli strains were isolated from soil samples as previously described (23). Briefly, 8 g sieved soil was diluted 1:10 in Escherichia coli medium with 4-methylumbelliferyl-β-d-glucuronide broth (EC-MUG). To maximize genetic diversity among recovered E. coli isolates, the suspension was subdivided among four 96-well microtiter plates for a total of 384 subsamples of approximately 180 μl each. Microtiter plates were incubated at 37°C. Bacteria from fluorescent wells were isolated on EC-MUG agar plates and were further tested with a standard biochemical assay for glutamate decarboxylase and β-glucuronidase activity. Isolates that were positive for these two tests were presumptively identified as E. coli, which was confirmed by subsequent gene sequencing as detailed below. No E. coli isolates were detected in samples from 3 produce sites (field 7, field 9, and field 10) and 2 forest sites (forest F4 and forest F18).

DNA extraction, MLST genotyping, and clonal groups.

Genomic DNA was extracted from E. coli by alkaline lysis of biomass in 50 mM NaOH at 95°C. Two genes (mdh and uidA) were sequenced first from all isolates. Then, only the unique two-gene sequence types from each sample were subjected to additional sequencing for five additional genes (aspC, clpX, icd, lysP, and fadD) by Sanger sequencing, performed by the Cornell University Life Sciences Core Laboratories Center. Evaluation of sequence read quality and assembly of forward and reverse reads were performed using Perl scripts, which iterated runs of phred and CAP3, respectively. Sequences with a probability of error of >0.005 (Q score < 23) in terms of read quality were edited manually, where possible, or discarded. Assembled sequences of each MLST locus were aligned and trimmed to standard base positions matching the E. coli K-12 sequence type from the STEC Center website (http://www.shigatox.net) (23). Alignments of assembled sequences for isolates from Flint Creek and the Hoosic River are available on GitHub (https://github.com/pbergholz/Dispersal-cost-modeling). The clonal groups of E. coli strains for Flint Creek and the Hoosic River were determined based on MLST sequence types using the goeBURST analysis program (70) at the single-locus-variant level.

Remotely sensed data and soil property data.

GPS (global positioning satellite) coordinates of sites were imported into the Geographical Resources Analysis Support System (GRASS) geographic information system (GIS) environment. Map layers for land cover (National Land Cover Database [NLCD], 2006 [https://www.mrlc.gov/data/nlcd-2006-land-cover-conus]) and the digital elevation model (DEM; Shuttle Radar Topography Mission; 1-arc-s data set) were acquired from the U.S. Geological Survey (USGS) Earth Explorer geographical data bank (https://earthexplorer.usgs.gov/). Map layers for soil characteristics were acquired from the U.S. Department of Agriculture Soil Survey Geographic (SSURGO) database (https://websoilsurvey.sc.egov.usda.gov/App/HomePage.htm). Road and hydrologic line graphs were obtained from the Cornell University Geospatial Information Repository (CUGIR; http://cugir.mannlib.cornell.edu/).

Percent land cover and adjacency were estimated by using FRAGSTATS v. 3.3 to analyze landcover within a 2-km buffer surrounding Flint Creek and the Hoosic River, respectively (71). All land that NLCD maps identified as pasturage was reclassified as cropland for our analyses. Percent adjacency was calculated as the proportion of pixels in the NLCD map that were adjacent forest and produce field, compared to the total of non-self-adjacencies in the 2-km buffer surrounding the waterway. For example, an adjacency_{produce|forest} of 10% would indicate that 10% of the edges of produce fields abutted forest in a given area.

Organic matter, moisture, pH, aluminum, arsenic, boron, barium, calcium, cadmium, cobalt, chromium, copper, iron, potassium, magnesium, manganese, molybdenum, sodium, nickel, phosphorus, lead, sulfur, strontium, and zinc contents of soil samples were measured at the Cornell Nutrient Analysis Lab.

Distribution of E. coli clonal groups and its relationship with geographic location.

The Mann-Whitney test was used to determine whether the number of clonal groups differed significantly between soil samples from produce field sites and forest sites for Flint Creek and the Hoosic River. Principal-coordinate analysis (PCoA) was conducted using the phyloseq package in R 3.6.0 to visualize the distribution of E. coli clonal groups among sites, based on Bray-Curtis distance. The 95% confidence ellipse in the PCoA plot assumes a multivariate normal distribution. Permutational multivariate analysis of variance (PERMANOVA) (72) was employed using the adonis function in R 3.6.0’s vegan package to test whether the centroids and dispersion of sample groups as defined by land use (produce field or forest) are equivalent for both groups based on Bray-Curtis distance of E. coli clonal groups. The PERMANOVA test statistic (F) and P value were obtained by 9,999 permutations. Analysis of similarities (ANOSIM) (72) was employed using the anosim function in R 3.6.0’s vegan package to test whether there is a significant difference between two groups (produce field sties and forest sites) of sampling units based on the Bray-Curtis distance of E. coli clonal groups. The ANOSIM test statistic (R) and P value were obtained by 9,999 permutations. To test the sampling bias which possibly caused the large variation of E. coli clonal groups observed across sampling sites within each watershed, sites with ≤3 clonal groups (field 6, field 8, and forest F9) were excluded, and PCoA, PERMANOVA, and ANOSIM were repeated for this subset of samples.

