In 1996, highly pathogenic avian influenza of subtype H5N1 gave rise to the progenitor of the present H5N1 HPAI subtype in Guangdong, China (A/Goose/Guangdong/1/96[H5N1]) (Duan et al., 2007). Initially, restricted to Southern China, the virus started spreading in 2003 and by 2008, it had spread to more than 60 countries (Food and Agriculture Organization of the United Nations, 2016), persisting now in only a few. Transmission of infection from birds to humans was also reported, causing disease in 850 confirmed human cases including 449 deaths, as of May 2016, making this virus a continuing source of human health concerns (WHO/GIP, 2016; Lai et al., 2016). Re-assortment with other influenza viruses led to the replacement of most internal viral genes of the original H5N1 virus. However, the haemagglutinin (HA) gene H5 has remained present in all isolates and was therefore used to develop a standardised ‘clade’ nomenclature, first adopted in 2008, based on the evolution and divergence of H5N1 viruses that evolved from the original HA gene of the 1996 H5N1 virus (WHO/OIE/FAO H5N1 Evolution Working Group, 2008). In the initial years from 1996 to 2008, 10 distinct clades (0–9) had been generated and by 2012, 11 distinct actively circulating clades had been identified (World Health Organization/World Organisation for Animal Health/Food and Agriculture Organization (WHO/OIE/FAO) H5N1 Evolution Working Group, 2014).

Between 2009 and 2013, H5Nx HPAI viruses from the clade 2.3.4 showed an apparent geographical range expansion and were not only of the H5N1 subtype. Continuous live-poultry market surveillance in China identified novel clade 2.3.4 reassortant viruses of different H5N2, H5N5 and H5N8 subtypes, alongside H5N1 (Gu et al., 2011; Zhao et al., 2012, 2013). All these viruses were part of an H5 monophyletic group of viruses that shared the H5 gene of an H5N1 clade 2.3.4 variant with neuraminidase (NA) genes from different viruses (Gu et al., 2013). Consequently, the nomenclature of H5Nx viruses that clustered in this divergent HA group was updated as a new clade 2.3.4.4, in addition to two other new clades (Smith and Donis, 2015). From January 2014 onward, viruses of clade 2.3.4.4 started spreading internationally. The first H5N8 HPAI virus outbreaks outside of China were reported in South Korea and Japan in spring 2014 (Hill et al., 2015). In May 2014, a novel H5N6 clade 2.3.4.4 reassortant caused outbreaks in China and Lao PDR (Wong et al., 2015) and thereafter from Viet Nam and Myanmar. In November 2014, H5N8 HPAI viruses were reported from Germany, Netherlands, UK, Italy and the Russian Federation in rapid succession. In the same autumn and winter 2014/2015, H5N2 HPAI were reported from outbreaks in British Colombia, Canada. This virus contained genes similar to those of the Eurasian clade 2.3.4.4 alongside genes from North American wild bird lineages (World Organisation for Animal Health, 2016 ). In Taiwan, novel (H5N2, H5N3) reassortants also caused several outbreaks in 2014 (Lee et al., 2016). In December 2014, the new H5N8 and H5N2 HPAI viruses were detected in wild birds in Washington USA, before being found in poultry. By February 2015, the H5N2 HPAI virus had triggered a true epidemic in commercial poultry in the US, with nearly 43 million chickens and 7 million turkeys killed or culled across more than 20 different states (Poultry Science Association, 2016). All these HPAI H5N1, H5N2, H5N6 and H5N8 viruses found in Eurasia and North America shared an H5 gene segment belonging to clade 2.3.4.4 (Claes et al., 2014; Food and Agriculture Organization of the United Nations, 2016).

In summary, we can describe two periods and groups of viruses. From 2003 to 2010, the H5 HPAI viruses responsible for international spread and most outbreaks in poultry were of the N1 type, with continuous evolution of the H genes into different sub-lineages and gradual changes in its internal genes yielding clades and sub clades. From 2010 onward, H5Nx clade 2.3.4 viruses reassorted with several other avian influenza viruses leading to generation of a diversity of H5 clade 2.3.4.4 viruses bearing NA other than N1 in Asia. These novel reassortant viruses then began spreading internationally in 2013, in some cases further reassorting with viruses from other geographic lineages to yield new viruses, all bearing an H5 clade 2.3.4.4 haemagglutinin.

Following the spatio-temporal pattern of H5N1 HPAI spread, several spatial analytical studies were conducted to identify risk factors associated with H5N1 HPAI presence. The majority of these have been country-level studies in Thailand (Gilbert et al., 2006), Viet Nam (Minh et al., 2009), China (Martin et al., 2011), Bangladesh (Ahmed et al., 2012), Indonesia (Yupiana et al., 2010) and India (Dhingra et al., 2014). Several studies have also been conducted at regional (Adhikari D, 2009; Gilbert et al., 2008; Williams and Peterson, 2009) and continental levels (Hogerwerf et al., 2010; Peterson and Williams, 2008). Spatial risk factors associated with H5N1 HPAI presence through different studies were reviewed in 2012 (Gilbert and Pfeiffer, 2012) and the study highlighted domestic duck density, indices of water presence (distance to rivers and proportion of land occupied by water) and anthropogenic variables (human population density and distance to roads) to be the most consistent risk factors across studies, countries and scales. However, studies comparing different sets of factors were never carried out at a global scale, and none made a distinction between clades and sub-lineages.

In this analysis, we aimed to produce a first global suitability map for H5N1 HPAI virus sustained transmission, to establish its capacity to provide reliable spatial extrapolations at large spatial scales and to compare different sets of spatial predictor variables in their predictive capacity. Machine learning techniques have become very powerful in reproducing observed distribution patterns with sets of predictor variables, but their skill in spatial extrapolation is rarely quantified and could help better discriminate among sets of important predictor variables. In addition, the very fast recent spread of clade 2.3.4.4 H5Nx viruses (H5N1, H5N2, H5N6 and H5N8), associated with multiple reassortments was unprecedented (De Vries et al., 2015) and warranted further examination. A separate analysis of how 2.3.4.4 H5Nx viruses had spread in the geographical and environmental space was hence carried out in comparison to the HPAI H5N1 viruses.

Boosted Regression Trees (BRT) models were developed to predict the global suitability of H5N1 HPAI and H5Nx clade 2.3.4.4 presence. The predictor variables were categorised into four sets (Table 1) of variables. The Set 1 variables included the host variables of extensive and intensive chicken densities, human population density, and a variable to account for the effect of mass vaccination of poultry in China (IsChina). Set 2 included land cover variables with IsChina. Set 3 included Fourier-transformed climatic variables of land-surface temperature (LST) and Normalised Difference Vegetation Index (NDVI) with IsChina. Finally, Set 4 variables included Set 1 variables in addition to selected variables from the earlier sets that were selected on the basis of prior epidemiological knowledge. The models were subjected to three different types of cross validations to measure their goodness-of-fit (GOF) and transferability: (i) standard cross-validation (CV) with a random and stratified divide between training and validation sets, (ii) a calibrated cross-validation to account for the spatial sorting bias (SSB) sensu Hijmans (2012) i.e. the tendency to have distance between training-presence and testing-presence sites to be smaller than the distance between training-presence and testing-absence sites, and (iii) a spatial cross-validation (Spatial CV) to spatially separate the training and validation sets by large distances and measure the spatial extrapolation capacity of the models.

The bootstrapped goodness of fit values for the H5N1 HPAI and H5Nx HPAI clade 2.3.4.4 models for the different sets of covariates and cross validation methods are shown in Figure 1. For the H5N1 HPAI global model, all overall GOF metrics were good with predictive accuracy Area Under the Curve (AUC) values higher than 0.9 when evaluated through standard CV (Figure 1). The reduction in GOF taking into account the SSB was minor and followed the same pattern. However, when evaluated through spatial CV, the different sets of covariates showed contrasting AUC values. The land-use (Set 2) and eco-climatic (Set 3) based models extrapolated poorly, and the Set 1 and Set 4 performed best. It is also noteworthy, that the combination of Sets 1 and 2 (Set 2.1), or Sets 1 and 3 (Set 3.1) did not result in significantly better models than Set 1 alone (Figure 1—figure supplement 1), and even tended to reduce the average AUC of spatial CV.

Figure 1 with 1 supplement see all

Download asset Open asset

The models for the H5Nx clade 2.3.4.4 virus also had high GOF metrics estimated by standard CV (Figure 1). Here too, a significant amount of predictive power was already obtained with the models containing only Set 1 variables, with AUC values close to 0.9. There was a strong impact of spatial CV on the GOF metrics, with a drastic reduction in predictive power when extrapolating over large distances (Figure 1). Throughout the different spatial CV metrics, Set 2, and 4 showed better AUC values than Set 1, and given that Set 4 was more parsimonious, with fewer predictor variables, it was kept as the final model for H5Nx clade 2.3.4.4 suitability. Similar conclusions could be drawn from models using combinations of Set 1 and Set 2 (Set 2.1), or Set 1 and Set 3 (Set 3.1) (Figure 1—figure supplement 1).

The relative contribution (RC) of the predictor variables of Set 1 and Set 4 for H5N1 HPAI and the H5 HPAI clade 2.3.4.4 models are presented in Figure 2. The most noticeable difference concerned the role of domestic duck density, human population density and chicken density. The H5Nx HPAI clade 2.3.4.4 showed much higher RC for human population density and intensively raised chickens than the H5N1 HPAI one. Conversely, a comparatively much higher RC of domestic duck density and extensively raised chicken was observed for the H5N1 HPAI model than for the H5Nx HPAI clade 2.3.4.4 models. Upon the inclusion of additional predictors in Set 4 (Figure 2), the influence of these host-based predictor variables followed a similar pattern. In addition, annual mean temperature made a relatively high contribution in both models, and cultivated vegetation showed a much higher RC in the H5Nx HPAI clade 2.3.4.4 model than in the H5N1 HPAI one.

