Scale dependencies and generalism in host use shape virus prevalence

Abstract

Processes that generate the distribution of pathogens and their interactions with hosts are not insensitive to changes in spatial scale. Spatial scales and species traits are often selected intentionally, based on practical considerations, ignoring biases that the scale and type of observation may introduce. Specifically, these biases might change the interpretation of disease–diversity relationships that are reported as either ‘dilution’ or ‘amplification’ effects. Here, we combine field data of a host–pathogen community with empirical models to test the effects that (i) spatial scale and (ii) host range have on the relationship between plant–virus infection prevalence and diversity. We show that prevalence–diversity relationships are scale-dependent and can produce opposite effects associated with different habitats at sub-ecosystem scales. The total number of host species of each virus reflected generalism at the ecosystem scale. However, plasticity in host range resembled habitat-specific specialization and also changed model predictions. We show that habitat heterogeneity, ignored at larger (ecosystem) spatial scales, influences pathogen distributions. Hence, understanding disease distributions and the evolution of pathogens requires reconciling specific hypotheses of the study with an appropriate spatial scale, or scales, and consideration of traits, such as host range, that might strongly contribute to biotic interactions.

1. Introduction

Understanding the relationship between the distributions of pathogens in the environment and biodiversity is central to predicting disease threats to human and animal health, food security and wildlife conservation. Models for the prediction of disease risk that use biodiversity as an independent variable have been recognized as powerful tools for understanding transmission dynamics [1]. These models show that biodiversity may either increase or decrease disease risk and pathogen emergence [2–4]. A negative relationship between diversity and disease risk has been referred to as a ‘dilution’ effect. In this scenario, higher species diversity reduces disease risk [5,6]. The opposite relationship has been called an ‘amplification’ effect, where species diversity correlates positively with risk. Amplification effects may be related to pathogen ‘spillover’, when transmission from a reservoir promotes infection in other species [7,8]. However, it is not clear if ecological mechanisms can be generalized to the extent of justifying biodiversity as a key predictor of disease risk [4,9]. Instead, specific community composition characteristics rather than diversity per se are proposed to be better predictors of pathogen distributions [10–12]. For example, the experimental removal of non-competent lizard hosts of tick vectors of Lyme disease resulted in uneven host-shifting patterns among competent reservoirs and a net decrease in tick vectors and disease risk [13]. The opposite effect occurred when non-reservoir rodents were experimentally removed from sites resulting in hantavirus infection prevalence increases in wild reservoir species [14]. As such, disease–diversity relationships may be largely spatially dependent, because as the sampling area increases, then so does the observed diversity, influencing the relative abundance of suitable hosts and ecological processes including transmission [4,9].

Spatial variation in distributions partly determines whether species interact strongly with one another or not. The extent to which species distributions are realized in a given area is called a species–area relationship, and is central to ecology [15]. Species diversity is a function of their richness, abundance and evenness (the relative abundance of each species), and is a component of community structure. When the abundance of a species is a positive function of sampling area, local extinction is more likely in small areas and rates of colonization are higher in larger areas. In another sense, variability in diversity decreases from local to regional scales, creating a negative variability–area relationship [16–18]. Therefore, the number of habitat types in a community will determine species diversity. Although spatial scale relationships have been recognized in the field of ecology since the nineteenth century [15,18], they have been rarely considered explicitly and tested in pathogen systems [19–21]. However, knowledge of the spatial scales at which processes such as pathogen transmission occur is critical to identifying key disease risk factors. For instance, assessing West Nile virus transmission dynamics required that the spatial scale of the study capture the heterogeneity in vector host feeding and the variation in host resistance [22]. Observations over large areas have been required to ascertain the spatial scale at which transmission patterns explain host switching and emergence of Canine distemper virus in multi-species communities [23]. However, disease–diversity relationships estimated at large spatial scales might not reflect local-scale mechanisms where phenotypes become strong explanatory variables. For example, plant species adapted to higher nitrogen availability exhibit release from pathogens in exotic compared with their native ranges [24]. The aggregation of fungal pathogens that cause powdery mildew in a herbaceous perennial has been related to the genetic structure of host populations, providing evidence of selection for resistance at local spatial scales [19]. The spatial scale at which environmental heterogeneity promotes selection (e.g. for pathogenicity) will depend largely on the pathogen's potential to disperse between habitats suitable for its population persistence [25,26]. Therefore, evolutionary constraints in host range may not explain disease risk when ecological changes promote high rates of novel encounters, for instance under anthropogenic disturbance of biotic communities [27,28].

