Mapping the zoonotic niche of Ebola virus disease in Africa

1Ebola viruses have for the last forty years been responsible for a number of outbreaks of Ebola virus disease (EVD) in humans, with high case fatality rates typically around 60–70%, but potentially reaching as high as 90%. The most recent outbreak began in Guinea in December 2013 and has subsequently spread to Liberia, Sierra Leone and Nigeria. The unprecedented size and scale of this ongoing outbreak has the potential to destabilise already fragile economies and healthcare systems, and fears of international spread of a Category A Priority Pathogen have made this a massive focus for international public health. This has led to the current outbreak being declared a Public Health Emergency of International Concern on the 8 August 2014.

2The Filoviridae, of which Ebolavirus is a constituent genus, belong to the order Mononegavirales. Two other genera complete the family: Marburgvirus, itself responsible for a number of outbreaks of haemorrhagic fever across Africa and Cuevavirus, recently isolated from bats in northern Spain. Five species of Ebolavirus have been isolated to date; the earliest recognised outbreaks of EVD were reported in Zaire (now the Democratic Republic of the Congo [DRC]) and Sudan in 1976. The causative viruses were isolated and later identified to be distinct species, Zaire ebolavirus (EBOV) and Sudan ebolavirus (SUDV). A third species of Ebolavirus, Reston ebolavirus, was isolated from Cynomologus monkeys imported from the Philippines to a facility in the United States, where they experienced severe haemorrhaging. Whilst serological evidence of infection with this species has been reported in individuals in the Philippines, no pathogenicity has been reported beyond primates and porcids. In 1994 a fourth species, Tai Forest ebolavirus was isolated from a veterinarian who had autopsied a chimpanzee in Côte d’Ivoire, though the virus has not been detected subsequently. The final species, Bundibugyo ebolavirus, was responsible for an outbreak of EVD in Uganda in 2007, as well as a more recent outbreak in the DRC.

3Initial analysis suggested that the viruses isolated from the current outbreak, originating in Guinea, formed a separate clade within the five Ebolavirus species. Subsequent re-analysis of the same sequences however, indicated that these isolates instead nest within the Zaire ebolavirus lineage, and diverged from Central Africa strains approximately ten years ago.

4Which reservoir species are responsible for maintaining Ebola transmission between outbreaks is not well understood, but over the last decade significant progress has been made in narrowing down the list of likely hosts (Figure 1). Primates have long been known to harbour filoviral infections, with the first Marburg strains identified in African green monkeys in 1967. Significant mortality has also been reported in wild primate populations across Africa, most notably in gorilla (Gorilla gorilla) and chimpanzee (Pan troglodytes) populations. The high case fatality rates recorded in the great apes combined with their declining populations and limited geographical range, indicate they are likely dead-end hosts for the virus and not reservoir species. A large survey of small mammals in and around Gabon identified three species of bats which were infected with Ebola viruses—Hypsignathus monstrosus, Epomops franqueti and Myonycetris torquata. Subsequent serological surveys and evidence linking the potential source of human outbreaks to bats lend support to the hypothesis of a bat reservoir. This, coupled with repeated detection of Marburgvirus in the fruit bat Rousettus aegypticus and the only isolations of Cuevavirus also from bats (specifically Llovia virus), all support the suspicion that Chiroptera play an important role in the natural life-cycle of the filoviruses.

5Humans represent a dead-end host for the virus, with only stuttering chains of transmission reported between humans in the majority of previous outbreaks and no indication that humans can reintroduce the virus back into reservoir species. The incubation period in humans ranges from two days to three weeks, after which a variety of clinical symptoms arise, affecting multiple organs of the body. At the peak of illness, haemorrhaging shock and widespread tissue damage can occur and can eventually lead to death within 6–16 days. Human-to-human transmission is mainly through direct unprotected contact with infected individuals and cadavers, with infectious particles detected in a number of different body fluids. The typical outbreak profile is defined by an index individual that has recently come into contact with the blood of another mammal through either hunting or the butchering of animal carcasses. Whilst it has been difficult to identify the zoonotic source for the index cases of some outbreaks, a recurring theme of hunting and handling bushmeat is suspected (Table 1;;,; ). For some outbreaks, including the most recent, the initial source of zoonotic transmission has not been identified. In subsequent human-to-human transmission, the highest risk activities are those that bring humans into close contact with infected individuals. These include medical settings where insufficient infection control precautions have been taken, as well as home care and funeral preparations carried out by families or close friends. As the conditions required for transmission are culturally and contextually dependent, opportunities for sustained transmission are highly heterogeneously distributed. Typically, chains of infection do not exceed three or four sequential transmission events, although occasionally (and particularly in the early stages of infection) a single individual may be responsible for directly infecting a large number of others. In the outbreak in Gabon in 1996, a single person was responsible for infecting ten other individuals whilst in the 1995 outbreak in the DRC, thirty five cases resulted from one individual. Secondary transmission can be restricted by effective case detection and isolation measures. Where this cannot be achieved, either due to a lack of infrastructure, poor understanding of the disease, or distrust of medical practices, secondary cases can continue to occur. As the number of infections grows, the ability of healthcare systems to control the further spread diminishes and the risk of a large outbreak increases.