Figure 2

Download asset Open asset

Partial dependence plots of the BRT models allow the contribution of a particular variable to be depicted on the fitted response after taking into account the effect of all the other predictors in the model (Figure 3, and Figure 3—figure supplement 1). The main difference between the partial dependence plots of the different variables was for the density of extensively raised chickens, which showed a positive association with H5N1 HPAI presence contrasting with an absence of association with the H5Nx HPAI clade 2.3.4.4 presence (Figure 3). Other profiles were somewhat comparable for the two groups of viruses and showed a positive association between virus presence and duck density, intensively raised chicken density, human population density, a negative association with the IsChina variable (Figure 3) and an optimum for percentage of cropland and temperature (Figure 3—figure supplement 1). It should be kept in mind that their relative contributions, i.e. their weight in the final prediction strongly differed between the two groups of viruses. It is noteworthy that the models outlined above were built using optimal number of trees estimated through spatial CV instead of standard CV, and this resulted in much lower optimal number of trees compared to standard CV models (Figure 3—figure supplement 2), suggesting that standard CV may be over fitting local clusters of presence points rather than making reliable large-distance predictions. The suitability maps of the models are presented in Figure 4. To interpret the extrapolation capacity of these suitability maps, multivariate environmental similarity surfaces (MESS) (Elith et al., 2010) were computed (Figure 4—figure supplement 3) giving information on where the models extrapolate within the range of predictor variables in the occurrence points. As observed, both models extrapolate prediction in areas with similar environmental conditions, as depicted by positive MESS values. However, the geographical space with high similarity to the occurrence point is comparatively wider for the HPAI H5N1 model, than for the H5Nx clade 2.3.4.4 models.

Figure 3 with 2 supplements see all

Download asset Open asset

Figure 4 with 3 supplements see all

Download asset Open asset

Figure 4—source data 1

Suitability predictions for the HPAI H5N1 best model (GeoTiff format).

https://doi.org/10.7554/eLife.19571.010

Download elife-19571-fig4-data1-v4.zip

Figure 4—source data 2

Suitability predictions for the H5Nx clade 2.3.4.4 best model (GeoTiff format).

https://doi.org/10.7554/eLife.19571.011

Download elife-19571-fig4-data2-v4.zip

As expected, high suitability values for the H5N1 HPAI model (Figure 4) are found in several parts of Asia, including China (when the effect of the IsChina variable is removed). Other areas where H5N1 HPAI has spread extensively are highlighted, such as eastern Indo-Gangetic plain, Thailand central plain, south Myanmar and the Red river and Mekong deltas of Vietnam, the island of Java in Indonesia and the Nile Delta in Egypt. The model also highlights areas where H5N1 HPAI was introduced but did not persist over long periods of time, such as in South Korea, Japan, Ukraine and Romania. Areas of western Africa, such as Nigeria, where the H5N1 HPAI outbreaks have been unfolding since late 2014 have been predicted as suitable by the model. Isolated parts in Eastern Europe, North America, Mexico, Dominican Republic and South America, are also deemed suitable for H5N1 establishment.

The suitability map for the H5Nx HPAI clade 2.3.4.4 virus is somewhat different, highlighting more isolated areas (Figure 4). The spatial extrapolation capacity of this model was low, and predictions made at large distances from known points of presence should be interpreted with caution. As this clade is still spreading, there may still be large areas of the landscape where it could potentially become established and where the model predictions may be inaccurate. In areas close to presence points, the predictions are believed to be robust, with several areas within Asia, such as China, South Korea, Japan and Taiwan depicted as suitable. The well-known virus ‘reassortment-sink’ areas of the Indo-Gangetic plains, the river deltas of Vietnam, southern Myanmar and Java, Indonesia, are also highlighted as areas of suitability. In Africa, the Nile delta is depicted as suitable for establishment. In North America, the high suitability areas match the intensive poultry areas of the Midwestern and southern states of USA. The Netherlands, Belgium and northwest France are highlighted with high suitability in Europe. In Australia, the commercial poultry rearing areas of Victoria and New South Wales are predicted as suitable; even though HPAI subtype H5 has never been reported in Australia.

The suitability predictions for the HPAI H5N1 and H5Nx clade 2.3.4.4 best models using Set 1 and Set 4 predictor variables are in Figure 4—source data 1 and 2, respectively. The suitability maps for H5Nx HPAI and H5 clade 2.3.4.4 for Sets 2 and 3 are presented in Figure 4—figure supplement 1 and Figure 4—figure supplement 2, respectively.

A first important result of this study is that it was possible to build a global suitability model for HPAI H5N1 virus with a high extrapolation capacity robustly established through spatial cross-validation. Interestingly, H5N1 HPAI outbreaks appeared to be best modelled by predictor variables relating to host distribution. Alternative models based on land use or eco-climatic variables showed marginally better accuracy metrics when evaluated with standard CV, but significantly lower extrapolation capacity than the host-only variable. Even the models combining host variables with other environmental predictors did not produce significantly better results when evaluated through spatial CV. This observation matches earlier observations that association with eco-climatic variables were not consistently reproducible across countries and studies (Gilbert and Pfeiffer, 2012) and that H5N1 HPAI is probably not as strongly environmentally constrained as other authors have suggested (Williams and Peterson, 2009; Zhang et al., 2014). This strongly contrasts with vector-borne diseases, where clear eco-climatic boundaries of vectors can be mapped, and where climate has a strong influence on vector seasonality and population dynamics (McMichael and Lindgren, 2011; Morin and Comrie, 2013). In the case of a directly contagious disease such as avian influenza, successful transmission and clinical outbreaks have been observed over a wide range of temperature and humidity conditions (e.g. Russia, Nigeria, Egypt, Northern China, Indonesia). Our results suggest that the main large-scale constrains to suitability for H5N1 HPAI occurrence are related to the distribution of hosts; densities of chickens and ducks raised in different systems, and to the density of the human population, probably as a surrogate measure for various anthropogenic transmission mechanisms. For the new H5Nx clade 2.3.4.4 viruses, we found a somewhat different result, with a clear improvement of the extrapolation capacity of models using a set of variables combining host distribution and environmental variables. However, these models were of relatively low overall predictive power, most likely because the virus has not yet had a chance to extend fully to its potential range of occurrence as compared to H5N1 HPAI, and false pseudo-absences may have had a strong impact on the construction of models and, therefore, on the accuracy of predictions. For this model, and given its low extrapolation capacity, we emphasise that predictions made at long distances from points of presence should be interpreted with caution, as there may still be large areas where it could potentially become established and where the our model predictions may be inaccurate.

A second important result of this study is to demonstrate the importance of spatial CV in building and validating avian influenza suitability models over large geographical extents. The difference between standard and spatial CV evaluation of GOF was already quite significant for the HPAI H5N1 models. The difference was even more striking for the H5Nx clade 2.3.4.4 models, which all appeared very good when evaluated through standard CV. However, they had poor extrapolation accuracy, sometimes even not much better than a null model when evaluated through spatial CV. Machine-learning techniques used in species distribution modelling have become incredibly powerful at reproducing a pattern from a given set of occurrence data and this is essentially what standard CV measures. However, as our results demonstrate, default cross-validation technique is a very misleading measure of their geographical extrapolation capacity. Spatial CV was found not only to be important for evaluating the extrapolation capacity of a given model, but also to be the only way to truly discriminate our model outputs based on different sets of predictor variables. The focus on extrapolation capacity for selecting predictor variables is driven by the assumption that a model that includes statistical relationships linked to causal mechanisms should spatially extrapolate well, as the cause-consequence statistical associations have a greater chance to apply well in different places than those that are coincidental. Of course, cause-consequence relationships may vary in space too, but the underlying assumption of suitability modelling extrapolation is that these remain constant over the spatial domain within which the model is applied. So, models based on coincidental statistical associations are expected to extrapolate poorly in the geographical domain, and these losses of predictability can hardly be quantified through standard CV because of spatial correlations between training and validation sets.

A third set of important results consisted in the comparison of the H5N1 HPAI and H5Nx clade 2.3.4.4 models, which showed areas of convergences and differences in the geographic and predictor variables spaces. Domestic duck density was the most important variable for both models, though with a lower RC for the H5Nx clade 2.3.4.4 model in Set 4. Ducks have always been strongly associated with areas of persistence and evolution of H5N1 HPAI (Gilbert and Pfeiffer, 2012), which relates to their capacity to act as an intermediate, domestic reservoir between wild Anatidae, the main wild reservoir of avian influenza viruses, and domesticated poultry. Ducks have been referred to as the ‘Trojan horses’ for H5N1 HPAI H5N1 presence (Kim et al., 2014) on account of their role in virus introduction, evolution, transmission and persistence (Hulse-Post et al., 2005), which has been demonstrated in both host pathogenicity (Cornelissen et al., 2013; Smith and Donis, 2015) and geospatial studies (Gilbert and Pfeiffer, 2012). The absence of duck density may in fact explain a lot of the difference in extrapolating capacity found between the host-model (Set 1) and the land-use and eco-climatic models (Set 2 and 3) that cannot discriminate areas with similar land-use and eco-climatic conditions, but that have very different duck densities. For example, India is predicted at relatively high suitability by the land-use model (Set 2, Figure 4) at a very low suitability by the host-based (except around Bangladesh) reflecting the near-absence of significant domestic duck densities in much of the country, in accordance with previous results (Gilbert et al., 2010).

The finding of a strong association between H5Nx clade 2.3.4.4 and ducks was somewhat less expected as the disease was found mostly in chicken farms in more intensive poultry production areas, but results are, however, in line with those of Hill et al. (2015) who found through phylogeographic analysis that the introduction of H5Nx clade 2.3.4.4 to South Korea was associated with areas where domestic ducks and wild waterfowl intermingled. Complex reassortment of multiple subtypes may also occur in areas where domestic ducks and migratory birds have an opportunity to share food, water and habitat, creating opportunities for virus transmission between different species, co-infection of individual animals with different influenza viruses and subsequent gene reassortment (Deng et al., 2013). It would be prudent for countries to put such areas under active surveillance for early detection of HPAI introductions and for monitoring of virus evolution. This would include the countries of the Americas and African continent where duck rearing is not as common as in South East Asia. It is noteworthy that one of the most severe recent H5 HPAI epidemics that started in 2015 in Dordogne region of France, a traditionally important duck rearing area with some of the highest duck densities in the country, even if the outbreaks were apparently caused by distinct H5 viruses from those circulating in Asia. So, the association found with domestic duck densities fits with existing knowledge of H5N1 spatial epidemiology and was a major predictor in both the H5N1 HPAI and H5Nx clade 2.3.4.4 models.