The degree of specificity between pathogens and hosts has given rise to models explaining the genetic basis of infection [29], and has consequences for infection dynamics. In particular, low specificity and broad host ranges have been correlated with increases in emergence or re-emergence in a diversity of pathogen phyla [30–33]. Movement ecology, biodiversity and the spatial scale at which selection is operating determine host range breadth [34,35]. Despite the widespread occurrence of host generalists, the selection for generalism and the evolution of pathogen host range in multi-species systems are poorly understood [33]. Constraints on the evolution of host range breadth are believed to result from selection for specialization in host traits [30]. In a homogeneous landscape, across-host fitness variation of generalists should theoretically favour selection for narrow host ranges [20]. According to this prediction, studies have demonstrated local-scale coevolution among genotypes [36,37] favouring narrow host ranges [21]. Because of across-host fitness trade-offs, the evolution of narrow host ranges has been linked to transmission between closely related host species [8,38]. Broad host ranges are possibly maintained by selection across heterogeneous environments [27]. Pathogen generalism implies that transmission regularly occurs among different host species [31,32,39]. Spatial relationships are therefore critical to realizing the extent of interactions generalist pathogens potentially encounter. If a generalist is subject to negligible fitness costs infecting multiple host species [40], the primary forces acting on host range evolution might therefore be ecological drift, dispersal and mutation rate [41,42]. The broader niche of generalists and concomitant phenotypic plasticity in trait interactions arising from habitat heterogeneity [43–45] should, therefore, be considered an important factor in transmission dynamics.

In light of seemingly unclear disease–diversity relationships, and of the recognized need to consider different spatial scales to understand disease dynamics, we decided to study how disease risk predictions changed according to the spatial scale at which diversity was measured in a multi-host–multi-pathogen system. Specifically, we tested the following two hypotheses: (i) that the relationship between infection prevalence and diversity changes depending on spatial scale, and (ii) that this relationship changes depending on virus host range. A measure of disease risk is prevalence, the proportion of individuals that are infected at a given time [46]. In this study, we used field observations of interactions between 11 virus and 83 plant species to test infection prevalence dependencies on plant diversity, spatial scale and virus host range. Standard ecological approaches were used to characterize environmental heterogeneity associated with different vegetation types in an agro-ecosystem. Species relative abundances were quantified at either the large (ecosystem) or partitioned at small (habitat) spatial scales to estimate diversity, and multivariate models were used to test the effect spatial scale had on prevalence. Environmental heterogeneity was, therefore, ignored at the ecosystem scale, and explicitly considered when diversity was partitioned among habitats. On top of this, host range observations of each virus were either included or omitted from the models to test the effect generalism in host use had on prevalence–diversity relationships.

2. Material and methods

(a). Sampling

Sampling of wild plant leaf tissue took place from 2000 to 2002 in Central Spain (see the electronic supplementary material for site details) with surveys made in each season (spring, summer, autumn and winter) from two sites in each of four vegetation types [32,47], henceforth referred to as habitats: (i) the boundary immediately adjacent to cultivated fields (edge), (ii) agricultural fields between crops (fallow), (iii) sites with no specific use interspersed among crops (wasteland), and (iv) cultivated melon fields (crop). Crops were sampled in two seasons only because they are absent from the ecosystem during winter and spring. Fallow plant communities reassemble de novo at different times through the year owing to tillage cycles. Edge communities are not ploughed over at any time and have relatively stable substrates. Wasteland remains largely untouched by anthropogenic disturbance. Habitats were dispersed over an area of 10 km²: the spatial relationships between edge and wasteland were constant, but the location of fallow and crop varied every year. As the homogeneity of infection prevalence is difficult to test a priori, sample size was based on extensive experience with disease systems in Central Spain and assessed using standard ecological approaches (see below). Briefly, leaf samples were collected from 25 plants at fixed points along transects at each site regardless of showing symptoms. Samples were not collected if a plant was not present at a fixed transect point. Under this systematic sampling, we assume that comparative frequencies of the observed infections represent realized interactions.

