In China, where approximately 15 billion head of poultry are produced annually with a standing population of 5.6 billion chickens, 760 million ducks and 300 million geese, major regional differences are apparent in ecological systems, husbandry practices, cultural behaviors and economic development with a consequential impact on the distribution of infectious diseases including HPAI H5N1, as well as their maintenance and spread and therefore on disease control options.
To date, spatial studies aiming at identifying HPAI H5N1 risk factors have been undertaken in many countries where the disease was introduced such as Thailand and Vietnam 
, Korea 
, India and Bangladesh 
, Romania 
or Africa 
. Only three studies analyzed the distribution of HPAI H5N1 outbreaks in China 
. Of these, only the study by Fang et al. 
attempts to map the distribution of HPAI H5N1 risk. Whilst highly valuable given that it is the first analysis, the output predicts areas at high risk in ecological areas that would not support the maintenance and transmission of the virus such as in the extremely large desert regions of Inner Mongolia, Tibet and Xinjiang autonomous regions.
In our study, we reported different results for the analyses based on outbreak and risk-based surveillance data. The distribution of reported HPAI H5N1 outbreaks was found to be primarily associated with lowland regions with high human population and chicken density. In contrast, HPAIV H5N1 presence detected through risk-based surveillance activities was found to be associated with regions with high waterfowl densities and were covered by high proportions of surface water. This result is very interesting since it may be a reflection of differences in HPAIV H5N1 pathogenicity between chickens and ducks, combined with environmental and host population conditions supporting virus spread and clinical disease outbreak occurrence as distinct from clinically silent virus persistence.
HPAI H5N1 is far more pathogenic in chickens than in ducks 
, though there is also evidence of significant variability in virulence at the species level 
. In the absence of control or prevention measures, the spread of HPAIV H5N1 and occurrence of clinical disease outbreaks is facilitated in regions where the density of chickens is particularly high, especially in intensive and industrial conditions where high numbers of animals are together facilitating transmission. Such regions are encountered in the north-eastern part of China, where the low cost of grain feed production and a fast-rising demand for poultry meat has supported the rapid development of intensive chicken production. Those intensive poultry production systems invest significant resources in disease prevention measures, and will apply mass-vaccination of their flocks, thereby preventing HPAIV H5N1 spread within and between farms. However, it is likely given the exceptionally high density of chickens and farms that occasional, albeit rare, lapses in vaccination coverage result in a small number of outbreaks. The human population density risk factor can be interpreted as a proxy of several epidemiological processes that are more likely to occur in highly-populated areas, such as a higher likelihood of outbreak detection and higher possibilities of HPAIV H5N1 transmission through trade and farming-related activities.
In contrast, long-term persistence of HPAIV H5N1 can only be possible if the virus can circulate without being detected or reported. Domestic ducks have been shown to be able to excrete large amounts of virus whilst remaining apparently healthy 
. Regions rich in domestic waterfowl are hence more prone to long-term persistence of HPAIV H5N1. This can be further exacerbated in geographical areas with an abundance of surface water. Permanent water bodies, rivers, rice paddy fields and canals are the habitat of wild and domestic ducks. One may speculate that water facilitates the transmission between hosts without direct contact through the faecal excretion of the virus, its persistence in the water, and the oral infection of other susceptible hosts sharing the same pond or downstream canal or river. The results indicating that HPAIV H5N1 presence detected through risk-based surveillance is associated with areas that have high waterfowl densities and a high proportion of surface water allows thus a straightforward interpretation. Associations between HPAIV H5N1 and domestic duck density had already been identified in other countries 
. However, no difference between outbreaks and clinically-silent infections was made in these earlier studies, which indeed becomes essential when analyzing HPAIV H5N1 distribution in the context of mass-vaccination such as in China.
Interestingly, the farming and cultural practices encountered in these regions were already described by Shortridge 28 years ago as an avian influenza breeding ground 
. Among others, Southern China still hosts a massive duck population raised on ponds and rice fields, facilitating frequent faecal-oral transmission of multiple influenza subtypes leading to a year-round and inter-epidemic occurrence of influenza viruses. Historically, agricultural practices in China have developed from the need to feed the people as efficiently as possible, using all available resources, and with little recourse to modern farming methods. Domestic ducks were first moved from rivers to cultivated rice fields at the start of the Qing dynasty in the middle of the 17th century 
to help protect the growing rice from pests. This practice reduces farmers' dependence on chemical insecticides, herbicides, fertilizers and mechanical farming aids and provides a close association between bird, water, rice and people. Ducks raised in ponds are also an important feature in the villages and communities of China, especially in southern China and coastal areas including the waterways of the Pearl River delta which are ideal for rice and fish farming 
. Furthermore, southern China has always been the focus of influenza experts' attention and often been referred as a hypothetical epicenter of AI pandemic strains. The foundation of this concept was originally raised by Webster et al. 