In contrast, the association with extensively and intensively raised chickens provided different results for the H5N1 HPAI and H5Nx clade 2.3.4.4 models, with the latter being more strongly associated with intensified chicken production systems, found in intensive crop production areas with high human population densities. An interesting hypothesis to explain this pattern would be a greater fitness of H5Nx clade 2.3.4.4 viruses to spread through intensive chicken production and poultry trade systems (Claes et al., 2016). We still lack extensive published experimental infection results of the new clade in poultry, but preliminary results are indicative of a lower pathogenicity of the H5Nx clade 2.3.4.4 virus in chickens compared to H5N1 HPAI, with longer survival and shedding period (Kim et al., 2014; Swayne et al., 2015). A lower virulence in chicken was also found for the reassortant H5N2, H5N6 and H5N8 clade 2.3.4.4 viruses compared to previous 2.3.4 HPAI H5N1 viruses (Sun et al., 2016), although they remained highly pathogenic. A lower mortality and longer period of infectivity may assist the virus in circulating longer and within intensified poultry production and trading systems, leading to increased opportunities for onward transmission. Evolution towards reduced pathogenicity would appear an asset in improving farm-to-farm transmission and long-term persistence even in the absence of domestic ducks. This could partly explain the stronger association of H5Nx clade 2.3.4.4 viruses with intensive chicken production areas in eastern Asia and in the US.

Our analyses have focussed on poultry outbreak locations and are therefore of more limited use in identifying the locations of initial introduction of avian influenza viruses, or places where viruses may undergo more frequent reassortment events leading to the local emergence of new viruses. Future work may look more explicitly into those aspects and could lead to better prevention at the sources of virus introduction and emergence.

1. 1. Adhikari D CA

   (2009)

   Modeling the ecology and distribution of highly pathogenic avian influenza (H5N1) in the indian subcontinent

   Current Science 13:2391–2397.

   • Google Scholar
2. 1. Ahmed SSU
   2. Ersbøll AK
   3. Biswas PK
   4. Christensen JP
   5. Hannan ASMA
   6. Toft N

   (2012) Ecological Determinants of Highly pathogenic avian Influenza (H5N1) Outbreaks in Bangladesh

   PLoS One 7:e33938.

   https://doi.org/10.1371/journal.pone.0033938

   • Google Scholar
3. 1. Castellan DM
   2. Hinrichs J
   3. Fusheng G
   4. Sawitri E
   5. Dung do H
   6. Martin V
   7. McGrane J
   8. Bandyopadhayay FS
   9. Inui K
   10. Yamage M
   11. Ahmed GM
   12. Macfarlane L
   13. Williams T
   14. Dissanayake R
   15. Akram M
   16. Kalpravidh W
   17. Gopinath CY
   18. Morzaria S

   (2014) Development and application of a vaccination planning tool for avian influenza

   Avian Diseases 58:437–452.

   https://doi.org/10.1637/10827-032414-Reg.1

   • PubMed
   • Google Scholar
4. 1. Claes F
   2. Kuznetsov D
   3. Liechti R
   4. Von Dobschuetz S
   5. Truong BD
   6. Gleizes A
   7. Conversa D
   8. Colonna A
   9. Demaio E
   10. Ramazzotto S
   11. Larfaoui F
   12. Pinto J
   13. Le Mercier P
   14. Xenarios I
   15. Dauphin G
   16. Dinh Truong B

   (2014) The EMPRES-i genetic module: a novel tool linking epidemiological outbreak information and genetic characteristics of influenza viruses

   Database 2014:bau008.

   https://doi.org/10.1093/database/bau008

   • PubMed
   • Google Scholar
5. 1. Claes F
   2. Morzaria SP
   3. Donis RO

   (2016) Emergence and dissemination of clade 2.3.4.4 H5Nx influenza viruses-how is the Asian HPAI H5 lineage maintained

   Current Opinion in Virology 16:158–163.

   https://doi.org/10.1016/j.coviro.2016.02.005

   • PubMed
   • Google Scholar
6. 1. Cornelissen JB
   2. Vervelde L
   3. Post J
   4. Rebel JM

   (2013) Differences in highly pathogenic avian influenza viral pathogenesis and associated early inflammatory response in chickens and ducks

   Avian Pathology 42:347–364.

   https://doi.org/10.1080/03079457.2013.807325

   • PubMed
   • Google Scholar
7. 1. De Vries E
   2. Guo H
   3. Dai M
   4. Rottier PJ
   5. van Kuppeveld FJ
   6. de Haan CA

   (2015) Rapid Emergence of highly pathogenic Avian Influenza Subtypes from a Subtype H5N1 Hemagglutinin variant

   Emerging Infectious Diseases 21:842–846.

   https://doi.org/10.3201/eid2105.141927

   • PubMed
   • Google Scholar
8. 1. Deng G
   2. Tan D
   3. Shi J
   4. Cui P
   5. Jiang Y
   6. Liu L
   7. Tian G
   8. Kawaoka Y
   9. Li C
   10. Chen H

   (2013) Complex reassortment of multiple subtypes of avian influenza viruses in domestic ducks at the Dongting Lake Region of China

   Journal of Virology 87:9452–9462.

   https://doi.org/10.1128/JVI.00776-13

   • PubMed
   • Google Scholar
9. 1. Dhingra MS
   2. Dissanayake R
   3. Negi AB
   4. Oberoi M
   5. Castellan D
   6. Thrusfield M
   7. Linard C
   8. Gilbert M

   (2014) Spatio-temporal epidemiology of highly pathogenic avian influenza (subtype H5N1) in poultry in eastern India

   Spatial and Spatio-Temporal Epidemiology 11:45–57.

   https://doi.org/10.1016/j.sste.2014.06.003

   • Google Scholar
10. 1. Domenech J
    2. Dauphin G
    3. Rushton J
    4. McGrane J
    5. Lubroth J
    6. Tripodi A
    7. Gilbert J
    8. Sims L

    (2009) Experiences with vaccination in countries endemically infected with highly pathogenic avian influenza: the Food and Agriculture Organization perspective

    Revue Scientifique Et Technique De l'OIE 28:293–305.

    https://doi.org/10.20506/rst.28.1.1865

    • Google Scholar
11. 1. Duan L
    2. Campitelli L
    3. Fan XH
    4. Leung YH
    5. Vijaykrishna D
    6. Zhang JX
    7. Donatelli I
    8. Delogu M
    9. Li KS
    10. Foni E
    11. Chiapponi C
    12. Wu WL
    13. Kai H
    14. Webster RG
    15. Shortridge KF
    16. Peiris JS
    17. Smith GJ
    18. Chen H
    19. Guan Y

    (2007) Characterization of low-pathogenic H5 subtype influenza viruses from Eurasia: implications for the origin of highly pathogenic H5N1 viruses

    Journal of Virology 81:7529–7539.

    https://doi.org/10.1128/JVI.00327-07

    • PubMed
    • Google Scholar
12. 1. Elith J
    2. H. Graham C
    3. P. Anderson R
    4. Dudík M
    5. Ferrier S
    6. J. Hijmans R
    7. J. Hijmans R
    8. R. Leathwick J
    9. R. Leathwick J
    10. Lehmann A
    11. G. Lohmann L
    12. A. Loiselle B
    13. A. Loiselle B
    14. Manion G
    15. Moritz C
    16. Nakamura M
    17. McC. M. Overton J
    18. Townsend Peterson A
    19. J. Phillips S
    20. J. Phillips S
    21. Richardson K
    22. E. Schapire R
    23. E. Schapire R
    24. Soberón J
    25. S. Wisz M
    26. E. Zimmermann N
    27. E. Zimmermann N
    28. Overton JMM
    29. Peterson AT

    (2006) Novel methods improve prediction of species’ distributions from occurrence data

    Ecography 29:129–151.

    https://doi.org/10.1111/j.2006.0906-7590.04596.x

    • Google Scholar
13. 1. Elith J
    2. Kearney M
    3. Phillips S

    (2010) The art of modelling range-shifting species

    Methods in Ecology and Evolution 1:330–342.

    https://doi.org/10.1111/j.2041-210X.2010.00036.x

    • Google Scholar
14. 1. Elith J
    2. Leathwick JR
    3. Hastie T

    (2008) A working guide to boosted regression trees

    Journal of Animal Ecology 77:802–813.

    https://doi.org/10.1111/j.1365-2656.2008.01390.x

    • PubMed
    • Google Scholar
15. Website

    1. Food and Agriculture Organization of the United Nations

    (2016) EMPRES-i - Global Animal Disease Information System

    Accessed 2016.

    http://empres-i.fao.org/eipws3g/ 
16. 1. Friedman JH
    2. Meulman JJ

    (2003) Multiple additive regression trees with application in epidemiology

    Statistics in Medicine 22:1365–1381.

    https://doi.org/10.1002/sim.1501

    • PubMed
    • Google Scholar
17. 1. Gaughan AE
    2. Stevens FR
    3. Linard C
    4. Jia P
    5. Tatem AJ

    (2013) High resolution population distribution maps for Southeast Asia in 2010 and 2015

    PLoS One 8:e55882.

    https://doi.org/10.1371/journal.pone.0055882

    • PubMed
    • Google Scholar
18. 1. Gilbert M
    2. Chaitaweesub P
    3. Parakamawongsa T
    4. Premashthira S
    5. Tiensin T
    6. Kalpravidh W
    7. Wagner H
    8. Slingenbergh J