(b). Virus detection

Plant collections were analysed using double-antibody sandwich enzyme-linked immunosorbent assay (DAS-ELISA: [48] to detect 11 ssRNA viruses: alfalfa mosaic virus (AMV), beet western yellow virus (BWYV), bean yellow mosaic virus (BYMV), cucumber mosaic virus (CMV), lettuce mosaic virus (LMV), papaya ringspot virus (PRSV), potato virus Y (PVY), tomato spotted wilt virus (TSWV), turnip mosaic virus (TuMV), watermelon mosaic virus (WMV) and zucchini yellow mosaic virus (ZYMV). These viruses (electronic supplementary material, table S1) are transmitted by aphid vectors in a non-persistent manner, except BWYV, which is transmitted by aphids in a persistent, non-propagative manner, and TSWV, which is transmitted by thrips in a persistent propagative way [49]. A sample was considered as infected by a certain virus if the absorbance in the ELISA assay was greater than or equal to 2× negative controls. For each virus, absorbance of negative controls across host-plant species belonged to the same distribution, so that no strong bias is expected in the detection of any virus in the different potential hosts. The ELISA infection data comprised more than 16 148 observations over 3 years (2000–2002) in each season, habitat and for 83 plant species. The ELISA detection data included 468 rows where infection was not determined (n.d.) in at least one virus, leaving a total of 1468 rows (‘complete’ infection records for 11 viruses) comprising 71 plant species. Of the 1468 rows, 608 supported infections of at least one of the 11 viruses and were used to construct the interaction networks (see below). From the 83 plant species collected, infection was detected in 47 by at least one virus.

(c). Environmental heterogeneity and host range analyses

Sampling bias was evaluated using spatial autocorrelation testing, rarefaction analyses and by comparing observed richness with richness estimates derived from bootstrap and non-parametric tests. Correspondence analysis (CA) of a habitat-species abundance matrix was used to assess distinctiveness among the four habitats. We examined the probability distributions of the predictor and response variables using quantile comparison plots and Bartlett tests to assess variance between groups. Pearson's product-moment coefficient was used to estimate correlations among variable distributions to assess their suitability as predictors in the linear models (figure 1). The statistics were conducted using the R [50] packages car [51] and MASS [52]. The infection data associated with each habitat were graphed to visualize infection patterns and used to calculate host ranges. Links between 11 virus nodes and 47 host-plant nodes represented each observed infection (figure 2; electronic supplementary material, figure S1). Network graphs were implemented in the R package [53]. We used the neighbourhood analysis function to count the number of links between viruses and plants, to estimate virus host range and the abundance of each host taxon (electronic supplementary material, figure S2). Kernel density estimation (KDE) was used to visualize changes in the distribution of infections by a virus across habitats with distinct plant compositions; a host range effect. It is a useful non-parametric technique that makes no assumptions about probability distributions to estimate density functions that are scaled, allowing direct comparisons. The kernels are scaled by standard deviation of the smoothing kernel and integrate to one. We selected Silverman's rule-of-thumb to estimate the bandwidth in the KDE analysis. The low smoothing bias results in overfitting the function making it easy to compare differences in (non-Gaussian) infection distributions among habitats.

Figure 1.

Virus prevalence versus host abundance and host richness. Higher numbers of individual hosts (regardless of species distinctions or infection status) correlated with the infections a virus achieves at the ecosystem scale. Log mean virus prevalence was positively correlated with the (log) number of hosts (Pearson's r(9) = 0.772, p = 0.005, t = 3.6508) and (log) richness (Pearson's r(9) = 0.889, p = 0.0002, t = 5.8323). (Online version in colour.)

