When is biodiversity data relevant to guiding disease surveillance?

The diversity of host communities can be a central determinant of pathogen dynamics and prevalence [4,31,42–45]. One of the leading ecological determinants of spillover risk is the density of infected hosts in a location proximate to human populations [2]. Unfortunately, this itself is often not actionable because it relies on spatially explicit data about abundance of — and pathogen prevalence in — all relevant hosts, which is rarely (if ever) available. Therefore, estimates of spillover risk are limited both by lack of knowledge on disease dynamics and the distribution and abundance of host species.

We address this data limitation by using weighted estimates of host composition to prioritize sampling locations. Further, this determines which type of sampling (species occurrence vs. pathogens prevalence) is most informative at a given location. While predicted presence of host species does not fully capture the risk of disease transmission, it hints at the potential for pathogen presence. This is especially the case when different species can be weighted by their relative importance as reservoir hosts, i.e., accounting for interspecific variation in host competence, population density, and probable role in pathogen maintenance.

By contrast with a situation where all species are assumed to be equally important, prior knowledge about the relevance of different hosts for pathogen transmission can be incorporated in these models as a relative weight of host sampling priority. This data can come from phylogenetic similarity to known hosts (potentially incorporating phylogenetic distance to humans, a known predictor of zoonotic potential [41]), existing records of infection, or expert knowledge on the system in question [46]. In the extreme case where there is no prior information to establish weights for host species, this approach lends itself to an adaptation of the well-established practice of estimating local host diversity by stacking predicted species distributions [47], which is particularly appropriate when using statistical predictions of species presence [48]. Although stacking species distribution models (SDMs) is often biased towards higher estimates of local species richness [49], in our context this bias is unobjectionable as it leads towards more cautious recommendations that in the worst case overestimate the ranges of reservoirs, and the stacking of species-specific SDMs has been shown to produce better estimates of richness compared to joint (multi-species) SDMs [50]. For a full set of guidelines for interpreting the outputs of SDMs, see Box 1.

What are host distribution models telling us about pathogens?

The circulation of pathogens (such as viruses or obligate intracellular bacterial infections) through an animal population can vary significantly over time and is influenced by host species behavior, immunology, and environmental factors, among other drivers. Over long time periods, pathogens are thought to circulate stably within their primary host populations throughout a landscape according to the contacts of individual members of a population. These pathogens can also circulate through other susceptible animal populations if the two competent animal populations overlap in the landscape, e.g. “bridge hosts” [65] can enable the persistence of pathogens in “maintenance hosts.” For example, possums (Trichosurus vulpecula) are the primary host of bovine tuberculosis (Mycobacterium bovis), but infection in red deer and ferrets and subsequent spillback enables pathogen persistence at larger spatial scales [66].

In places where no susceptible or potentially susceptible hosts are present, it is not possible to find infected individuals. Thus, the extent of a pathogen can be thought of as the aggregation of ranges for each host species, and a rough estimate of pathogen habitat suitability across space can be obtained by weighting host susceptibility to infection [67,68]. The ensemble distribution then represents the landscape that the pathogen faces, which is not directly formed by the physical environment, but is instead determined by the distribution of hosts in a landscape.

Biodiversity dose, host weighting, and uncertainty

Throughout this manuscript, we refer to the estimate of local host composition that reflects both host richness and a priori estimates of host importance as the “biodiversity dose” — the ecological capacity for a location to maintain the pathogen. This is related to the existing concept of community competence, but distinct as the biodiversity dose is calculated as a weighted average of SDM habitat suitability scores (we provide a glossary of terms and their definitions in S3 Table). Dose therefore captures both the likely presence of host reservoirs and their relative importance in disease transmission. Biodiversity dose is defined for a group of hosts, ideally encompassing the full set of hosts for a given pathogen, but also accounting for the subset of hosts of interest for a given monitoring program.

This still leaves the task of assigning species weights. If no information is available about species importance as a host for a pathogen, a naive assumption of equal weights can be applied. If the capacity of a species to serve as a reservoir for a given pathogen is predicted with a model [39], the relative confidence in the prediction can be used as a weight.