6The recent outbreak in Guinea and surrounding countries indicate that the previous paradigm for Ebola outbreaks is shifting. The last 40 years of EVD outbreaks were accompanied by considerable changes in demographic patterns throughout Africa. There has been a large increase in population size coupled with increasing urbanisation. African populations have also become better connected internally and internationally. Only recently have we begun to understand the dynamic nature of these travel patterns which have been clearly demonstrated to influence disease transmission over different temporal and spatial scales. Changes in land use and penetration into previously remote areas of rainforest bring humans into contact with potential new reservoirs, while changes in human mobility and connectivity will likely have profound impacts on the dispersion of Ebola cases during outbreaks. These conditions are thought to have a major role in setting the stage for the current outbreak.

7This paper aims to define the areas suitable for zoonotic transmission of Ebolavirus (i.e., those routes defined in Figure 1 excluding human-to-human transmission) through species distribution modelling techniques. The fundamental niche of a species can be conceptualised as the confluence of environmental conditions that support its presence in a particular location. Species distribution models quantitatively describe this niche based on known occurrence records of the organism and their associated environmental conditions, enabling predictions of the likely geographic distribution of the species in other regions. The era of satellites and geographical information systems has made high resolution global data on environmental conditions increasingly available. Species distribution modelling using flexible machine learning approaches have been successfully applied to map the global distributions of disease vectors and pathogens such as dengue, influenza and leishmaniasis.

8Previous studies applied the GARP (Genetic Algorithm for Rule-set Production) species distribution modelling approach to the locations of 12 Ebola outbreaks in humans between 1976 and 2002 to map the likely distribution of Ebola viruses and as a mechanism to identify potential reservoir hosts. Here we update and improve the maps of the zoonotic transmission niche of EVD by: (i) incorporating more recent outbreak data from outside the formerly predicted niche of EVD; (ii) integrating for the first time data on outbreaks in primates and the occurrence of the virus in the suspected Old World fruit bat (OWFB) reservoirs; (iii) using new satellite-derived information on bespoke environmental covariates from Africa, including new distribution maps of the OWFB; and (iv) using new increasingly flexible niche mapping techniques in the modelling framework. To elucidate the relevance of these maps for transmission, we have also calculated the population at risk of primary spillover outbreaks from the zoonotic niche of EVD in Africa, and we investigated the changing nature of the populations within this niche.

9In total, 23 outbreaks of Ebola virus were identified in humans across Africa, consisting of a hypothesised 30 independent primary infection events (Table 1; Figure 2). These outbreaks span the last 40 years from the first outbreaks in 1976 to the five outbreaks that have occurred since 2010 (Table 1). The locations of the index cases span from West Africa, with the most westerly outbreak ongoing in Guinea, to Gabon, the Republic of Congo (ROC), the DRC, South Sudan and Uganda. Before December 2013, a total of 2322 cases had occurred from Ebolavirus infections, a number already overtaken by the likely underreported current case count of the ongoing outbreak >2250 (Figure 2A). Of the four viruses circulating in Africa, Zaire ebolavirus has been responsible for the most outbreaks (13), followed by Sudan ebolavirus (7) and Bundigbuyo ebolavirus with just two outbreaks in 2007/8 and 2012. Tai Forest has caused one confirmed infection in humans, from which the patient recovered. Although outbreaks have been reported since 1976, there was an absence of reported outbreaks in humans for 15 years between 1979 and 1994 (although antibodies in humans were identified over the period) and the frequency of outbreaks has increased substantially post 2000 (Figure 2A).