    (2006) Free-grazing ducks and highly pathogenic avian influenza, Thailand

    Emerging Infectious Diseases 12:227–234.

    https://doi.org/10.3201/eid1202.050640

    • PubMed
    • Google Scholar
19. 1. Gilbert M
    2. Conchedda G
    3. Van Boeckel TP
    4. Cinardi G
    5. Linard C
    6. Nicolas G
    7. Thanapongtharm W
    8. D'Aietti L
    9. Wint W
    10. Newman SH
    11. Robinson TP

    (2015) Income disparities and the global distribution of intensively farmed chicken and pigs

    PLoS One 10:e0133381.

    https://doi.org/10.1371/journal.pone.0133381

    • PubMed
    • Google Scholar
20. 1. Gilbert M
    2. Golding N
    3. Zhou H
    4. Wint GR
    5. Robinson TP
    6. Tatem AJ
    7. Lai S
    8. Zhou S
    9. Jiang H
    10. Guo D
    11. Huang Z
    12. Messina JP
    13. Xiao X
    14. Linard C
    15. Van Boeckel TP
    16. Martin V
    17. Bhatt S
    18. Gething PW
    19. Farrar JJ
    20. Hay SI
    21. Yu H

    (2014) Predicting the risk of avian influenza A H7N9 infection in live-poultry markets across Asia

    Nature Communications 5:4116.

    https://doi.org/10.1038/ncomms5116

    • PubMed
    • Google Scholar
21. 1. Gilbert M
    2. Newman SH
    3. Takekawa JY
    4. Loth L
    5. Biradar C
    6. Prosser DJ
    7. Balachandran S
    8. Subba Rao MV
    9. Mundkur T
    10. Yan B
    11. Xing Z
    12. Hou Y
    13. Batbayar N
    14. Natsagdorj T
    15. Hogerwerf L
    16. Slingenbergh J
    17. Xiao X

    (2010) Flying over an infected landscape: distribution of highly pathogenic avian influenza H5N1 risk in South Asia and satellite tracking of wild waterfowl

    EcoHealth 7:448–458.

    https://doi.org/10.1007/s10393-010-0672-8

    • PubMed
    • Google Scholar
22. 1. Gilbert M
    2. Pfeiffer DU

    (2012) Risk factor modelling of the spatio-temporal patterns of highly pathogenic avian influenza (HPAIV) H5N1: a review

    Spatial and Spatio-Temporal Epidemiology 3:173–183.

    https://doi.org/10.1016/j.sste.2012.01.002

    • PubMed
    • Google Scholar
23. 1. Gilbert M
    2. Xiao X
    3. Pfeiffer DU
    4. Epprecht M
    5. Boles S
    6. Czarnecki C
    7. Chaitaweesub P
    8. Kalpravidh W
    9. Minh PQ
    10. Otte MJ
    11. Martin V
    12. Slingenbergh J

    (2008) Mapping H5N1 highly pathogenic avian influenza risk in Southeast Asia

    PNAS 105:4769–4774.

    https://doi.org/10.1073/pnas.0710581105

    • PubMed
    • Google Scholar
24. 1. Gu M
    2. Liu W
    3. Cao Y
    4. Peng D
    5. Wang X
    6. Wan H
    7. Zhao G
    8. Xu Q
    9. Zhang W
    10. Song Q
    11. Li Y
    12. Liu X

    (2011) Novel reassortant highly pathogenic avian influenza (H5N5) viruses in domestic ducks, China

    Emerging Infectious Diseases 17:1060–1063.

    https://doi.org/10.3201/eid/1706.101406

    • PubMed
    • Google Scholar
25. 1. Gu M
    2. Zhao G
    3. Zhao K
    4. Zhong L
    5. Huang J
    6. Wan H
    7. Wang X
    8. Liu W
    9. Liu H
    10. Peng D
    11. Liu X

    (2013) Novel variants of clade 2.3.4 highly pathogenic avian influenza A(H5N1) viruses, China

    Emerging Infectious Diseases 19:2021–2024.

    https://doi.org/10.3201/eid1912.130340

    • PubMed
    • Google Scholar
26. 1. Hijmans RJ

    (2012) Cross-validation of species distribution models: removing spatial sorting bias and calibration with a null model

    Ecology 93:679–688.

    https://doi.org/10.1890/11-0826.1

    • PubMed
    • Google Scholar
27. 1. Hill SC
    2. Lee Y-J
    3. Song B-M
    4. Kang H-M
    5. Lee E-K
    6. Hanna A
    7. Gilbert M
    8. Brown IH
    9. Pybus OG

    (2015) Wild waterfowl migration and domestic duck density shape the epidemiology of highly pathogenic H5N8 influenza in the Republic of Korea

    Infection, Genetics and Evolution 34:267–277.

    https://doi.org/10.1016/j.meegid.2015.06.014

    • Google Scholar
28. 1. Hogerwerf L
    2. Wallace RG
    3. Ottaviani D
    4. Slingenbergh J
    5. Prosser D
    6. Bergmann L
    7. Gilbert M

    (2010) Persistence of highly pathogenic avian influenza H5N1 virus defined by agro-ecological niche

    EcoHealth 7:213–225.

    https://doi.org/10.1007/s10393-010-0324-z

    • PubMed
    • Google Scholar
29. 1. Hulse-Post DJ
    2. Sturm-Ramirez KM
    3. Humberd J
    4. Seiler P
    5. Govorkova EA
    6. Krauss S
    7. Scholtissek C
    8. Puthavathana P
    9. Buranathai C
    10. Nguyen TD
    11. Long HT
    12. Naipospos TS
    13. Chen H
    14. Ellis TM
    15. Guan Y
    16. Peiris JS
    17. Webster RG

    (2005) Role of domestic ducks in the propagation and biological evolution of highly pathogenic H5N1 influenza viruses in Asia

    PNAS 102:10682–10687.

    https://doi.org/10.1073/pnas.0504662102

    • PubMed
    • Google Scholar
30. 1. Kim YI
    2. Pascua PN
    3. Kwon HI
    4. Lim GJ
    5. Kim EH
    6. Yoon SW
    7. Park SJ
    8. Kim SM
    9. Choi EJ
    10. Si YJ
    11. Lee OJ
    12. Shim WS
    13. Kim SW
    14. Mo IP
    15. Bae Y
    16. Lim YT
    17. Sung MH
    18. Kim CJ
    19. Webby RJ
    20. Webster RG
    21. Choi YK

    (2014) Pathobiological features of a novel, highly pathogenic avian influenza A(H5N8) virus

    Emerging Microbes & Infections 3:e75.

    https://doi.org/10.1038/emi.2014.75

    • PubMed
    • Google Scholar
31. 1. Lai S
    2. Qin Y
    3. Cowling BJ
    4. Ren X
    5. Wardrop NA
    6. Gilbert M
    7. Tsang TK
    8. Wu P
    9. Feng L
    10. Jiang H
    11. Peng Z
    12. Zheng J
    13. Liao Q
    14. Li S
    15. Horby PW
    16. Farrar JJ
    17. Gao GF
    18. Tatem AJ
    19. Yu H

    (2016) Global epidemiology of avian influenza A H5N1 virus infection in humans, 1997-2015: a systematic review of individual case data

    The Lancet Infectious Diseases 16:e108–e118.

    https://doi.org/10.1016/S1473-3099(16)00153-5

    • PubMed
    • Google Scholar
32. 1. Lee MS
    2. Chen LH
    3. Chen YP
    4. Liu YP
    5. Li WC
    6. Lin YL
    7. Lee F

    (2016) Highly pathogenic avian influenza viruses H5N2, H5N3, and H5N8 in Taiwan in 2015

    Veterinary Microbiology 187:50–57.

    https://doi.org/10.1016/j.vetmic.2016.03.012

    • PubMed
    • Google Scholar
33. 1. Linard C
    2. Gilbert M
    3. Snow RW
    4. Noor AM
    5. Tatem AJ

    (2012) Population distribution, settlement patterns and accessibility across Africa in 2010

    PLoS One 7:e31743.

    https://doi.org/10.1371/journal.pone.0031743

    • PubMed
    • Google Scholar
34. 1. Liu CM
    2. Lin SH
    3. Chen YC
    4. Lin KC
    5. Wu TS
    6. King CC

    (2007) Temperature drops and the onset of severe avian influenza a H5N1 virus outbreaks

    PLoS One 2:e191.

    https://doi.org/10.1371/journal.pone.0000191

    • PubMed
    • Google Scholar
35. 1. Martin V
    2. Pfeiffer DU
    3. Zhou X
    4. Xiao X
    5. Prosser DJ
    6. Guo F
    7. Gilbert M

    (2011) Spatial distribution and risk factors of highly pathogenic avian influenza (HPAI) H5N1 in China

    PLoS Pathogens 7:e1001308.

    https://doi.org/10.1371/journal.ppat.1001308

    • PubMed
    • Google Scholar
36. 1. McMichael AJ
    2. Lindgren E

    (2011) Climate change: present and future risks to health, and necessary responses

    Journal of Internal Medicine 270:401–413.

    https://doi.org/10.1111/j.1365-2796.2011.02415.x

    • PubMed
    • Google Scholar
37. 1. Minh PQ
    2. Morris RS
    3. Schauer B
    4. Stevenson M
    5. Benschop J
    6. Nam HV
    7. Jackson R

    (2009) Spatio-temporal epidemiology of highly pathogenic avian influenza outbreaks in the two deltas of Vietnam during 2003-2007

    Preventive Veterinary Medicine 89:16–24.

    https://doi.org/10.1016/j.prevetmed.2009.01.004

    • PubMed
    • Google Scholar
38. 1. Morin CW
    2. Comrie AC

    (2013) Regional and seasonal response of a West Nile virus vector to climate change

    PNAS 110:15620–15625.

    https://doi.org/10.1073/pnas.1307135110

    • PubMed
    • Google Scholar
39. 1. Peterson AT
    2. Williams RAJ

    (2008) Risk Mapping of highly pathogenic avian influenza distribution and spread

    Ecology and Society 13:15.