The method used to identify infection can also be used as a basis for species weighting: for example, PCR and serology provide qualitatively different information about the host-pathogen relationship. PCR can only identify an ongoing infection during viral shedding, whereas serology can determine if a pathogen was ever present in a host. However, positive serological tests only indicate that a host was exposed to the pathogen, and does not definitively indicate the host can or did transmit the pathogen. The form of collected data should be considered, along with these factors, when assigning weights. When there is uncertainty about the capacity of a host to be infected by a pathogen, methods that assign weights based on phylogenetic proximity to known hosts [69] can also provide usable information, particularly as distance to humans is a strong predictor of zoonotic potential [70].

At some level, weights will reflect expert understanding of host-pathogen relationships and judgement of risk. While explicit recommendations of formulae to construct weights from phylogenies, predicted host-pathogen association networks, and serological evidence remain outside the scope of this paper, future work could apply comparative methods to test the efficacy of different weight-construction methods on simulated data under different levels of data availability to assess the robustness of weight construction in practice.

Assessing the robustness of sampling priority to both different weight selection and host inclusion is essential to build confidence that this protocol provides the most informative data possible. We discuss the use of sensitivity analyses to better understand the consequences of these choices on the recommended configuration of the monitoring network in the final subsection of the protocol. If there is considerable epistemic uncertainty regarding assigned host weights, or inclusion of specific hosts, priority maps from multiple candidate sets of weights can be created, and the regions where sampling priority is high under all candidate priority maps can be used as the region to target sampling.

Another crucial element for guiding sampling is the uncertainty about the distribution of the biodiversity dose, i.e., the weighted average of the uncertainty associated with the SDM for each host. There are several different ways to derive uncertainty from an SDM, and they tend to reflect different aspects of model fit — see Box 2 for more information on the various forms of SDM uncertainty that can be used. High uncertainty points to an increased priority for sampling host occurrence to better refine the prediction of host presence and overall biodiversity dose.

The information about biodiversity dose, dose uncertainty, and prevalence data (if available) can be used to make a map of sampling priority to decide where and what to sample — see Fig 1. It is critical to emphasize that while the diagram in Fig 1 is useful for planning the type of data to sample when prioritizing data collection, it cannot be used to infer risk or guide outbreak response. Instead, this protocol helps identify what type of sampling is most appropriate at a given stage of surveillance, based on host biodiversity distributions and assumptions about host importance. When the sampling objective is population surveillance or estimating pathogen prevalence within hosts, areas of high biodiversity dose (Fig 1B, Quadrants A & B) should be prioritized. However, when the goal is to improve host occurrence data, sampling should focus on areas of high dose uncertainty (Fig 1B, Quadrants B & D).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. A protocol for biodiversity-informed wildlife disease surveillance.

(A): The conceptual framework for optimizing sampling priority using host species distribution models, and pathogen prevalence data if available (left). (B): The type of sampling that is most informative (pathogen prevalence vs. host occurrence), depending on a location’s level of biodiversity dose (weighted host habitat suitability) and uncertainty about that dose.

https://doi.org/10.1371/journal.pesy.0000023.g001

Based on the goal of sampling, we can combine dose and uncertainty predictions into an overall priority score at each location x as

where is the SDM score of species i at location x, is the uncertainty of this score, is the weight for the biodiversity dose, is the weight for the uncertainty such that , and is the weight for species i, such that .

Integration of prevalence data into monitoring network design

When available, data on pathogen prevalence can refine sampling priorities, but the best way to use this information for sampling is contingent on the spatial and temporal scale of the prevalence data and the goals of sampling. For example, continued surveillance of populations known to have high prevalence may be useful if the goal is to determine if this population serves as a persistent reservoir for a given pathogen, or if it was undergoing a transient outbreak when the existing data was collected.