Figure 2.

Plant–virus infection network graphs in the four habitat categories. The graph edges are weighed according to infection frequency. The graph vertices are weighted by plant abundance and virus host range across the entire ecosystem. The rings surrounding the graphs differentiate plants (mid-grey) from viruses. Species are positioned identically between each graph.

(d). Data and statistical modelling

To estimate prevalence, the number of infections and the total number of individuals can either be summed across viruses for each host (host prevalence) or across host species for each virus (virus prevalence) [32]. Each of these summations requires that the data be formatted differently (as ‘long’ and ‘wide’, respectively). To relate diversity directly to disease prevalence, we used the wide format. Additionally, confidence intervals of prevalence estimates of each virus were improved by collapsing by season, decreasing sample bias. The wide format also allows the inclusion in the model of a fixed variable for each virus host range. Host range was estimated at different spatial scales; first by summing infections over all observed plant taxa exploited by a virus in the ecosystem (i.e. fundamental host range), and second by summing over spatially defined subsets of the data associated by habitat type (i.e. realized host range). In this way, host ranges were estimated to compare the habitat spatial scale with quantities estimated from the ecosystem. The data for the virus mean prevalence models were summed without factors for year or season, and estimates of diversity were calculated either at the habitat (i.e. four fixed factors in one model) or ecosystem (i.e. one fixed factor in another model) spatial scales.

Another model structure with the long format and a host prevalence response was used to assess temporal effects. This allows diversity to be estimated with the inclusion of species where no infection was detected. Diversity was estimated either for each year-season-habitat or year-season combination to adjust the ‘scale of sampling’. The long data format forces diversity estimates to be constant across all the viruses, but differing for each year, season or habitat combination depending on the model (see the electronic supplementary material for details).

Generalized linear models (GLM) using the wide format were used to compare the explanatory power of Shannon–Wiener (H) and Tsallis entropy (S_q; see the electronic supplementary material) diversity estimates on a virus prevalence response. The Tsallis estimate S_q of diversity was selected because prevalence–diversity hypotheses might rely on quantifications that are sensitive to species evenness [54,55]. Variables for abundance, richness, H, and S_q diversity indices were used as fixed factors. The predictor variables in the virus prevalence models (abundance, richness, host range, H, S_q) were scaled to adjust unit values. Bartlett tests did not reject the null hypothesis that variances of each variable from each habitat were the same (richness, p = 0.946; abundance, p = 0.4097; H, p = 0.6907; S_q, p = 0.7533). Explicit tests of the diversity–prevalence relationship between habitats and at the ecosystem spatial scale were also conducted using GLMs. Generalized linear mixed models (GLMM) were used to test the effects that random factors for year and season had on host prevalence response (i.e. with the long format). These models also included a random factor for virus, and fixed factors for Tsallis S_q. Models with or without a fixed factor for habitat were used to compare diversity at each spatial scale. All linear modelling used gamma variance and log-link functions (GLM and GLMM) and was conducted using the R packages car and lme4 [56].

3. Results

(a). Scale-dependent model effects on virus prevalence

Ordination and rarefaction approaches were used to characterize the distinctiveness of each habitat (see the electronic supplementary material for details). Chao and bootstrap richness estimates agreed with our sample estimates (electronic supplementary material, table S2) and rarefaction curves were close to asymptotic (electronic supplementary material, figure S3). There was no indication of spatial autocorrelation. CA showed that seasonal collections of each habitat type clustered strongly in distinct groups indicating mostly similar species compositions among seasons (electronic supplementary material, figure S4). The CA indicates that each habitat shared species compositions across seasons and sites, but with abundances of one or a few species shared across habitats producing an over-dispersed pattern between seasons (second axis) in fallow. Infection network graphs indicated skewed infection frequency distributions (as indicated by wider edges) between particular virus–plant pairs in a given habitat (figure 2). For instance, CMV infection occurred with high frequency in Diplotaxis erucoides in all habitats except edge, in Amaranthus sp. in crop and moderately in fallow, and in Medicago sativa in edge and M. sativa supported high frequency of AMV infection in the edge and D. erucoides in fallow. There were several high frequency infections of AMV, BWYV, CMV and TSWV in various hosts in fallow. By contrast, infection frequencies were more evenly distributed in wasteland. Additionally, the relatively low abundance of hosts such as D. erucoides in crop and Cirsium arvense in fallow associated with high frequency of CMV infection indicates a host preference given the presence of other host species within its range in these habitats. These patterns were consistent with shifts in the number of infections by AMV, CMV and TSWV using kernel density estimates across habitats (figure 3).