    https://doi.org/10.5751/ES-02532-130215

    • Google Scholar
40. 1. Phillips SJ
    2. Dudík M
    3. Elith J
    4. Graham CH
    5. Lehmann A
    6. Leathwick J
    7. Ferrier S

    (2009) Sample selection bias and presence-only distribution models: implications for background and pseudo-absence data

    Ecological Applications 19:181–197.

    https://doi.org/10.1890/07-2153.1

    • PubMed
    • Google Scholar
41. 1. Pigott DM
    2. Golding N
    3. Mylne A
    4. Huang Z
    5. Henry AJ
    6. Weiss DJ
    7. Brady OJ
    8. Kraemer MU
    9. Smith DL
    10. Moyes CL
    11. Bhatt S
    12. Gething PW
    13. Horby PW
    14. Bogoch II
    15. Brownstein JS
    16. Mekaru SR
    17. Tatem AJ
    18. Khan K
    19. Hay SI

    (2014) Mapping the zoonotic niche of Ebola virus disease in Africa

    eLife 3:e04395.

    https://doi.org/10.7554/eLife.04395

    • PubMed
    • Google Scholar
42. Website

    1. Poultry Science Association

    (2016) Poultry News Archive

    Accessed 2016.

    http://www.poultryscience.org/poultrynews_archive.asp?autotry=true&ULnotkn=true
43. 1. Randin CF
    2. Dirnböck T
    3. Dullinger S
    4. Zimmermann NE
    5. Zappa M
    6. Guisan A

    (2006) Are niche-based species distribution models transferable in space?

    Journal of Biogeography 33:1689–1703.

    https://doi.org/10.1111/j.1365-2699.2006.01466.x

    • Google Scholar
44. 1. Rijks J
    2. ElMasry I

    (2009)

    Characteristics of poultry production in Egyptian villages and their effect on HPAI vaccination campaign results - Results of a participatory epidemiology study

    Food and Agriculture Organisation Report.

    • Google Scholar
45. 1. Robinson TP
    2. Wint GR
    3. Conchedda G
    4. Van Boeckel TP
    5. Ercoli V
    6. Palamara E
    7. Cinardi G
    8. D'Aietti L
    9. Hay SI
    10. Gilbert M

    (2014) Mapping the global distribution of livestock

    PLoS One 9:e96084.

    https://doi.org/10.1371/journal.pone.0096084

    • PubMed
    • Google Scholar
46. 1. Sawitri ES
    2. Darminto J
    3. Weaver
    4. Bouma A

    (2007)

    The vaccination program in Indonesia. In: Vaccination: a tool for the control of avian influenza, Proc. of the Joint OIE/FAO/IZSVe Conference, Verona, Italy March 2007

    Istituto Zooprofilattico Sperimentale Delle Venezie, Dodet and the Scientific and Technical Department of the O.I.E., Eds 130:151–158.

    • Google Scholar
47. 1. Scharlemann JP
    2. Benz D
    3. Hay SI
    4. Purse BV
    5. Tatem AJ
    6. Wint GR
    7. Rogers DJ

    (2008) Global data for ecology and epidemiology: a novel algorithm for temporal Fourier processing MODIS data

    PLoS One 3:e1408.

    https://doi.org/10.1371/journal.pone.0001408

    • PubMed
    • Google Scholar
48. 1. Shaman J
    2. Kohn M

    (2009) Absolute humidity modulates influenza survival, transmission, and seasonality

    PNAS 106:3243–3248.

    https://doi.org/10.1073/pnas.0806852106

    • PubMed
    • Google Scholar
49. 1. Smith GJD
    2. Donis RO

    (2015) Nomenclature updates resulting from the evolution of avian influenza A(H5) virus clades 2.1.3.2a, 2.2.1, and 2.3.4 during 2013–2014 - Smith - 2015 - Influenza and Other Respiratory Viruses - Wiley Online Library

    Influenza Other Respiratory Viruses 9:271–276.

    https://doi.org/10.1111/irv.12324

    • Google Scholar
50. Website

    1. Socioeconomic Data and Applications Center

    (2016) Population Count Grid, v3: Gridded Population of the World (GPW), v3 | SEDAC

    Accessed 2016.

    http://sedac.ciesin.columbia.edu/data/set/gpw-v3-population-count
51. 1. Sorichetta A
    2. Hornby GM
    3. Stevens FR
    4. Gaughan AE
    5. Linard C
    6. Tatem AJ

    (2015) High-resolution gridded population datasets for Latin America and the Caribbean in 2010, 2015, and 2020

    Scientific Data 2:150045.

    https://doi.org/10.1038/sdata.2015.45

    • PubMed
    • Google Scholar
52. 1. Sun H
    2. Pu J
    3. Hu J
    4. Liu L
    5. Xu G
    6. Gao GF
    7. Liu X
    8. Liu J

    (2016) Characterization of clade 2.3.4.4 highly pathogenic H5 avian influenza viruses in ducks and chickens

    Veterinary Microbiology 182:116–122.

    https://doi.org/10.1016/j.vetmic.2015.11.001

    • PubMed
    • Google Scholar
53. Website

    1. Swayne DE
    2. Bertran K
    3. Kapczynski DR
    4. Pantin Jackwood MJ
    5. Spackman E
    6. Suarez DL

    (2015) Reduced experimental infectivity and transmissibility of intercontinental H5 (H5N8 and H5N2) compared to eurasian H5N1 highly pathogenic avian influenza viruses for chickens, turkeys, and japanese quail

    Accessed 2015.

    http://www.ars.usda.gov/research/publications/publications.htm?SEQ_NO_115=315175
54. 1. Tuanmu M-N
    2. Jetz W

    (2014) A global 1-km consensus land-cover product for biodiversity and ecosystem modelling

    Global Ecology and Biogeography 23:1031–1045.

    https://doi.org/10.1111/geb.12182

    • Google Scholar
55. 1. Wenger SJ
    2. Olden JD

    (2012) Assessing transferability of ecological models: an underappreciated aspect of statistical validation

    Methods in Ecology and Evolution 3:260–267.

    https://doi.org/10.1111/j.2041-210X.2011.00170.x

    • Google Scholar
56. Website

    1. WHO/GIP

    (2016) Cumulative number of confirmed human cases for avian influenza A(H5N1) reported to WHO

    Accessed 2016.

    http://www.who.int/influenza/human_animal_interface/2016_02_25_tableH5N1.pdf?584 ua=1
57. 1. WHO/OIE/FAO H5N1 Evolution Working Group

    (2008) Toward a unified nomenclature system for highly pathogenic avian influenza virus (H5N1)

    Emerging Infectious Diseases 14:e1.

    https://doi.org/10.3201/eid1407.071681

    • Google Scholar
58. 1. Williams RA
    2. Peterson AT

    (2009) Ecology and geography of avian influenza (HPAI H5N1) transmission in the Middle East and northeastern Africa

    International Journal of Health Geographics 8:47.

    https://doi.org/10.1186/1476-072X-8-47

    • PubMed
    • Google Scholar
59. 1. Wong FY
    2. Phommachanh P
    3. Kalpravidh W
    4. Chanthavisouk C
    5. Gilbert J
    6. Bingham J
    7. Davies KR
    8. Cooke J
    9. Eagles D
    10. Phiphakhavong S
    11. Shan S
    12. Stevens V
    13. Williams DT
    14. Bounma P
    15. Khambounheuang B
    16. Morrissy C
    17. Douangngeun B
    18. Morzaria S

    (2015) Reassortant highly pathogenic influenza A(H5N6) virus in Laos

    Emerging Infectious Diseases 21:511–516.

    https://doi.org/10.3201/eid2103.141488

    • PubMed
    • Google Scholar
60. 1. World Health Organization/World Organisation for Animal Health/Food and Agriculture Organization (WHO/OIE/FAO) H5N1 Evolution Working Group

    (2014) Revised and updated nomenclature for highly pathogenic avian influenza A (H5N1) viruses

    Influenza and Other Respiratory Viruses 8:384–388.

    https://doi.org/10.1111/irv.12230

    • Google Scholar
61. Website

    1. World Organisation for Animal Health

    (2016) OIE World Animal Health Information System

    Accessed 2016.

    http://www.oie.int/wahis_2/public/wahid.php/Diseaseinformation/diseasehome
62. 1. Yupiana Y
    2. de Vlas SJ
    3. Adnan NM
    4. Richardus JH

    (2010) Risk factors of poultry outbreaks and human cases of H5N1 avian influenza virus infection in West Java Province, Indonesia

    International Journal of Infectious Diseases 14:e800–e805.

    https://doi.org/10.1016/j.ijid.2010.03.014

    • PubMed
    • Google Scholar
63. 1. Zhang Z
    2. Chen D
    3. Chen Y
    4. Wang B
    5. Hu Y
    6. Gao J
    7. Sun L
    8. Li R
    9. Xiong C

    (2014) Evaluating the impact of environmental temperature on global highly pathogenic avian influenza (HPAI) H5N1 outbreaks in domestic poultry

    International Journal of Environmental Research and Public Health 11:6388–6399.

    https://doi.org/10.3390/ijerph110606388

    • PubMed
    • Google Scholar
64. 1. Zhao G
    2. Gu X
    3. Lu X
    4. Pan J
    5. Duan Z
    6. Zhao K
    7. Gu M
    8. Liu Q
    9. He L
    10. Chen J
    11. Ge S
    12. Wang Y
    13. Chen S
    14. Wang X
    15. Peng D
    16. Wan H
    17. Liu X

    (2012) Novel reassortant highly pathogenic H5N2 avian influenza viruses in poultry in China

    PLoS One 7:e46183.

    https://doi.org/10.1371/journal.pone.0046183

    • PubMed
    • Google Scholar
65. 1. Zhao K
    2. Gu M
    3. Zhong L
    4. Duan Z
    5. Zhang Y
    6. Zhu Y
    7. Zhao G
    8. Zhao M
    9. Chen Z
    10. Hu S
    11. Liu W
    12. Liu X
    13. Peng D
    14. Liu X

    (2013) Characterization of three H5N5 and one H5N8 highly pathogenic avian influenza viruses in China

    Veterinary Microbiology 163:351–357.

    https://doi.org/10.1016/j.vetmic.2012.12.025

    • PubMed
    • Google Scholar

Thank you for submitting your article "Global mapping of highly pathogenic avian influenza H5N1 and H5Nx clade 2.3.4.4 viruses with spatial cross-validation" for consideration by eLife. Your article has been favorably evaluated by Wendy Garrett (Senior Editor) and three reviewers, one of whom served as a guest Reviewing Editor. One of the three reviewers, Richard J Webby, has agreed to share his name.