There are numerous factors that impact how prevalence data can be used to inform spatial sampling prioritization. Pathogen prevalence data is often reported at coarse spatial scales associated with regional administrative boundaries. Alternatively, if existing prevalence data is precise with high enough spatial coverage that it enables spatially explicit estimates of prevalence, sampling can most effectively improve spatial predictions of prevalence by targeting regions where the degree of prevalence is uncertain, but for which we are confident there is a high biodiversity dose (as we do in the second case study). This is useful for locating existing reservoirs across space to better map zoonotic hazard and spillover risk. Another consideration is the sample size from which existing prevalence estimates are derived. For example, prevalence estimates derived from 10 total tested individuals are far more uncertain than 500 total individuals, and this uncertainty can be used to inform sampling priority. For a full overview of potential methods to use depending on the scale of existing prevalence data see S1 Table, and for guidelines on how to target different forms of sampling regions based on their predicted dose, dose uncertainty, and predicted prevalence, see S2 Table.

Another important consideration is the variability in what existing prevalence data is saying. For example, serological diagnostic tests indicate exposure to a pathogen at any point in the past, and do not directly indicate the individual’s ability to transmit that pathogen, whereas PCR can detect both active infection and viral load as a measure of transmissibility. Together, these can provide complementary forms of information about both historical infection dynamics and active prevalence. Fully representative sampling of wildlife populations is rarely feasible, so estimates of prevalence are obtained from a small subsample of the entire population of unknown size. Both of these factors induce error in the resulting prevalence estimates; however, there are numerous methods from population ecology for estimating abundance with imperfect detection (e.g., with N-Mixture Models, [71]) and to infer population trends over time [72]. These can naturally be extended to the context of imperfect detection of both host abundance and pathogens within hosts in infectious disease modeling [73].

Sampling point selection

Once we have a sampling priority map, we want to use it to guide sampling site selection. There are many options for point selection algorithms rooted in sampling theory, which itself has a well-developed theory of spatial sampling. A full review of spatial sampling theory is outside of the scope of this paper, but see [74] and [25], and for a comprehensive review of the general theory of sampling, see [75].

A crucial point here is that because we are interested in targeting areas of high priority, we are constrained to a set of point-selection methods for unequal probability sampling, meaning each location in space can have a distinct probability of being included in the sample. A naive approach would be to draw samples with probability directly proportional to priority value, but due to the autocorrelation associated with biodiversity dose and uncertainty, this would result in many sampling points clumped close together, which would likely provide redundant information, and be an inefficient use of sampling effort. This is a well-established issue in spatial sampling, and as a result there are many algorithms for selecting spatially balanced samples — for example: Generalized Random Tessellation Sampling (GRTS; [76]), the Pivotal Method [77], and Balanced-Acceptance Sampling (BAS; [78]). These methods are all included in the BiodiversityObservationNetworks.jl package in Julia, which we utilize in the case studies.

Recent results suggest that most site selection algorithms achieve equivalent performance for the same problem [79], which allows users to pick an algorithm based on specific features, such as the support for auxiliary environmental data. For our purposes, we find BAS the most flexible method because it is both the most computationally efficient and achieves better spatial balance than the alternatives [78].

In practice, algorithms for generating spatially balanced samples targeted toward regions of high priority may not be skewed “enough” toward high priority sites for a given use case, so we also tend to tilt the distribution of priority scores to make the values of high priority even higher. This is done using an exponential transform, , where is the original priority at a given location x, and the parameter controls how much to tilt the adjusted map toward areas of high priority.

Another common method for designing samples is stratification. In the context of spatial sampling, stratification consists of splitting the spatial domain into discrete regions (corresponding with specific strata), each of which contains a user-determined number of samples. In our case, stratification can be used to divide space into the different data collection regimes (sampling for host presences, prevalence or both; Fig 1B) and distribute sampling effort across these strata based on the goals of the monitoring program, with spatially balanced point selection methods being used within each stratum. For further guidelines for interpretation and practical considerations when using selected sampling points, see Box 3.

Sensitivity/robustness analysis

This protocol relies on several choices of weights based on the goals of sampling and a priori estimates of host relevance for the pathogen(s) of interest. Assessing the sensitivity of the priority map to weight choices is important to ensure the priority map is robust to small tweaks to the weights that do not reflect meaningful biological differences in host relevance.