Figure 3.

Kernel density estimates of the number of individuals in each host infected by AMV, CMV, PVY and TSWV in the four habitats. It was more probable to observe a species with one individual infected than many individuals of a species being infected. The probability distribution changes among habitats is an indication of fluctuating host range breadth. The functions integrate to one and were scaled for comparison. For example, AMV in the edge habitat has a relatively high probability of many infected individuals of a species, and similarly with CMV in crop. *No distribution; only one host species in crop.

The 11 viruses analysed had fundamental host ranges between 8 and 42 species, with a total of 47 host species for all viruses (electronic supplementary material, figure S1). We used KDE to visualize differences in realized host ranges by comparing probability distributions of the number of infections by each virus over species among the habitats (figure 3). For example, the probability density distributions of AMV, CMV, TSWV and PVY generally showed higher probabilities of single or few infections per species in each habitat. The realized host ranges of the first three viruses shifted from a high density of few infections to a relatively high density of infection in few host species between the different habitats (figure 3). Although not indicated in these graphs, the low frequency infections tended to occur across many species, while the high frequency infections occurred in a single, or very few species per habitat. The higher probability for large numbers of infections by CMV in fallow resulted from infecting Amaranthus sp. and D. erucoides. Similarly, AMV-infected M. sativa in large numbers in edge. By contrast, the probability distribution of infections by PVY across habitats was relatively constant. Together these results demonstrate that the viruses with wide fundamental host ranges infect comparatively large numbers of individuals only in one or few species associated with a given habitat.

(b). Virus prevalence models

We looked at model fit comparing abundance, richness, Shannon's (H) and Tsallis diversity (S_q) predictors of virus prevalence (see Material and methods) in separate analyses without and with fixed factors for habitat and host range. All GLMs with habitat-sensitive predictor variables showed very high residual deviance test (RDT) values close to 1.0 (electronic supplementary material, table S3) indicating good model fit. Abundance and Tsallis S_q had consistently better explanatory power across model designs (electronic supplementary material, table S3). The addition of host range effects in the abundance and S_q models substantially improved the Akaike information criteria (AIC) scores especially for those including more predictor variables across the habitats. In particular, the abundance and S_q models that included a fundamental host range effect resulted in higher significance for nearly all effects and the best-fit RDT scores. Tsallis S_q was chosen as the most appropriate estimate of diversity as it produced better model fit values compared with Shannon's H and was used in all subsequent linear modelling.

At the ecosystem-level scale, S_q had a positive effect on virus prevalence (table 1). However, opposite sign effects among habitats were evident in the habitat-sensitive models when a fundamental host range predictor was included (table 2): negative in edge and wasteland (Wald p = 0.028; Wald p = 0.010, respectively) and positive in fallow and crop (Wald p = 0.006; Wald p = 0.008, respectively). Inclusion of realized host range predictors in the habitat-level models resulted in a significant negative S_q effect (Wald p = 0.030) on virus prevalence, which was due only to edge (electronic supplementary material, table S4). Together, the models show that the relationship between virus prevalence and diversity is sensitive to spatial scale and dependent on differences among the habitats. Estimating ecological variables at the ecosystem scale all resulted in positive prevalence–diversity relationships. At the habitat scale, there were significant opposite sign effects indicating context-dependency.

Table 1.