10A total of 51 surveyed locations reporting infections in animals were identified in the literature since the discovery of the disease (Table 2; Figure 3). These comprised 17 infections in gorillas (Gorilla gorilla), nine infections in chimpanzees (Pan troglodytes), 18 in OWFB and 2 in duikers (Cephalophus spp.). A large proportion of the great ape cases originated from the ROC/Gabon border, coinciding with the main known distributions of both chimpanzees and gorillas and representing a period of well-documented great ape Ebola outbreaks in and around the Lossi Animal Sanctuary. All animal isolations of Ebola viruses have come from countries that have also reported index cases of human outbreaks, with the exception of several seropositive bats from a survey in southern Ghana.

11Three species of bats, Hypsignathus monstrosus, Myonycteris torquata and Epomops franqueti, were identified as the most likely candidates to be reservoir species for Ebola viruses due to high seroprevalence and the isolation of RNA closely related to Zaire ebolavirus. In total, 239 locations were identified from the Global Biodiversity Information Facility (GBIF): 67 for H. monstrosus (Figure 4A), 52 for M. torquata (Figure 4B) and 120 for E. franqueti (Figure 4C). Distribution models for all three species demonstrated predictive skill (indicated by an area under the curve (AUC) greater than 0.5) as follows: H. monstrosus AUC 0.63 ± 0.04; M. torquata AUC = 0.59 ± 0.04; E. franqueti AUC = 0.58 ± 0.03, n = 50 submodels for all three species. In addition, each species was broadly predicted within its considered expert opinion range (Figure 4A–C). The marginal effect plots (not shown) were strongly influenced by land surface temperature (LST) and vegetation (as measured by the enhanced vegetation index [EVI]). The predicted combined distribution of these species (Figure 4D), covers West and Central Africa, specifically the moist forests of the northeastern, western and central Congo basin, and Guinea, as well as the Congolian coastal forest ecoregions.

12The predicted environmental niche for zoonotic transmission of EVD is shown in Figure 5. All countries with observed index cases of EVD (n = 7, hereafter Set 1) have areas of the highest environmental suitability (see list in Table 1). In addition, areas of high environmental suitability for zoonotic transmission are predicted in a further 15 countries where, to date, index cases of the four African species of Ebolavirus have not been recorded. These are Nigeria, Cameroon, Central African Republic (CAR), Ghana, Liberia, Sierra Leone, Angola, Tanzania, Togo, Ethiopia, Mozambique, Burundi, Equatorial Guinea, Madagascar and Malawi (hereafter Set 2).

13The AUC for the Ebola model was relatively high (AUC = 0.85 ± 0.04, n = 500 submodels) indicating that the model could strongly distinguish regions of environmental suitability for EVD. Enhanced vegetation index had the greatest impact on the distribution (relative contribution [RC] of 65.3%) followed by elevation (RC = 11.7%), night-time land surface temperature (LST) (RC = 7.7%), potential evapotranspiration (PET) (RC = 5.7%) and combined bat distribution (RC = 3.8%). Marginal effect plots are presented in Figure 5—figure supplement 2.

14In total, 22.2 million people are predicted to live in areas suitable for zoonotic transmission of Ebola. The vast majority, 21.7 million (approximately 97%), live in rural areas, as opposed to urban or peri-urban areas. Of these, 15.2 million are in Set 1 and 7 million are in Set 2. In terms of ranked populations at risk, DRC, Guinea and Uganda are highest in Set 1 and Nigeria, Cameroon and CAR are top in Set 2. For a full listing of these populations living in areas of risk, see the stacked bar plot in Figure 5A.

15Over the 40 year period since discovery of EVD, the total population living in those countries predicted to be within the zoonotic niche has nearly tripled (from 230 million to 639 million) and the proportion of the population in these countries living in an urban (rather than rural) setting has changed from 25.5% to 59.2% (Figure 6).

16Data on the connectivity of human populations over this period were not available. We can infer however, intuitively, empirically and theoretically that rates of population movement within a country will scale directly in proportion to population growth.

17Records of passenger seat capacity are available since 2000 and show substantive increases over the period in Set 1 (from 2.96 to 4.77 million, a fractional change of 1.61) and Set 2 (from 5.6 to 15.6 million, a change of 2.8) (Figure 7A). More specific data on passenger volumes show almost universally similar increases since 2005 with Set 1 nations changing from 2 million to 2.5 million, a fractional change of 1.22 and Set 2 changing from 5 million to 7.9 million, a change of 1.57 (Figure 7B).