For a given set of weights, , a natural way to assess the sensitivity of the overall priority is to examine the mean absolute change in the priority map, P, when a small random perturbation is applied to the weights. This gives us an interpretable metric of how sensitive the resulting priority map is at a chosen weight value. We create a “nudged” version of the weights, , by adding noise , which is a vector of i.i.d. of the same length as , and then renormalized (so also sums to 1). We then construct the priority map, P, for the original weights, and the priority map from the “nudged” weights. The overall sensitivity, S, of the priority map at a given weight value can be computed as the sum of the absolute difference across all locations x:

If the selected weights for a given use case are close to high sensitivity regions in the space of possible weights, end users should consider the additional caveat that the selected weights are very sensitive to small changes, and ensure the selected weight values are reflective of meaningful biological differences in host relevance. We provide an example of this analysis in S1 Fig, assessing the sensitivity of weights for the first case study.

Creating dose and uncertainty maps

For both case studies, we use the same SDM methodology outlined here. We emphasize that the specific approach to building SDMs is not the focus — sampling prioritization is only going to be as good as the SDM, and we present guidelines for the interpretation of SDM outputs in Box 1.

We obtain occurrence records for each host species from GBIF (https://doi.org/10.15468/dl.b3z33h for India, https://doi.org/10.15468/dl.wzcxp9 for South Korea). As predictor variables, we use the 19 bioclimatic variables from CHELSA [85] at 1 km x 1 km resolution for South Korea, and from WorldClim [86] at 2.5 arcminute resolution (approximately 5 km x 5 km at the equator) for India (to keep the raster size reasonable). We generate pseudoabsences [87] using background thickening [88], with a buffer around each occurrence record where no pseudoabsences can go (25 km for India, 8 km for South Korea).

Models were trained using SpeciesDistributionToolkit.jl [89] and EvoTrees.jl [90] in Julia. Specifically, we use a Boosted Regression Tree (BRT; [91]) with a Gaussian loss metric, meaning the value of each node in the tree is fit to a Gaussian using maximum-likelihood estimation. Therefore, the BRT provides both a suitability prediction score, and an uncertainty value associated with it, for each pixel.

In order to standardize the dose values for each species, we compute the empirical cumulative distribution function (ECDF) on raw output predictions, ensuring all prediction layers are on the same scale.

Case Study 1: Designing a network without prevalence data

Our first case study considers the situation where we are primarily limited by a lack of prevalence data. This case study was inspired by [92], published 9 years after Lassa virus (Mammarenavirus lassaense) was first identified in Nigeria. In the immediate aftermath of the discovery of a virus of severe public health importance, [92] were the (self-identified) first to sample for Lassa virus outside of Africa. Lassa virus circulates in hosts with widespread distributions, and at the time, Lassa virus was difficult to distinguish in clinical settings from other infectious diseases present in India (e.g., typhoid fever, malaria). Therefore it was not unreasonable to suspect it may be present in rodent and shrew populations in India.

Today we know that Lassa virus is confined to Western Africa, but this case study serves as an example for how we use biodiversity data to prioritize sampling in the context of a recently identified virus without existing reliable information about its geographic extent, prevalence, or known reservoir species. We construct a priority map for sampling using the host species tested by [92]: Funambulus tristriatus, Hystrix indica, Rattus rattus, and Suncus murinus. Note that in a modern context, other forms of data could be used to supplement the list of potential hosts that were not available to [92], e.g., phylogenies constructed with modern sequencing methods, and crowdsourced occurrence data from GBIF of other rodents and shrews in the region. For the sake of example, we choose to stick with the hosts considered in the original study.

To generate a priority map for sampling, we follow the SDM procedure outlined above, and for the sake of example assign species weights at random (resulting in 0.11 for Suncus murinus, 0.34 for Rattus rattus, 0.2 for Hystrix indica, and 0.35 for Funambulus tristriatus).