Generalized linear models of ecosystem-level relationships between virus infection prevalence responses (n = 11) with fixed effects for Tsallis (S_q) diversity with type II Wald χ²-tests. (Fundamental host range effects were compared to models without it. A residual deviance test (RDT) = 1 indicates the model fit, given the data, is highly likely. Values in bold indicate significance.)

	estimate	s.e.	χ2	d.f.	p > χ2	AIC	RDT
TSALLIS Sq
(intercept)	−4.330	0.190	—	—	—	−66.24	0.4060
total Sq	1.302	0.200	50.434	1	0.0000
(intercept)	−4.201	0.242	—	—	—	−65.30	0.4141
total Sq	0.557	0.846	0.441	1	0.5065
host range	0.624	0.688	0.824	1	0.3639

Table 2.

Generalized linear models of habitat-level relationships between virus infection prevalence responses (n = 11) with fixed effects for Tsallis (S_q) diversity with type II Wald χ²-tests. (Fundamental host range effects were compared to models without it. A residual deviance test (RDT) = 1 indicates the model fit, given the data are highly likely. Values in bold indicate significance.)

	estimate	s.e.	χ2	d.f.	p > χ2	AIC	RDT
TSALLIS Sq
(intercept)	−4.455	0.255	—	—	—	−62.44	0.9984
E Sq	0.308	0.630	0.246	1	0.6196
F Sq	1.429	1.009	1.767	1	0.1838
W Sq	−0.768	1.338	0.317	1	0.5734
C Sq	0.470	0.649	0.499	1	0.4798
(intercept)	−3.623	0.361	—	—	—	−70.63	0.9993
E Sq	−2.165	1.009	4.803	1	0.0284
F Sq	2.094	0.741	7.587	1	0.0059
W Sq	−3.564	1.424	6.654	1	0.0099
C Sq	1.677	0.651	7.072	1	0.0078
host range	2.612	0.971	7.679	1	0.0056

Models formatted with host prevalence responses indicated no significant effect of Tsallis S_q regardless of the scale of the relationship. In all cases, the contribution to total variance by random factors for year, season and virus was negligible compared with the residual error (electronic supplementary material, table S5). The addition of a fixed factor for habitat to one for S_q, improved the model fit with a significant (Wald p = 0.038) habitat effect (electronic supplementary material, table S6).

In summary, both host and virus prevalence models (i.e. long and wide format) were consistent in that habitat heterogeneity contributed significantly to prevalence. Also, diversity effects on prevalence depended on the spatial scale at which diversity was estimated.

4. Discussion

This is the first field study that we are aware of that (i) uses a suite of generalist plant–virus species, to (ii) provide empirical evidence of spatial scale and host range phenotype dependencies in predicting infection prevalence in a multi-host–multi-pathogen species system. Our standout finding was that spatial scales affect the relationship between biodiversity and infection prevalence. When habitat heterogeneity was explicit in the model structure, prevalence became conditional on diversity and the prevalence–diversity relationship was habitat specific, showing both amplification and dilution effects at the smaller spatial scale of the habitat. When habitat was not considered as a factor in the model structure, prevalence was conditional on diversity at the larger spatial scale of the ecosystem, but only an amplification effect was detected. The inclusion of a factor for each virus host range also altered the statistical significance, predictive power and generality of the prevalence–diversity models. Our findings agreed with other studies which showed that pathogen distributions in heterogeneous environments responded to host composition in a context-specific manner [9,36,57] and to species-specific host range variation [12,58]. Spatial-scale dependencies and complexity in biotic interactions represent large caveats in the generalization of disease–diversity relationship. The realized host ranges of each virus were consistent with facultative generalism [59] and highly aggregated resource use in heterogeneous environments [19].