18Global analysis of airline passenger volumes demonstrates that international connectivity has increased amongst all global regions and national income strata (Figure 8). Total passenger volumes have increased by a third from 9.5 to over 14 million during the eight year window (2005–2012) where records are available. The largest increases have occurred in WHO regions outside of the sub-Saharan African region (AFRO) (Figure 8A,B). In 2012, almost half of the final destinations of those travelling from these at-risk countries were to other AFRO nations (47%). Other frequent destinations were in Europe (EURO; 27%) and the Eastern Mediterranean (EMRO; 13%). Similarly, analysis of passenger volumes by World Bank national income groupings (Figure 8C,D) show that in 2012 40% of all passenger final destinations were to low or low-middle income countries.

19We have re-evaluated the zoonotic niche for EVD in Africa. In doing so we have (i) used all existing outbreaks to assemble an inventory of index cases (n = 30); (ii) added to this all confirmed records of Ebola virus in animals (n = 51); (iii) assembled more accurate and contemporary environmental covariates including new maps of the distribution of confirmed OWFB reservoirs of the disease; and (iv) used the latest niche modelling techniques to predict the geographic distribution of potential zoonotic transmission of the disease. Using these predictions we have estimated the populations at risk of EVD both in countries which have confirmed index cases (Set 1, n = 7) and those for which we predict strong environmental suitability for outbreaks (Set 2, n = 15). In all countries at risk we show that since the discovery of EVD in 1976, urban and rural populations have increased and have become more interconnected both within and across national borders. During the last 40 years the increasing size and connectivity of these populations may have facilitated the subsequent spread of EVD outbreaks. These factors underline a change in the way in which EVD interacts with human populations.

20The remote and isolated nature of Ebola zoonotic transmission events, paired with the relatively poor diagnostics and understanding of the disease transmission routes in early outbreaks, mean that under-reporting of previous outbreaks is probable. An increasing understanding and description of a broader range of symptoms used in case definitions of EVD also increase the possibility that past outbreaks may have been misattributed to different diseases. This poor detectability of EVD also clearly limits capacity to accurately identify the locations and transmission routes of index cases. We must assume, as has been done previously, that the first reported cases are representative of the true location of the index cases. Where possible we have represented this geographic uncertainty by attributing the index case to a wide-area polygon which then incorporated this uncertainty into the mapping process (see ‘Materials and methods’).

21The relationship between the EVD niche and the environmental covariates (Figure 5—figure supplement 2), particularly the high relative contribution of the vegetation index, underscore that there are clear environmental limits to transmission of the virus from animals to humans, and that ecoregions dominated by rainforest are the primary home of such zoonotic cycles. Our analysis has shown that the zoonotic niche of the pathogen is more widespread than previously predicted or appreciated, most notably in West Africa.

22This analysis used information from all human outbreaks and animal infections to delineate the likely zoonotic niche of the disease. Further analysis, excluding the existing outbreak focussed in Guinea from the dataset used to train the model (Figure 5—figure supplement 3), still resulted in prediction of high suitability in this region, with the presumed index village located within 5 km of an at-risk pixel. This implies that the eco-epidemiological situation in Guinea is very similar to that in past outbreaks, mirroring phylogenetic similarity in the causative viruses. The ecological similarity between the past and current outbreaks also lends support to the notion that the scale of this outbreak is more heavily influenced by patterns of human-to-human transmission than any expansion of the zoonotic niche.

23It is important to appreciate that this zoonotic niche map delineates areas in which populations are at-risk of zoonotic transmission of EVD (Figure 5B). It does not predict the likelihood of EVD spillover, the likelihood of an outbreak establishing, or its subsequent rate of spread within a population. Increasing human encroachment and certain cultural practices sometimes linked with poverty, such as bushmeat hunting, result in increasing exposure of humans to animals which may harbour diseases including Ebola. Increasing human population may accelerate the degree of risk through these processes but spatially refined information on these factors is not available comprehensively. It is hoped that as the understanding of the risk factors for zoonotic transmission of Ebolavirus to humans increases, it will be possible to incorporate this information into future risk mapping assessments.