Fig 2A shows a bivariate map of both the dose and uncertainty aggregated across host species. The map of India highlights three large-scale regions of interest: the south, where the model is confident that many host species are present, with increasing uncertainty moving northward; the northwest, where the model is uncertain about many or all species; and the northeast, where the model is also relatively confident that host species are present. We say a region in the top 50% of uncertainty for a given species is for “discovery sampling”, and a region in the top 50% of biodiversity dose is for “prevalence sampling”. In Fig 2B we see the total number of species that fall into each of these categories across space. Fig 2C shows our priority map, with both and equal to 0.5 (reflecting equal priority toward discovery and surveillance). The white points reflect sampling locations generated using Balanced Acceptance Sampling applied to the tilted priority map with (see Sampling Point Selection subsection above). This tilting value was chosen arbitrarily to result in a set of sampling locations concentrated at high priority values. As the tilting value increases, the median priority score at selected sites increases, and the variance of the priority score at selected sites decreases (S2 Fig). In practice, the tilting value should be adjusted to reflect confidence and uncertainty in the decisions of host inclusion and weighting made during the prioritization process — if there is little uncertainty associated with the hosts and weighting, high tilting values will ensure all sites are very high priority, where as more uncertainty may warrant a diversity of mid-to-high priority values, which will be selected by lower tilting values.. The teal markers reflect the four historical sampling locations from [92]. Finally, in Fig 2D, we see the host species that contributed the most to the priority score across space (computed as the maximum value of at location x across species i).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. The case study for India.

(A) Biodiversity dose + uncertainty bivariate plot. (B) Number of hosts in prevalence-regime (top 50% of habitat suitability for that species) vs. discovery (top 50% of uncertainty for that species), (C) Sampling priority map and BON sites. X’s are historical sampling locations from [92]. (D) The largest host contributor to priority. Administrative boundaries come from GADM.

https://doi.org/10.1371/journal.pesy.0000023.g002

Case Study 2: Accounting for prevalence data in network design

Our second case study revolves around incorporating prevalence data into sampling prioritization. In contrast to the prior example where very little is known about the prevalence of a given pathogen, we now demonstrate the protocol for the surveillance of two hantaviruses in South Korea (Orthohantavirus hantanense & Orthohantavirus seoulense), which are two of the main causative agents of hemorrhagic fever with renal syndrome (HFRS; [93]), and the recently characterized Imjin virus (Thottimvirus imjinense, [94]), which, although the pathogenicity in humans is unknown, has been shown to replicate in human macrophages and endothelial cells [95]. This is reflective of biosurveillance targeting multiple objectives: both known zoonoses and a recently discovered virus with limited knowledge about the potential to infect humans. Typically, the hosts of these viruses are rodents and shrews, and in our study we focus on three small-mammal hosts: Rattus norvegicus, Apodemus agrarius, and Crocidura lasiura. These host-virus pairs were selected as the database contained prevalence records across the country at a broad enough spatial scale to enable spatially explicit prediction of prevalence.

The data comes from Project ArHa, a database of Arena- and Hantaviridae host-virus associations [84]; specifically the following studies from South Korea with explicit prevalence estimates: [94,96–107].

Prevalence information is incorporated into the prioritization by spatially explicitly estimating prevalence across the extent using Gaussian Process regression (also known as kriging) for each host-virus pair with at least 5 populations sampled across space. This is done using a Squared-Exponential covariance kernel with a fixed length scale parameter using GaussianProcesses.jl [108], and the hyperparameters for the model are optimized using the L-BFGS algorithm in Optim.jl [109]. This provides both spatial estimates of prevalence, but crucially, spatially explicit uncertainty maps associated with this prevalence estimate.

Fig 3 shows both the bivariate plot of biodiversity dose vs. dose uncertainty (Fig 3A) and dose vs. prevalence uncertainty (Fig 3B). This identifies a clear geographic trend in dose and uncertainty. Dose is highly concentrated in the southeast of the Korean peninsula, with dose uncertainty (uncertainty regarding the presence of hosts) concentrated in the northwest, near Seoul and many other large cities. In contrast, prevalence uncertainty is primarily concentrated in the region of highest dose, the southeast. This indicates that southeastern Korea is a priority region for for prevalence sampling, as it contains many hosts with high confidence, but lacks existing data on the prevalence of these pathogens.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. The case study for South Korea.