Mantel tests were performed using the vegan package in R 3.6.0 to assess the relationship between the biological dissimilarity of E. coli and geographic distance (9,999 permutations). Biological dissimilarity of E. coli clonal groups was calculated in Bray-Curtis distance using vegan package in R 3.6.0. Geographic distance between isolates was calculated from latitude and longitude, using the geopy module in Python 3.6.8. Linear regression analysis of biological dissimilarity of E. coli and geographic distance was performed in R 3.6.0. Distance-decay relationship was inferred from the slope and R² of the linear regression. A steeper slope with a larger R² value suggests a stronger distance-decay relationship.

Relationship between E. coli clonal groups and soil variables.

Due to a lack of mineral soil to measure soil properties after combusting away the organic matter, forest F6, forest F11, and forest F16 were not included in analyses on the relationship between E. coli clonal groups and soil variables. After screening for covariation, soil variables with low levels of covariation (r < 0.7 in Pearson’s correlation analysis) were selected for variation partitioning analysis (VPA) using the vegan package in R 3.6.0 to quantify the relative contribution of the environment effect and the geographical effect on the dissimilarity of E. coli clonal groups based on Bray-Curtis distance (73). Principal coordinates of neighbor matrices (PCNM) were used to represent spatial patterns based on GPS coordinates (74). By using the ordistep function in the vegan package, a subset of PCNM variables which significantly explained variation in the dissimilarity of E. coli clonal groups was included in VPA. Four components of variations were calculated in VPA: (i) pure contribution of environmental effect (R²_A – R²_G); (ii) pure geographical effect (R²_A – R²_E); (iii) spatially structured environmental effect (R²_G + R²_E – R²_A); and (iv) unexplained effect (1 – R²_A). R²_A, R²_G, and R²_E represent variation of the dissimilarity of E. coli clonal groups explained by all variables, spatial variables, and environmental variables, respectively.

A partial Mantel test was performed to examine the correlation between environmental dissimilarity of each soil variable and the dissimilarity of E. coli clonal groups independent of geographical influence using the vegan package in R 3.6.0 (9,999 permutations). The dissimilarity of E. coli clonal groups was calculated in Bray-Curtis distance, and environmental dissimilarity was calculated in Euclidian distance. Soil variables with a P value of <0.05 in partial Mantel tests were defined as key soil variables. Mann-Whitney tests were further performed to determine if key soil variables differed significantly between soil samples from produce field sites and forest sites.

Dispersal model formulation and selection.

To predict the dispersal of E. coli across watershed landscapes, multiple dispersal models were developed to describe landscape effects by integrating remotely sensed and field-collected data into resistance surfaces for wildlife vehicles. Four common classes of wildlife vehicles, including (i) large nuisance wildlife species, (ii) small mammals, (iii) small flocking insectivores/granivores, and (iv) migratory bird flocks, were selected in this study.

Predicted dispersal among sites was calculated as follows: D_i,j = (L_i × A_j)/C_i,j, where D_i,j is the dispersal rate among sites i and j, L_i is the E. coli load from the source site (i.e., the starting point), A_j is the attraction (gravity) coefficient of the sink site (i.e., the stopping point) to the vehicle, and C_i,j is the least-cost distance between sites i and j. E. coli load (L_i) expresses the expected mobility of E. coli from these areas as a function of expected prevalence. Expected prevalence was inferred from random forest analysis of E. coli prevalence based on sampling excursions. One load map was generated per watershed. The attraction (gravity) coefficient (A_j) describes the tendency of a dispersal vehicle to move toward an area on the landscape and the expected residence time of the dispersal vehicle after it arrives at a location. Attraction was primarily a function of percent favored land cover for each of the vehicles and interspersion of land cover types. The least-cost distance (C_i,j) describes the movement preferences of a dispersal vehicle in terms of a friction surface (borrowed from circuit theory) that predicts resistance of the landscape to movement of dispersal vehicles. The cost surfaces were a function of baseline resistance (dependent on the dispersal vehicle), riparian corridor effect (i.e., the tendency of wildlife to prefer movement through riparian forests), dispersal barrier effect (i.e., the strength of barriers to movement, such as major road- and waterways), and proximity effect (i.e., the strength and type of edge interactions among forests, produce fields, pasturage, and urban areas). The least-cost distance was measured as the distance along the path that accrued the least cumulative cost between pairs of movement start and stop sites. The characteristics of the dispersal vehicles, E. coli load model (L_i), attraction model (coefficient A_j), cost model (i.e., riparian corridor effect, dispersal barrier effect, and proximity effect) are shown in Table 1 and are summarized on the basis of published literature (75–86). Based on these characteristics, the most likely attraction model and cost model were selected for each class of vehicle, generating the predicted dispersal models (Table 2).

For each of the predicted dispersal models for the four classes of wildlife vehicles, an association matrix (D_i,j) containing predicted dispersal rates along lowest-cost paths among all pairs of sites was generated. This was accomplished by using a set of scripts developed in the GRASS GIS v. 6.4.3 programming environment; Perl scripts were used to automate calculations in the GIS. Scripts are available on GitHub (https://github.com/pbergholz/Dispersal-cost-modeling). Mantel tests were employed to estimate the correlation between predicted dispersal models and biological dissimilarity of E. coli clonal groups among sampled sites in each watershed using R version 3.6.0. Statistical significance of model fits was estimated by 9,999 permutations. The wildlife vehicle for which the predicted model had the highest significant correlation coefficient was deemed to represent the dominant dispersal vehicle for E. coli.