24Previous considerations of the geographic distribution of EVD have used human outbreaks alone. We have updated this work to include the last decade of outbreaks, as well as disaggregated outbreaks where evidence suggests multiple independent zoonotic transmission events overlap in space and time. Furthermore, our modelling process accommodates uncertainty in geopositioning of these index cases by utilising both point and polygon data. In addition, we include occurrence of infection in wildlife, important to the wider scale of zoonotic transmission (Figure 1), which in total has increased the dataset used in the model to 81 occurrences. The rareness of EVD outbreaks and the prevalence of detectable Ebola virus in reservoir species suggests that there will always be a limited set of observation data when compared to mapping of more prevalent zoonoses. The results demonstrate predictive skill using a stringent validation procedure, however, indicating strong model performance even with this relatively limited observation dataset.

25A broad zoonotic niche is predicted across 22 countries in Central and West Africa. Whilst several of these countries have reported index cases of EVD, others have not, although serological evidence in some regions points to possible underreporting of small-scale outbreaks. With improved ecological understanding, particularly with improvements to our knowledge of specific reservoir species and their distributions, it may be possible to delineate areas not at risk due to the absence of these species.

26Despite relatively a large population living in areas of risk and the widespread practice of bushmeat hunting in these predicted areas, Ebolavirus is rare both in suspected animal reservoirs and in terms of human outbreaks (Table 1). There is some indication however, that the frequency of Ebola outbreaks has increased since 2000, as shown in Figure 2A. We have shown that the human population living within this niche is larger, more mobile and better internationally connected than when the pathogen was first observed. As a result, when spillover events do occur, the likelihood of continued spread amongst the human population is greater, particularly in areas with poor healthcare infrastructure.

27Whilst rare in comparison to other high burden diseases prevalent in this region, such as malaria, Ebola outbreaks can have a considerable economic and political impact, and the subsequent destabilisation of basic health care provisioning in affected regions increases the toll of unrecorded morbidity and mortality of more common infectious diseases, throughout and after the epidemic period. The number of concurrent infections during the present outbreak represents a significant strain on healthcare systems that are already poorly provisioned and many other Set 1 and Set 2 countries rank amongst the lowest per capita healthcare spenders. These considerations should be paramount when international organizations debate the financing requirements for the improvement of healthcare needed in the region and the urgency with which new therapeutics and vaccines can be brought into production.

28Together, these considerations necessitate prioritisation of efforts to reinforce and improve existing surveillance and control, and encourage the development of therapeutics and vaccines. The national population at risk estimates presented here would be a strong rationale for improving, prioritising and stratifying surveillance for EVD outbreaks and diagnostic capacity in these countries. We believe it would be prudent to test OWFB species in Set 2 countries for Ebola virus, particularly during the bat breeding season to maximise chances of isolation in order to clarify the outbreak risk in these countries. In all Set 1 and Set 2 countries, raising awareness about the risk presented by reservoir bats and incidental primate hosts and the modes of transmission of this disease could be of value. Finally, increasing our capacity to rapidly map ever changing biological threats is also a core need.

29The increasing connectedness of the Africa region means that EVD is now a problem of international concern. While most EVD secondary transmission occurs locally and is likely transported via ground transit, the potential for international spread of infection is possible, as demonstrated by the importation of the disease from Liberia to Nigeria, culminating in further secondary transmission in Lagos. The aetiology of EVD infection and disease progression means that an international outbreak propagated by air travel remains unlikely, particularly in high-income countries better able to handle EVD cases. Nevertheless, a non-negligible threat remains, particularly in the low and middle income destinations and the rapid increase in global connectivity of these at-risk regions indicates that international airports could see more imported cases.

30We have focussed on reanalysing the zoonotic niche for EVD transmission and the characterisation of the populations at risk to improve the landscape in which future risk and impact of EVD outbreaks can be discussed. During the current emergency much of the work will concentrate on routes of secondary transmission in the human population—conceptually the red arrow of the H box in Figure 1. An important task is to stratify the risk of EVD spread both within and between countries and identify the most likely pathways of spread for characterisation and surveillance. Our next priority therefore is to investigate aspects of secondary human-to-human transmission by documenting the rate of geographic spread of EVD during the past and ongoing epidemics to help understand changes in these patterns in the historical record. Simulating these movements in a real landscape of population movement patterns, inferred from population movements assessed by mobile phones and other data, as well as parametric movement models is a logical next step, and can be used in future targeting of interventions and potential new treatments for both the current and future outbreaks.

31As previously discussed, whilst there is the risk of human travel during the latent phase of infection, and therefore potential for international spread, the high pathogenicity during infectiousness (immobilising infected persons) and the likely rapid and effective isolation measures implemented in regions with strong health care systems, limit the pandemic potential of EVD. Nevertheless, improvement of international containment plans and informed discussions of potential risks to airline carriers and populations of other regions will be supported by knowledge of local, regional and international population flows. Assessing these flows by air traffic volumes is an ongoing priority.