(a). Scale dependencies and prevalence–diversity relationships

The effect of spatial scale on the disease–diversity relationship was recognized by Ostfeld & Keesing [2] and was critical to the interpretation of biotic interactions in our study system. At the ecosystem scale (table 1), GLMs indicated significant support (Wald p < 0.001) for the positive relationship between virus prevalence and diversity (S_q), an amplification effect. By contrast, no significant effects of diversity occurred at the smaller spatial scale of the habitat. However, significant prevalence–diversity relationships were found at the habitat scale when a variable for fundamental host range was added to the model, indicating that virus phenotypes change the prevalence–diversity relationship. With the addition of a fixed factor for host range, a dilution effect occurred in both wasteland and edge (table 2; electronic supplementary material, table S3). The effect of realized host range also altered the prevalence–diversity relationship, with edge producing the only significant dilution effect (electronic supplementary material, table S4). The differences in realized and fundamental host range that we observed indicate phenotypic variation in host resource use by plant viruses across heterogeneous environments. The broad host ranges of generalists affected prevalence–diversity patterns because the numbers of infections were less dependent on species richness across spatial scales, compared to what might be expected of viruses that specialize on very few species. In other words, the realized host ranges of the generalist plant viruses show that hosts were infected at low or high abundance, and host species identity was less important compared with specialists.

The distinction between the wider fundamental and narrower realized host ranges of the generalist plant viruses responding to habitat heterogeneity in this study were consistent with plastic responses to community host composition [43–45]. The Tsallis S_q index provided an indication of how much infection was attributed to the presence of either rare or dominant species in a given habitat at the smaller spatial scale. Keesing et al. [3] envisaged the importance of species evenness in disease–biodiversity relationships. The value of the evenness parameter of the Tsallis S_q index, q, indicated that habitats included dominant and a large proportion of rare host species. Most relationships between a virus and its host range were characterized by q < 1, indicating rare host species contributed strongly to infection prevalence responses (electronic supplementary material, table S7). Low abundance host species were also susceptible to large numbers of infections. For example, viruses with the widest fundamental host ranges (e.g. AMV, BWYV and CMV) showed high infection frequencies in relatively low abundant species such as C. arvense in fallow and D. erucoides in crop habitats (figure 2). This suggests infections were not consistent with frequency or density-dependent models (e.g. [60]). The observation of both amplification and dilution effects across the habitat scale in our study were consistent with the negative variability–area relationship hypothesis [16,18]. That is, at small scales the variability in habitat diversity was large, therefore, prevalence–diversity relationships were also variable. Thus, the abundance of a species at a given spatial scale might not be an indication of its role as a competent host in the pathogen system when ecological organization permits cross-scale transmission. In other words, variation in the relative abundances of any host species between habitats, or aggregation effects, might have caused species to appear rare at the habitat scale, even when their relative abundance across the ecosystem was the same. Conversely, some host species with low relative abundance at the habitat scale might also be rare at the ecosystem scale. Host species that are rare at larger spatial scales might act as hosts when generalist viruses can infect them in the absence of other hosts within their range, or because of heterogeneity in transmission among host species [58] and fine-scale variation in vector ecology [22]. Rare host species have been acknowledged as important hosts of spillover dynamics when the ecological structure of surrounding flora and close phylogenetic relatedness are strong predictors of prevalence [8]. A hosts' competence to transmit might be dependent on within-host mechanisms related to virus and vector multiplication [8,61] and feeding preferences by vectors [10], but these effects are embedded in our observations of prevalence. Fine-scale heterogeneity in the agro-environment might favour the evolution of phenotypic plasticity capable of responding to relatively wide host trait variation [25]. These results suggest that a decoupling of the prevalence–diversity relationships occurred at the habitat scale as a result of the generalist plant viruses’ ability to infect, and transmit among a broad range of hosts.

(b). Environmental heterogeneity and generalist host use

We expected plant viruses with broad host ranges to encounter only a fraction of the species in their fundamental host range at smaller spatial scales, but still be able to infect them. Evidence of host range plasticity among habitats shown in this study is analogous to a definition for facultative generalist diets in mammals [59]. Under this definition, facultative generalists have the widest host range breadths, which can be very narrow on any particular resource without a relatively high fitness cost; i.e. the resource that provides fewer obstacles to be used will be favoured. Weak across-host fitness trade-offs and shifts to distantly related hosts are common [28], and suggest ecological differences between habitats and broad host ranges were driving prevalence patterns in our study (i.e. dilution or amplification effects were not generalizable). Facultative generalism implies that the mean fitness variance across heterogeneous environments need not necessarily be lower for generalist compared with specialist viruses, because the cost of adaptation is offset by phenotypic plasticity. This is in contrast to obligate generalists that theoretically use resources that are more costly to assimilate on average [59]. Heterogeneous environments theoretically promote relatively narrow resource use, but these models usually assume the same genes control fitness relationships across different environments [25]. Highly competent host populations, species or communities in one habitat might actually be poor hosts in another [44]. Heterogeneity in resources available to hosts that affect competence might be an important influence in the evolution of virulence and resistance mechanisms in ecosystems [43,62]. Furthermore, changes in resource availability to hosts might also influence vector behaviour and pathogen transmission.