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

Summary:

The manuscript describes the development of global risk maps for H5N1 and H5NX viruses. Building on previous work mapping risk by country and by region, the authors sought to build parsimonious global suitability models based on host and environmental factors. In order to discriminate among models incorporating different variables the authors used boosted regression trees and identified human and animal population densities as key factors influencing risk (as supported by previous work). While the basic factors identified are perhaps not unexpected, refinement and improvement to mapping practices are needed to help support global preparedness and response.

All reviewers agree that the data shed light on an important and interesting problem. However, before we can recommend acceptance, a number of issues should be addressed.

Essential revisions:

1) It is not clear why the distinction was made between clade 2.3.4.4 viruses and all H5 viruses others. As pointed out, the 2.3.3.4 viruses did have different neuraminidases and did transmit to North America, but there is no convincing evidence that we are aware of that supports the authors assumption that "they may have different phenotypic characteristics leading to different transmission patterns". There are very likely different risk factors combined in the "other clade" category. For example, clade 2.2 and clade 2.3.2.1 viruses spread out of Asia much like 2.3.4.4 (with obvious exception of migration to Americas) and at least appeared more associated with wild bird spread. Were any other subclade analyses done? Is there evidence, or at the least practical reason, to support clumping all other clades together?

2) The use of human population distribution for pseudo-absence generation needs to be better justified. Given that the purpose of pseudo-absence generation is to reflect the bias in sampling or reporting effort present, it is not obvious that human population density best reflects this difference. We suggest the authors consider approaches that integrate other strains of avian influenza as direct absence locations. Given EMPRESi as a data source, this should at least be feasible at least for H7 viruses. Otherwise, the authors need to provide more context for why such testing would scale with human population; it could be argued that testing could scale with avian distributions better than it does humans (e.g. in the United States). There may also exist other, more appropriate, high resolution proxies for this reporting bias.

3) As pointed out in the Methods section, the inference of pseudo-absences can be biased by the fact that the pathogen has not been introduced into a region. This is undoubtedly true but the assertion that this bias is likely to be stronger for H5Nx than H5N1 viruses is peculiar. The suitability maps in Figure 4 differ substantially for North America, presumably because of the lack of H5N1 infections and thus inferred absence. However, as in comment 1 above, we are unaware of evidence to suggest that there are phenotypic differences between the H5N1 and H5Nx viruses that would make North America suitable for H5Nx viruses but unsuitable for H5N1 viruses. It could be argued that H5N1 has also not yet had the opportunity to explore its full potential range.

4) In the sixth paragraph of the Results the authors discuss the results of Chinas inclusion in the high suitability zone, but essentially only when the IsChina variable was removed. What exactly does this mean regarding the influence of this variable? It is obviously clear China is very much a high suitability zone for these viruses. Along similar lines, if IsChina is meant to account for mass poultry vaccination in China, why have similar variables been ignored for other countries that have or have had mass vaccination campaigns – particularly Indonesia, Egypt, and Vietnam?

5) The authors produce predictions that are many miles away from any previously reported occurrences – whilst the dotted lines on Figure 4 help the reader be aware of this, the authors should also consider a multivariate environmental similarity surface approach to gauge the extent to which these records are extrapolation, and where predicted values fall within environmental space already considered by the presence points already in place. There is an example of its usage in Elith et al. 2010 http://onlinelibrary.wiley.com/doi/10.1111/j.2041-210X.2010.00036.x/full

6) We are concerned about the independence of the occurrence points, particularly with respect to secondary spread rather than an inherent feature of the underlying environment in these locations. An example of where this is problematic would be charting the progress of Nipah outbreaks in Malaysia in 1998 – pigs and humans over a widespread area were infected yet in terms of environmental suitability, only one location, where the index case arose, is suitable for inclusion. All other cases are more appropriately modelled by understanding the nature of swine connectivity via livestock trade and human-animal interactions. To what extent are these influenza cases located, and to what extent are they environmentally driven?

7) This exposes a broader issue relating to the inclusion of the poultry layers in the model. Niche maps are simply reflections on the potential for presence of the pathogen – by using data from poultry and then including poultry measures as predictors it is therefore unsurprising that they have a high explanatory power within the model, particularly where there may be ineffective control for the biasing effects this may have. As a consequence, the maps may be a more accurate prediction of where outbreaks are more likely to occur (or be reported) rather than an accurate depiction of potential occurrence (i.e. anything from singular cases to very many). It was interesting therefore to see how the inclusion/exclusion of different covariates impacted this – even exploring what impact that has when converted into a map would be interesting to see, particularly an output based just upon Sets 2 and 3 alone.

[Editors' note: further revisions were requested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "Global mapping of highly pathogenic avian influenza H5N1 and H5Nx clade 2.3.4.4 viruses with spatial cross-validation" for further consideration at eLife. Your revised article has been favorably evaluated by Wendy Garrett (Senior editor), a Reviewing Editor, and one of the original reviewers.

The manuscript has been improved but there are some remaining issues that need to be addressed before acceptance, as outlined below:

Reviewer #2:

The authors response to reviewer comments have clarified some initial concerns as well as provided interesting additional analyses. However, I do have some still outstanding issues, as well as some newly raised concerns given the authors response.

It was interesting to note the use of gazetteers to approximate disease occurrence since this adds an additional concern as to how reliable this classification is, and if in doubt, what steps were taken to mitigate (e.g. turning specific lat longs into larger polygons capturing such uncertainty).

The authors state that different pseudo-absence models produced "similar results", but provide no means for reviewers to gauge that similarity.

The clarification the authors present that these outputs are depicting suitability for infection, including secondary spread, to my mind is problematic, since this requires mixing different types of data together within the same model. It is therefore unclear to me how the disparate effects of environment (likely influencing primary infection) and human/movement/transmission networks (likely influencing secondary infection) will be disaggregated. Whilst the cross-validation may mitigate the potential clustering that primary and secondary cases have, I'm still not 100% convinced that this differing process can be adequately captured simultaneously. To my mind a more powerful demonstration of difference (or indeed lack of) would be a comparison of primary and secondary data alone.

It was good to see the inclusion of Set 2 and Set 3 results – what was disappointing however was that there was no discussion of the differences this produces (sometimes considerable e.g. India for HPAI). How much of this variation is due to differences in ratios of primary v. secondary cases considered as occurrence records across the different geographies for instance? I found it very interesting to see such a different picture result when the human/host variables were dropped.

1) It is not clear why the distinction was made between clade 2.3.4.4 viruses and all H5 viruses others. As pointed out, the 2.3.3.4 viruses did have different neuraminidases and did transmit to North America, but there is no convincing evidence that we are aware of that supports the authors assumption that "they may have different phenotypic characteristics leading to different transmission patterns". There are very likely different risk factors combined in the "other clade" category. For example, clade 2.2 and clade 2.3.2.1 viruses spread out of Asia much like 2.3.4.4 (with obvious exception of migration to Americas) and at least appeared more associated with wild bird spread. Were any other subclade analyses done? Is there evidence, or at the least practical reason, to support clumping all other clades together?

There were a few recent virology studies that suggest milder clinical disease in duck favoring long-distance transmission (e.g. DeJesus et al. 2016 doi:10.1016/j.virol.2016.08.036), but we agree that this statement of possible different phenotypic characteristics is a bit too speculative to justify the distinction. Making a full clade-level analysis would go beyond the scope of the present study, and would require in-depth data mining to be able to allocate clades to all known outbreaks. This is an approach we recently followed, over a much more limited geographical domain (Artois et al. 2016 doi 10.1038/srep30316).

However, the very fast spread of clade 2.3.4.4 H5Nx viruses and the fact that it involved multiple reassortments leading to many different viruses with different neuraminidase is quite unprecedented (de Vries et al. 2015 doi 10.3201/eid2105.141927), and warranted further examination of how it had spread in the geographical and environmental space, in comparison to previous known HPAI H5N1 viruses.

So, we removed the most speculative statements and tried to better justify this grouping based on the above argument. We also somewhat restructured the Introduction to make it clearer that the main focus of the paper is on the spatial epidemiology of HPAI H5N1 in general, and that the analysis of the 2.3.4.4 H5Nx viruses comes as a comparative addition.

2) The use of human population distribution for pseudo-absence generation needs to be better justified. Given that the purpose of pseudo-absence generation is to reflect the bias in sampling or reporting effort present, it is not obvious that human population density best reflects this difference. We suggest the authors consider approaches that integrate other strains of avian influenza as direct absence locations. Given EMPRESi as a data source, this should at least be feasible at least for H7 viruses. Otherwise, the authors need to provide more context for why such testing would scale with human population; it could be argued that testing could scale with avian distributions better than it does humans (e.g. in the United States). There may also exist other, more appropriate, high resolution proxies for this reporting bias.

This choice was made for several reasons. First, in a context of underreporting as happening in much of Asia, we assumed that under-reporting would be easier in remote areas with low population density than in regions with high population density, where the disposal or dead birds and carcasses would be less likely to go unnoticed. Second, the Empres-I database compiles outbreak locations data from very heterogeneous sources and the geo-referencing of individual cases, in the absence of explicit X-Y locations, is often made through place names gazetteers that will tend to place outbreak location in the closest village, or administrative unit main village rather in the exact location of the farm where the disease was found, which will introduce a bias correlated with human population density.