32There are several other zoonotic viral haemorrhagic fevers (for example Marburgvirus, Lassa fever, hantaviral infections and arenaviruses) that are of similar public health and biosecurity concern, due to their high virulence and mortality and their potential to cause outbreaks and spread internationally. Despite this their geographical distributions are poorly understood. Many of the methods applied here can be adapted to these diseases and improve our geographical understanding of the risk presented by these pathogens.

33We are in the midst of a public health emergency that will likely last for many more months and which has brought EVD to global attention. We emphasise that the maps of zoonotic transmission presented here do not enable assessment of secondary transmission rates in human populations, but they do act as an evidenced-based indicator of locations with potential for future zoonotic transmission and thus outbreaks. Interestingly, early reports of another independent zoonotic outbreak in the DRC are in predicted at-risk areas. An improved understanding of the spatial extent of the zoonotic niche can only help future efforts in biosurveillance.

34A boosted regression tree (BRT) modelling framework was used to generate predictive risk maps of the zoonotic Ebola virus niche in Africa. This methodology combines regression trees, where trees are built according to optimal decision rules based on how binary decisions best accommodate a given dataset, and boosting, which selects the tree that minimises the loss function. In doing so, a parameter space is defined which captures the greatest amount of variation present in the dataset. In order to train the model, four component datasets were compiled: (i) a comprehensive dataset of the reported locations of Ebola virus transmission from a zoonotic reservoir to a human; (ii) a dataset of the locations of Ebola virus infections in suspected reservoir and (non-human) susceptible host species (iii) a suite of ecologically relevant environmental covariates for Africa, including predicted distribution maps of suspected reservoir bat species and (iv) background (or pseudo-absence) records representing locations where zoonotic Ebola virus has not been reported. This study was limited to the African continent since no natural outbreaks of EVD have occurred outside the continent. Only Reston ebolavirus has a distribution reported outside of Africa, focussed in the Philippines, but has never been reported as pathogenic in humans; as a result this species was not included in the analysis.

35Tables detailing proven outbreaks of Ebola virus, initially sourced from the scientific literature and from health reporting organisations, were used to coordinate searches of the formal scientific literature using Web of Science and PubMed for each specific outbreak. Relevant papers were abstracted and where possible outbreak-specific epidemiological surveys were sourced. The citations in these references were obtained in order to reconstruct the outbreak in detail and to identify the most probable index case. Index cases were defined as any human infection resulting from interaction with non-human sources of the disease. Some of these cases arose from presumed interactions with zoonotic reservoirs or hosts, such as primates and other mammals during hunting trips (;,; WHO, 2003) or butchering of bats. Any cases arising from existing human infections are considered as secondary infections rather than index cases. Similar to methodology employed elsewhere, the site, or supposed site, of this zoonotic transmission event was geopositioned using Google Earth. For locations where precise geographic information (e.g., geographic coordinates of a hunting camp) was provided by the authors, these were used. Where precise geographic information could not be accurately geopositioned, a geographic area (termed a polygon) was defined covering the reported region. In several cases only the first reported patient could be identified, with the source of infection unknown. With these outbreaks the location of the first patient was geopositioned under the assumption that an initial zoonotic spillover event occurred in the vicinity of this location. In two outbreaks multiple independent zoonotic transmission events were identified (WHO, 2003;; ), and each unique event was geopositioned and included in the model as separate entities. Table 1 catalogues the outbreaks used in this study.

36A literature search was conducted in Web of Science using the search term “Ebola” that returned 8635 citations. The abstracts were examined and for those that contained possible data on animal Ebola infection, the full text was obtained. The sampling site or location of the animal in the study was identified and geopositioned using Google Maps. These locations were recorded either as precise locations or as polygons, as with human index cases. Records for which local transmission of Ebola virus was deemed unlikely (e.g., seropositive primates tested in containment facilities several years after their capture) were excluded from the study. The non-human Ebola virus occurrence data collected are detailed in Table 2, including the diagnostic methods used.

37The open access sequence database GenBank was searched using MESH Umbrella search terms for Ebola virus, returning 181 results. These were then manually cross-referenced with the existing human and animal Ebola information, collected above, and 30 duplicates were removed. For the remaining isolates, original references and GenBank information fields were examined, but as there was insufficient information to establish precise location of isolation and/or whether the isolate represented an index case for any of these data sources, they were excluded from subsequent analyses.