The movement ecology of pathogens has been tightly associated with the assumption that gene flow will overwhelm genetic divergence and local adaptation at fine spatial scales [35]. The realized host ranges of generalist viruses such as AMV and CMV observed in this study suggest virus gene flow might permit localized population divergence at fine spatial scales of 25 m. Local community effects associated with narrow resource use by pathogens are consistent with population divergence at local scales [33]. Resource heterogeneity and selection at fine scales has been demonstrated in micro-geographical [35] patterns of host resistance [19]. Genetic interactions between hosts and pathogens will fundamentally depend on encounter rates between genotypes [20,21]. There are several mechanisms by which micro-geographical population divergence can theoretically arise [35]. These include habitat (or host) preferences that might pertain to vector behaviour, genetic drift or mutation, but must ultimately rely on host susceptibility and competence. Temporal changes in a given space also have important consequences for disease severity in populations (e.g. [63]) and species [6], but we focus on spatial dependencies as this aspect is relatively neglected in the epidemiology literature. Generally, the biological complexity associated with a particular environment will determine the magnitude of trait interactions that are possible.

A number of empirical considerations arise when biologically complex multi-species host–pathogen interactions are the focus of an investigation. The zoonotic literature has provided great insight into multi-host–pathogen disease dynamics [2,41,64,65]. Typically, studies compare the diversity of multiple hosts with a single virus [6,7,14,40], between a single virus and a vector species [13], or among several viruses and a single host [5] by estimating prevalence in terms of the host or the pathogen, but not both. Only when many host and pathogen species are jointly analysed in the system does it become necessary to consider both perspectives, as done here, and only one of these may suit the hypothesis being tested. For example, when multiple species of both host and parasite have been considered [58], estimation of prevalence in each host was used to test whether differences in the contribution to transmission existed among host species (an explicit test of host heterogeneity effects). Furthermore, when phenotypes are considered, the model must be tractable to testing the relationship between the trait under consideration (e.g. virus host range) and other variables of interest. Implementing preventative measures requires the identification of key predictive variables that will also depend on the system being investigated.

In conclusion, by using species-specific traits and spatially explicit variables we found predictive models for disease based on diversity were improved, but also emphasized the context-specific nature of disease–diversity relationships. In this study, generalist viruses exhibited host associations in particular habitats that resembled specialization. The presence of these narrow associations disagrees with fitness costs expected of generalists by theory. The lack of fitness trade-offs among hosts typical of facultative generalists represents a stochastic factor in disease risk prediction. The spatial scale used to measure components of a system is critical to identifying either adaptive or non-adaptive mechanisms in disease emergence. To develop a mechanistic understanding of a system, it is critical to identify the spatial scales where host–pathogen interactions are reproducible, or stable in time, so to separate stochastic noise from deterministic cause. Future work should consider meta-community approaches to inform models of scale dependencies, and community ecology approaches to compare host–pathogen species convergence–divergence patterns among ecosystems.

Data accessibility

The data are available from the Dryad Digital Repository (http://dx.doi.org/10.5061/dryad.253hr [66]).

Authors' contributions

F.G.-A., A.F., S.S. collected the data and developed the conceptual framework of the project. M.M.L. developed the methodology, statistical analyses and wrote the manuscript and all authors contributed to the final version.

Competing interests

The authors are aware of no competing interests.

Funding

This work was in part supported by grant BFU2015-60418-R, Plan Estatal de I+D+I, Spain, to F.G.-A.