(A): Bivariate plot of dose (increasing green) and dose uncertainty (increasing blue). (B): Bivariate plot of dose (increasing green) and prevalence uncertainty (increasing orange; measured as the sum of estimated variance from kriging across each host-virus pair). (C): The spatial strata for which occurrence, prevalence, and both forms of sampling should occur (see main text for how these are computed). (D): The priority scores for regions where discovery sampling is a priority (increasingly blue) and prevalence sampling is the priority (increasingly purple) in their respective strata. (E): The total priority map with sites selected using Balanced Acceptance Sampling. The marker shape of each site corresponds to the type of sampling that should occur there, based on the stratum in (C) that each point falls within. Administrative boundaries come from GADM.

https://doi.org/10.1371/journal.pesy.0000023.g003

To better isolate regions where we are confident hosts are present for prevalence sampling, and regions of high host presence uncertainty for discovery sampling, we stratify the spatial extent into discrete regions for prevalence sampling and occurrence sampling as follows: the region that is both in the top 50% of biodiversity dose and top 50% of prevalence uncertainty is the stratum for prevalence sampling, and the region in the top 50% of dose uncertainty is the stratum for discovery sampling (Fig 3C). Spatial regions that fall into both categories are then targets for both forms of sampling. The choice of 50% quantiles is arbitrary for this example. The sampling priority within each of these strata can be seen in Fig 3D. This allows explicit quantification of the regions of highest priority for each type of sampling, and highlights that more occurrence data for hosts is the highest priority in the northwest, and that prevalence sampling is likely to be most fruitful in the southeast.

The overall sampling priority map can be seen in Fig 3E, along with the points sampled using Balanced Acceptance Sampling with a tilting value of (see previous case study for discussion of tilting value selection), with the marker indicating the type of sampling (prevalence, occurrence, or both) that should take place at each location.

This case study demonstrates how we can integrate host SDMs with existing prevalence information to design sampling strategies and split geographic space into discrete regions where each type of sampling (occurrence and prevalence) should take place.

Box 3. Guidelines for the interpretation of recommended sampling locations

We want to emphasize that sampling points, though given as exact (longitude, latitude) coordinates, are not to be interpreted without critical thought. Since the grain at which the underlying SDMs are operating is often significantly larger than the scale at which one may have to decide where to place a sampling trap, selected points should be thought of as most useful as the centroids of “local” regions (where “local” corresponds to the pixel size of the SDMs used for prioritization) wherein sampling effort could best improve knowledge about the ecological settings of viral transmission. To make this point explicit, if the model suggests sampling for Mus musculus in an urban area and the coordinate itself is in the middle of the center of a highway, the model’s output should be interpreted as indicating the general area where sampling should be done (see below figure).

For example, Fig 4 shows three selected sampling sites for the second case study on South Korea, with OpenStreetMap data corresponding to the cell in the priority map that was selected. As the original climate layers have pixels that are approximately 1 km², there is considerable heterogeneity within these pixels. Experts on particular taxa should assess the scope of the local region near each sampling point, and use it to guide the practical logistics of where effort (e.g., traps) are most likely to result in useful information.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Selected sites.

Three selected sites for the second case study (colored markers, left) overlaid on a map of South Korea with the 10 most populous cities shown in dark grey dots, and the major highway network shown in light grey (to get a sense of the accessibility of these sites). Right: Land cover data, derived from OpenStreetMap, for each site. Administrative boundaries come from GADM.

https://doi.org/10.1371/journal.pesy.0000023.g004

Beyond just the spatial scale at which recommendations are made, temporal variation in both prevalence and host occurrence is a well-known driver of disease dynamics, often driven by environmental variation tied to seasonality [110,111]. Factors like migration and birth rates can affect these within-year cycles [112]. Between-year fluctuations, particularly in rodents, are both dramatic and can also counter-intuitively have an inverse effect on populations’ prevalence [113]. Here, we do not consider temporal variation in host abundances in our sampling recommendation protocol, since to do so would require orders of magnitude more data than we consider available, and would likely require the use of a different class of models/approaches.