38A suite of ecologically relevant gridded environmental covariates for Africa was compiled, each having a nominal resolution of 5 km × 5 km. The environmental covariates used in this analysis were: elevation (from the shuttle radar topography mission); the mean value, and a measure of spatial variation (range, described below) between 2000 and 2012 of Enhanced Vegetation Index (EVI), daytime Land Surface Temperature (LST) and night-time LST; and mean potential evapotranspiration from 1950–2000 (Figure 5—figure supplement 1).

39The EVI and LST datasets were derived from satellite imagery collected by NASA's Moderate Resolution Imaging Spectroradiometer (MODIS) remote sensing platform. EVI is a metric designed to characterise vegetation density and vigour based on the ratio of absorbed photosynthetically active radiation to near infrared radiation. LST is a modelled product derived from emissivity as measured by the MODIS thermal sensor, which is correlated, though not synonymous with air temperature, and effective for differentiating landscapes based on a combination of thermal energy and properties of surface types. The MODIS datasets utilized in this research (EVI was derived from the MCD43B4 product and the MOD11A2 LST product was used directly) were acquired as composite datasets created using imagery collected over multiple days, a procedure that results in products with 8-day temporal resolutions. Despite compositing, the EVI and LST datasets contained gaps due to persistent cloud cover found in forested equatorial regions, and these gaps were filled using a previously described approach. The EVI and LST datasets were then aggregated from their native 1 km × 1 km spatial resolution to a final 5 km × 5 km resolution, calculating both the mean and the range of the values of the subpixels making up each larger pixel. These spatial summaries therefore characterise both the mean temperature in each location as well as the degree of spatial heterogeneity within that pixel. This is of interest as humans and susceptible species are more likely to come into contact in transitional areas (e.g., boundary areas between areas of highly suitable susceptible species habitat and areas heavily utilised by humans). The final covariate production step consisted of summarising temporally across the 13-year data archive to produce synoptic datasets devoid of annual or seasonal anomalies.

40Over recent years, significant research has been undertaken in investigating the role bats have to play in the transmission cycle of Ebola viruses and evidence of asymptomatic infection in fruit bats has been documented to varying extents. In order to incorporate this potential driver of Ebola virus transmission into the model we developed predicted distribution maps for three species of fruit bat implicated as primary reservoirs of the disease: Hypsignathus monstrosus, Epomops franqueti and Myonycteris torquata. The evidence was strongest for these three species having a reservoir role as Ebola virus RNA (all nested within the Zaire ebolavirus phylogeny) has been detected in all three. Whilst a handful of other bat species have been found to be seropositive, no further viral isolations have been recorded.

41Whilst expert opinion range maps for these species exist, there is some disagreement with independently-sourced occurrence data (all archived in the Global Biodiversity Information Facility). As a result, a predictive modelling approach was used to create a continuous surface of habitat suitability for these species which we then included as a predictor in the model. Occurrence data for all Megachiroptera in Africa was extracted from GBIF using the R packages dismo and taxize. To remove apparently erroneous records in the GBIF archive all occurrence records more than 100 km from the species known ranges, as determined by expert-opinion range maps, were excluded, as were duplicate records and those located in the ocean. This resulted in a total dataset of 1341 unique occurrence records.

42The occurrence database was then used to train separate boosted regression tree species distribution models to predict the likely distribution of each of these suspected reservoir species. For each model, occurrence records for the target species (H. monstrosus, n = 67; E. franqueti, n = 120; and M. torquata, n = 52) were considered presence records and occurrence records of all other species were used as background records. This procedure is designed to account for the potentially biasing effect of spatial variation in recording of Megachiroptera occurrences.

43For each species we ran fifty submodels each trained to a randomly selected bootstrap of this dataset, subject to the constraint that each bootstrap contained a minimum of 10 occurrence and 10 background records. Each submodel was fitted using the gbm.step subroutine in the dismo R package. In each submodel the background records were down weighted so that the weighted sum of presence records equalled the weighted sum of background records in order to maximise the discrimination capacity of the model. We generated a prediction map from each of these submodels and calculated both the mean prediction and 95% confidence interval around the prediction for each 5 km × 5 km pixel for each species.