Third, in much of Asia, rural population density is somewhat correlated with poultry density. However, we agree that this needed to be established more carefully, so we re-ran our models using poultry density instead of human population density to generate pseudo-absences, and we obtained very similar results.

We looked into the possibility of using H7 viruses, but the large majority of H7 records in Empres-I are in fact H7N9 positive samples from markets in China (Gilbert et al. 2014, doi 10.1038/ncomms5116), which are concentrated within cities. So, we felt that this would have the risk of introducing an unwanted bias.

The choice of human population density for pseudo-absence generation is now better justified in the manuscript (added in subsection “Modelling procedure”.)

3) As pointed out in the Methods section, the inference of pseudo-absences can be biased by the fact that the pathogen has not been introduced into a region. This is undoubtedly true but the assertion that this bias is likely to be stronger for H5Nx than H5N1 viruses is peculiar. The suitability maps in Figure 4 differ substantially for North America, presumably because of the lack of H5N1 infections and thus inferred absence. However, as in comment 1 above, we are unaware of evidence to suggest that there are phenotypic differences between the H5N1 and H5Nx viruses that would make North America suitable for H5Nx viruses but unsuitable for H5N1 viruses. It could be argued that H5N1 has also not yet had the opportunity to explore its full potential range.

We did not infer absence of HPAI H5N1 in North America, as our pseudo- absences were filtered by geographic distance from positives (as indicated in the Materials and methods), which prevented any pseudo-absence to fall with the Americas. So, the absence of significant suitability for HPAI H5N1 is not related to inferred absences there. We agree, however, that we can hardly exclude that H5N1 could do well in North America, but not necessarily so. One should keep in mind that HPAI H5N1 was seeded throughout Western Europe in several countries and locations and never turned into a real epidemic. Domestic duck density has a relatively high relative contribution in our models, and was shown in many previous researches to be associated with HPAI H5N1 spread in Asia at multiple spatial scale. The absence of high domestic duck densities may then partly explain the predicted low suitability in the Americas.

We removed the statement suggesting that this effect may be stronger for the H5Nx in the Materials and methods section and now only comment the study results, i.e. that the extrapolation capacity quantified through spatial cross validation was found to be lower for the H5Nx viruses. In addition, we updated the figure with pseudo-absences to better visualize the respective sets of pseudo- absences used for each model (HPAI H5N1 and H5Nx clade 2.3.4.4).

4) In the sixth paragraph of the Results the authors discuss the results of Chinas inclusion in the high suitability zone, but essentially only when the IsChina variable was removed. What exactly does this mean regarding the influence of this variable? It is obviously clear China is very much a high suitability zone for these viruses. Along similar lines, if IsChina is meant to account for mass poultry vaccination in China, why have similar variables been ignored for other countries that have or have had mass vaccination campaigns – particularly Indonesia, Egypt, and Vietnam?

China is very much in the high suitability zone for these viruses. To clarify, even with the inclusion of the IsChina variable, large parts of China still remain suitable for HPAI H5N1. But we have strong reasons to believe that the effect of mass-vaccination is far more pronounced in China than in these other countries, which justifies using this term in our models. Post vaccination seropositivity in China ranges between 80 and 95% in China (see Figure 5 in Martin et al. 2011 doi: 10.1371/journal.ppat.1001308), whereas papers having looked at post- vaccination seropositivity in Indonesia (Sawitri et al. 2007), Vietnam (Domenech et al. 2009), and Egypt (Rijks 2009) found it to be insufficient, often close to 30%. China is by far also the biggest user of vaccines: “125 billion doses of H5N1 vaccine were produced and deployed in total; in China (120 billion), Indonesia (3 billion), and Viet Nam (2 billion) between 2004 and 2012” (Castellan et al., 2014). This justification has been added to the manuscript (subsection “Spatial predictor variables”, fourth paragraph).

Sawitri, E. S., Darminto, J. Weaver, and A. Bouma. The vaccination program in Indonesia. In: Vaccination: a tool for the control of avian influenza, Proc. of the Joint OIE/FAO/IZSVe Conference, Verona, Italy March 2007. Istituto Zooprofilattico Sperimentale delle Venezie, Dodet and the Scientific and Technical Department of the O.I.E., eds. 130:151– 158.Karger, Basel, Switzerland. 2007.

Domenech, J., G. Dauphin, J. Rushton, J. McGrane, J. Lubroth, A. Tripodi, J. Gilbert, and L. D. Sims. Experiences with vaccination in countries endemically infected with highly pathogenic avian influenza: the Food and Agriculture Organization perspective. O.I.E. Rev. Sci. Tech. 28:293– 305. 2009.

Rijks J, ElMasry I. Food and Agriculture Organisation report. 2009. Characteristics of poultry production in Egyptian villages and their effect on HPAI vaccination campaign results – Results of a participatory epidemiology study.

Castellan, D.M., Hinrichs, J., Fusheng, G., Sawitri, E., Dung, D.H., Martin, V., McGrane, J., Bandyopadhayay, S., Inui, K., Yamage, M., Ahmed, G.M., Macfarlane, L., Williams, T., Dissanayake, R., Akram, M., Kalpravidh, W., Gopinath, C.Y., Morzaria, S., 2014. Development and Application of a Vaccination Planning Tool for Avian Influenza. Avian Dis. 58, 437–452. doi:10.1637/10827-032414-Reg.1

5) The authors produce predictions that are many miles away from any previously reported occurrences – whilst the dotted lines on Figure 4 help the reader be aware of this, the authors should also consider a multivariate environmental similarity surface approach to gauge the extent to which these records are extrapolation, and where predicted values fall within environmental space already considered by the presence points already in place. There is an example of its usage in Elith et al. 2010 http://onlinelibrary.wiley.com/doi/10.1111/j.2041-210X.2010.00036.x/full

Thank you for this suggestion. We estimated the MESS for our different set of predictors and for our two set of occurrence points, and those are now added as supplementary information. As can be seen from the plot, we had very limited geographical areas with high predicted suitability outside of the range of values found in the occurrence points, and the range in the full set of points (positive and pseudo-absence) would only be larger.

6) We are concerned about the independence of the occurrence points, particularly with respect to secondary spread rather than an inherent feature of the underlying environment in these locations. An example of where this is problematic would be charting the progress of Nipah outbreaks in Malaysia in 1998 – pigs and humans over a widespread area were infected yet in terms of environmental suitability, only one location, where the index case arose, is suitable for inclusion. All other cases are more appropriately modelled by understanding the nature of swine connectivity via livestock trade and human-animal interactions. To what extent are these influenza cases located, and to what extent are they environmentally driven?

We believe that there was an element of misunderstanding here. Our suitability maps depict suitability for infection, including secondary spread, and not HPAI H5N1 or H5Nx clade 2.3.4.4 introduction or emergence events alone. This is precisely why we included anthropogenic and poultry factors, i.e. these are more likely to explain patterns of spread than environment variables, as our results quantitatively demonstrate. We agree that there is spatial dependence in the occurrence points, but this is precisely why we implemented the spatial cross- validation, so that we would not overestimate the goodness of fit of our models through correlated validation sets.

7) This exposes a broader issue relating to the inclusion of the poultry layers in the model. Niche maps are simply reflections on the potential for presence of the pathogen – by using data from poultry and then including poultry measures as predictors it is therefore unsurprising that they have a high explanatory power within the model, particularly where there may be ineffective control for the biasing effects this may have. As a consequence, the maps may be a more accurate prediction of where outbreaks are more likely to occur (or be reported) rather than an accurate depiction of potential occurrence (i.e. anything from singular cases to very many). It was interesting therefore to see how the inclusion/exclusion of different covariates impacted this – even exploring what impact that has when converted into a map would be interesting to see, particularly an output based just upon Sets 2 and 3 alone.

We agree, but a particular aspect of HPAI H5N1 is that the probability of presence was never found to necessarily scale-up with chicken numbers (where the disease is visible because of high mortality), but rather scales up with duck density (where the disease circulates more silently). This was very apparent in the review by Gilbert & Pfeiffer (2012, doi 10.1016/j.sste.2012.01.002). In many separate countries and studies, the probability of presence of HPAI H5N1 at different levels was not associated with high chicken density, most likely because of many different production systems associating with different patterns of transmission. In fact, an important point of the paper is precisely that for an infectious disease like HPAI H5N1, host-related variables are much better at spatially predicting disease occurrence than eco-climatic ones. This was evidenced by the fact that models combining host variables with other environmental predictors did not produce significantly better results when evaluated through spatial CV.

We agree that the outputs of Set 2 and 3 alone would be of potential interest to the readers and have included them as supplementary information.

[Editors' note: further revisions were requested prior to acceptance, as described below.]

[…]

Reviewer #2:

The authors’ response to reviewer comments have clarified some initial concerns as well as provided interesting additional analyses. However, I do have some still outstanding issues, as well as some newly raised concerns given the authors response.

It was interesting to note the use of gazetteers to approximate disease occurrence since this adds an additional concern as to how reliable this classification is, and if in doubt, what steps were taken to mitigate (e.g. turning specific lat longs into larger polygons capturing such uncertainty).

This could have been a concern if the analysis was carried out with a finer grain, and in fact, we implemented bootstrapping to sample through polygons in a previous study carried out in the Mekong region only (Artois et al. 2016 doi 10.1038/srep30316), at a spatial resolution of 0.0083333 decimal degree (approx. 1 km at the equator). In this case, we are using 17,068 records spread over three continents, at a spatial resolution of 0.083333 decimal degrees (approx. 10 km), so the influence of the spatial uncertainty of a fraction of these points should be quite negligible. For example, only 3.5% of the records of Empres-I were reported at the coarsest administrative level 1, 5.8% at administrative level 2, and the rest of the records were either at administrative level 3, or from locations of unknown quality for which little can be said about the position uncertainty. We’ve looked at this more carefully for national data sets that we know better, and Thailand alone recorded over 1500 records, all linked to administrative unit level 3, which have a median size of 16 km², i.e. far smaller than our pixels. The scaling was somewhat similar for Vietnam, with outbreaks geo-referenced at the commune level. The addition of a bootstrapping procedure accounting for spatial uncertainty would be technically feasible, but would only apply for the records with a known quality, and we feel that given the strong spatial autocorrelation between neighboring pixels in most of the predictor variables, it would add a significant burden of processing for a negligible difference.