44Model accuracy was assessed by calculating the mean area under the curve (AUC) statistic for each submodel under a stringent 10-fold cross validation for each submodel and obtaining the mean and standard deviation across all 50 submodels. Under this procedure the dataset was split into ten subsets, each containing approximately equal numbers of presence and background points. The ability of a model trained on each subset to predict the distribution of the other 90% of records was assessed by AUC and the mean value taken. As so few presence records were used to train each fold model (i.e., around 5 presence records for M. torquata up to 12 for E. franqueti), this represents a very stringent test of the model's predictive capacity. Additionally, to prevent inflation of the accuracy statistics due to spatial sorting bias, these statistics were estimated using a pairwise distance sampling procedure. Consequently, the AUC statistics presented here are lower than would be returned by standard procedures but gives a more realistic quantification of the model's ability to extrapolate predictions to new regions. We also generated marginal effect plots with associated uncertainty intervals and relative contribution statistics (how often each covariate was selected during the model fitting process) as quantification of the sensitivity of the model to the different covariates. These allow us to make inferences about the ecological relationship between each species and its environment as well as to identify where this relationship is most uncertain.

45To generate a single surface describing the distribution of the bat reservoir species to be used as a covariate in the subsequent Ebola modelling, the three mean prediction distribution maps were merged by taking the average habitat suitability for each of the three bat species at each pixel.

46The Ebola virus occurrence dataset was supplemented with a background record dataset generated by randomly sampling 10,000 locations across Africa, biased towards more populous areas as a proxy for reporting bias. We fitted 500 submodels to bootstraps of this dataset. To account for uncertainty in the geographic location of those occurrences reported as polygons, for each submodel one point was randomly selected from each of these occurrence polygons. This Monte Carlo procedure enabled the model to efficiently integrate over the environmental uncertainty associated with imprecise geographic data. A bootstrap sample was then taken from each of these datasets and used to train the BRT model using the same procedure and weighting of background records as for the bat distribution models. Similarly, we generated a prediction map from each of these models and calculated both the mean prediction and corresponding 95% confidence intervals for each pixel and analysed prediction accuracy using the same stringent cross validation and sensitivity analysis procedure as for the bat distribution models (detailed above).

47The predicted distribution map produced by this approach represents the environmental suitability of each pixel for zoonotic Ebola virus transmission. This may be interpreted as a relative probability of presence in the sense that more suitable pixels are more likely to contain zoonotic transmission than less suitable pixels, though not an absolute probability that transmission occurs in a given pixel. Similarly, the presence of zoonotic transmission increases the risk of transmission to a human, though this is also contingent on how humans interact with these zoonotic pools, through hunting or other activities.

48In order to identify areas which are likely to be at risk of transmission of Ebolavirus from zoonotic reservoir hosts to humans, the continuous map of the predicted environmental suitability for zoonotic transmission (shown in Figure 5) was converted into a binary map classifying pixels as either at risk or not at risk. A pixel was assumed to be at risk if its predicted environmental suitability for zoonotic Ebola virus transmission was greater than 0.673, the lowest suitability value predicted at the locations of known transmission to humans (point records of human index cases). Countries containing at least one at-risk pixel are shown in Figure 5B—those countries that previously report an index case were defined as Set 1; countries with at least one at-risk pixel with no previous index cases of EVD were categorised as Set 2. The number of people living in at-risk areas in each of these countries was calculated by summing the estimated population of at-risk pixels using population density maps from the AfriPop project and the proportion of those living in urban, periurban and rural areas was evaluated using the Global Rural Urban Mapping Project.

49The R code used for all of the analysis has been made available on an open source basis (https://github.com/SEEG-Oxford/ebola_zoonotic).

50For three separate years (1976, 2000 and 2014), total national populations were retrieved and the proportion of rural to urban populations noted from World Bank statistics. To describe global air travel patterns from Set 1 and Set 2 countries, flight schedules data from the Official Airline Guide, reflecting an estimated 95% of all commercial flights worldwide, were analysed between 2000 and 2013 to calculate the annual volume of seats on direct flights that depart from each predicted country and which have an international destination. Complementing these seat capacity data, worldwide data on anonymised, individual passenger flight itineraries from the International Air Transport Association (2012) were analysed between 2005 and 2012 to calculate the annual volume of international passenger departures out of each Set 1 and Set 2 country. The IATA dataset represents an estimated 93% of the world's commercial air traffic at the passenger level and includes points of departure and arrival and final destination information for travellers as well as their connecting flights.