The authors state that different pseudo-absence models produced "similar results", but provide no means for reviewers to gauge that similarity.

We simply did not want to flood the reader with supplementary information material. A long and complex modeling procedure as this one entails many choices made at different points. Showing, all the possible consequences of the choices become intractable and of somewhat limited use. The figures below show the predicted distribution of the pseudo-absences distributed according to human populations (Author response image 1) and poultry (Author response image 2), with the IsChina term set to 0.

Author response image 1

Download asset Open asset

Author response image 2

Download asset Open asset

The differences are hardly noticeable (see, for example the small differences along the Iraq/Iran border) and we would be happy to add these predictions to the manuscript should the editor feel this is a necessity, alongside other results, but we feel this would simply add an extra load of results with limited added informational value.

The clarification the authors present that these outputs are depicting suitability for infection, including secondary spread, to my mind is problematic, since this requires mixing different types of data together within the same model. It is therefore unclear to me how the disparate effects of environment (likely influencing primary infection) and human/movement/transmission networks (likely influencing secondary infection) will be disaggregated. Whilst the cross-validation may mitigate the potential clustering that primary and secondary cases have, I'm still not 100% convinced that this differing process can be adequately captured simultaneously. To my mind a more powerful demonstration of difference (or indeed lack of) would be a comparison of primary and secondary data alone.

The distinction between primary and secondary infections is simply not straightforward for a HPAI virus. For example, HPAI H5N1 is depicted to have emerged in China in the late nineties, and only in the years 2003/2004 it started to spread internationally across Southeast Asia. What is the primary event in this situation? The emergence point in southern China? Or the first time it was introduced in these surrounding countries? If the former is considered, there’s only one event, i.e. the emergence point of the new HPAI H5N1 virus. If the latter is considered, these primary introductions in Vietnam, Thailand or even Indonesia may have as much to do with human and poultry trade network as they had with the environment. Or maybe the reviewer refereed to the first introduction of an HPAI H5N1 virus in a new area through wild birds? But how could one make such distinction out of pattern analysis?

A distinction between primary and secondary cases for HPAI would have to be based either be done using space-time filters to detect the first record in a particular area, which could introduce some bias by how it is defined (which area? how big? a continent? a country? a province? which time windows?), or would need to be based on some molecular definitions allowing to trace the hosts through which the virus transmitted, which would require individual sequences and extensive phylogeographic processing.

And even if this distinction was feasible, it wouldn’t necessarily allow differentiating those processes. Human movements and poultry trade transmission networks can also be at the origin of novel introductions in a country, and have been in several cases, and, conversely, short-distance contagious spread could also involve resident birds or environmental factors (e.g. water-borne, wind-borne).

This is probably why, out of the 47 papers on risk factor modeling of HPAI H5N1 reviewed in Gilbert & Pfeiffer (2012, doi: 10.1016/j.sste.2012.01.002), only one tried to make such distinction based on a space-time filter windows in Thailand. Studies that have looked at HPAI H5N1 outbreaks by type (in poultry vs. wild birds) are more frequent, but this doesn’t apply here, as our study only included records of HPAI found in poultry.

It was good to see the inclusion of Set 2 and Set 3 results – what was disappointing however was that there was no discussion of the differences this produces (sometimes considerable e.g. India for HPAI). How much of this variation is due to differences in ratios of primary v. secondary cases considered as occurrence records across the different geographies for instance? I found it very interesting to see such a different picture result when the human/host variables were dropped.

The paper is already fairly long and contains several messages, so we didn’t want to dwell too much into discussing the full set of results of models that we demonstrate to be poorer.

We don’t think it has to do with the ratio of primary/secondary cases. Rather, the most likely cause of the difference is the near absence of domestic ducks in the country, except in the few states surrounding Bangladesh. As discussed, domestic ducks play an important role in the epidemiology of HPAI H5N1 as silent carriers. So, when the poultry variables are dropped, the land-use model cannot make the difference between cropping/agricultural areas where ducks are abundant from those where they are scarce, and it therefore predicts higher risk in India than the host-based model, with poorer general performances. The resulting risk map is in fact in fairly good accordance with an analysis published previously and looking specifically at the situation in South Asia (Gilbert et al. 2010, doi 10.1007/s10393-010-0672-8 & Dhingra et al. 2014 doi 10.1016/j.sste.2014.06.003). See, in particular, Figure 1B in Gilbert et al. (2010) with the distribution of ducks in South Asia.

We have added a couple of lines to explain this in the Discussion (fifth paragraph) as we agree it is an illustrative example to explain why the host-based model could provide better extrapolations than the land-use one.

Author details

1. Madhur S Dhingra

   1. Spatial Epidemiology Lab, Université Libre de Bruxelles, Brussels, Belgium
   2. Department of Animal Husbandry and Dairying, Government of Haryana, Panchkula, India

   Contribution

   MSD, Acquisition of data, Analysis and interpretation of data, Drafting or revising the article

   Contributed equally with

   Jean Artois

   Competing interests

   The authors declare that no competing interests exist.

2. Jean Artois

   Spatial Epidemiology Lab, Université Libre de Bruxelles, Brussels, Belgium

   Contribution

   JA, Acquisition of data, Analysis and interpretation of data, Drafting or revising the article

   Contributed equally with

   Madhur S Dhingra

   Competing interests

   The authors declare that no competing interests exist.

3. Timothy P Robinson

   Livestock Systems and Environment, International Livestock Research Institute, Nairobi, Kenya

   Contribution

   TPR, Conception and design, Drafting or revising the article

   Competing interests

   The authors declare that no competing interests exist.

   "This ORCID iD identifies the author of this article:" 0000-0002-4266-963X

4. Catherine Linard

   1. Spatial Epidemiology Lab, Université Libre de Bruxelles, Brussels, Belgium
   2. Department of Geography, Université de Namur, Namur, Belgium

   Contribution

   CL, Acquisition of data, Drafting or revising the article

   Competing interests

   The authors declare that no competing interests exist.

5. Celia Chaiban

   Spatial Epidemiology Lab, Université Libre de Bruxelles, Brussels, Belgium

   Contribution

   CC, Analysis and interpretation of data, Drafting or revising the article

   Competing interests

   The authors declare that no competing interests exist.

6. Ioannis Xenarios

   1. Swiss-Prot and Vital-IT group, Swiss Institute of Bioinformatics, Lausanne, Switzerland
   2. Center for Integrative Genomics, University of Lausanne, Lausanne, Switzerland

   Contribution

   IX, Conception and design, Drafting or revising the article, Contributed unpublished essential data or reagents

   Competing interests

   The authors declare that no competing interests exist.

7. Robin Engler

   Swiss-Prot and Vital-IT group, Swiss Institute of Bioinformatics, Lausanne, Switzerland

   Contribution

   RE, Analysis and interpretation of data, Drafting or revising the article

   Competing interests

   The authors declare that no competing interests exist.

8. Robin Liechti

   Swiss-Prot and Vital-IT group, Swiss Institute of Bioinformatics, Lausanne, Switzerland

   Contribution

   RL, Drafting or revising the article, Contributed unpublished essential data or reagents

   Competing interests

   The authors declare that no competing interests exist.

9. Dmitri Kuznetsov

   Swiss-Prot and Vital-IT group, Swiss Institute of Bioinformatics, Lausanne, Switzerland

   Contribution

   DK, Drafting or revising the article, Contributed unpublished essential data or reagents

   Competing interests

   The authors declare that no competing interests exist.

10. Xiangming Xiao

    1. Department of Microbiology and Plant Biology, University of Oklahoma, Norman, United States
    2. Center for Spatial Analysis, University of Oklahoma, Norman, United States
    3. Institute of Biodiversity Science, Fudan University, Shanghai, China

    Contribution

    XX, Acquisition of data, Drafting or revising the article

    Competing interests

    The authors declare that no competing interests exist.

11. Sophie Von Dobschuetz

    Animal Production and Health Division, Food and Agriculture Organization of the United Nations, Rome, Italy

    Contribution

    SVD, Conception and design, Acquisition of data, Drafting or revising the article

    Competing interests

    The authors declare that no competing interests exist.

12. Filip Claes

    Emergency Center for Transboundary Animal Diseases, FAO Regional Office for Asia and the Pacific, Bangkok, Thailand

    Contribution

    FC, Conception and design, Drafting or revising the article

    Competing interests

    The authors declare that no competing interests exist.

13. Scott H Newman

    Emergency Center for Transboundary Animal Diseases, Food and Agriculture Organization of the United Nations, Hanoi, Vietnam

    Contribution

    SHN, Conception and design, Drafting or revising the article

    For correspondence

    scott.newman@fao.org

    Competing interests

    The authors declare that no competing interests exist.

14. Gwenaëlle Dauphin

    Animal Production and Health Division, Food and Agriculture Organization of the United Nations, Rome, Italy

    Contribution

    GD, Conception and design, Acquisition of data, Drafting or revising the article

    For correspondence

    Gwenaelle.Dauphin@fao.org

    Competing interests

    The authors declare that no competing interests exist.

15. Marius Gilbert

    1. Spatial Epidemiology Lab, Université Libre de Bruxelles, Brussels, Belgium
    2. Fonds National de la Recherche Scientifique, Brussels, Belgium

    Contribution

    MG, Conception and design, Acquisition of data, Analysis and interpretation of data, Drafting or revising the article

    For correspondence

    marius.gilbert@gmail.com

    Competing interests

    The authors declare that no competing interests exist.

    "This ORCID iD identifies the author of this article:" 0000-0003-3708-3359

• 3,945

  Page views

• 727

  Downloads

• 33

  Citations

Article citation count generated by polling the highest count across the following sources: Crossref, Scopus, PubMed Central.