and supported by the wide variety of influenza virus subtypes discovered in Southern China during decades 
. More specifically, the distribution of HPAIV H5N1 risk of persistence inferred from the risk-based virological surveillance data and using the logistic and BRT models is similar and highlights different levels of risk according to the following ecological regions (, bottom; see Figure S2 in Text S1
for a map of the zones):
Zone I) In Southern China a large potential zone of virus persistence extends from south of the Yangtze River. This area hosts the vast majority of the Chinese duck population comprising birds for meat or egg production. This area can be subdivided into three areas: I-a) an area which extends from the provinces of Jiangsu, Anhui, Hubei, Jiangxi, Hunan, Guangxi autonomous region down to Guangdong province. This might be one of the most important ecological zones where key epidemiological drivers for emergence, persistence and spread are present, including a huge reservoir population, traditional farming system, a high animal and human population density, some major wild bird congregation sites such as the Poyang lake located in Jiangxi province and an important North – South gradient of poultry trade which crosses this region. This supports the hypothesis of a wider and slightly displaced epicenter of influenza viruses, not only concentrated around the Pearl River delta in Guangdong province but extending south of the Yangtze River and including provinces such as Jiangxi where internal segments of the 1996 geese HPAI H5N1 virus may have originated 
. I-b) A coastal area stretching from Jiangsu to Guangdong provinces with a risk hotspot in Guangdong province along the Pearl River delta. This strip of coastal land also hosts the typical duck pond system where the risk of infection and disease is present. I-c) Few isolated areas within this geographical zone displaying an increased risk located in Yunnan, Guangxi autonomous region, Guizhou, Sichuan and Chongqing provinces which have experienced either outbreaks in the past (Guangxi autonomous region and Yunnan) or only reported viral circulation (Sichuan and Chongqing provinces).
Zone II) A vast geographical area in the West and North, displaying radically different geography, socio-economic and animal production features and characterized by scattered and isolated spots of higher predicted risk. This includes specifically southern Tibet autonomous region and scattered areas in the North and South of Xinjiang autonomous region where sporadic outbreaks have occurred in the past.
Zone III) In the North-East of the Yangtze River, a region where the contribution to disease persistence seems fairly limited while localized areas at higher risk of outbreaks encroach regions of intensive production where the disease could rapidly spread in case of virus introduction and breach in biosecurity. This region extends from Shangdong into Liaoning, Jilin and Helongjiang provinces. These provinces are characterized by denser human population and large-scale commercial poultry production, and were predicted as high risk based on the reported clinical disease outbreak data ( top). In these regions of North-Eastern China, chicken production and marketing systems are intensifying and concentrating in response to economic growth and urbanization. Substantial numbers of poultry are now processed at large-scale slaughterhouses in this region, while the majority of poultry are still sold through live poultry markets in the South of the country. In the colder north-eastern provinces water birds are also housed and kept more intensively.
The apparent persistence of HPAIV H5N1 in those regions has two main implications. First, given the possible presence of silent infection involving an extremely high population of domestic waterfowl, eradication of the virus through massive vaccination appears extremely difficult, although it has successfully reduced the number of outbreaks. Vaccination has been one major component of the government policy to curb the spread of the disease and reduce the incidence of outbreaks of clinical disease and of transmission of infection. China uses more vaccine against avian influenza than any other country and Chinese veterinary authorities base much of their control and preventive strategy around vaccination 
. More than 13 billion doses of AI vaccine have been used each year since 2007 
and the objectives of the national strategy are to reach a 100% vaccination coverage for the national poultry population and to ensure an effective immune response (defined as sero-conversion in bird with titres >4Log2 when measured by HI test) in more than 70% of the nationwide poultry population all year round. In this study, we also analyzed the post-vaccination surveillance data collected at provincial level since January 2007 and found that provinces where clinical HPAI H5N1 outbreaks had been reported or HPAIV H5N1 detected had a lower level of post-vaccination seropositivity, confirming that increased protection does indeed result in lower disease outbreak or infection risks but would require an approach better targeted at identified high risk areas to drastically reduce the viral load in the environment.
Second, the different regions of China are not independent and are possibly epidemiologically linked through poultry trade and likely also through wild bird migrations. High production-demand discrepancies lead to long-distance trade of poultry products (e.g. chicken from the north exported to southern provinces, or duck meat exported from the south to the north). In addition, areas such as the Poyang lake, where a large population of domestic waterfowl is raised in close proximity to thousands of over-wintering wild waterfowl, could favour the transmission between wild and domestic waterfowl and lead to long-distance transmission of the virus. As a consequence, the persistence of HPAIV H5N1 in some particular regions may influence the chances of introduction into other more distant regions. For instance, the wild bird 2.2 clade which was associated with the origin of the Qinghai lake epidemic in 2005 in West China was responsible for a major outbreak during the same year in domestic poultry in Liaoning province, in the north eastern part of the country. Likewise, the 2006 Shanxi strain also grouped into the clade 7 cluster present in North-Central China has been found in Jiangsu province in South Eastern China. There is a complex pattern of links that exists between these different ecological regions that offers hiatuses for viruses to escape their reservoir areas and invade others.
Continued efforts pursued by the Ministry of Agriculture, its affiliated research centers and local veterinary authorities to strengthen the HPAI national surveillance program and its control strategy have resulted in a steady decrease in the number of outbreaks reported since 2004 and a better understanding of HPAIV H5N1 infection distribution in space, time and within traditional marketing systems known as live bird markets.
However, national surveillance programs have also demonstrated that HPAIV H5N1 continues to circulate in poultry on a regular basis. Since 2007, an apparent increase in virus detection is believed to represent the result of increased and intensive efforts made by the Ministry of Agriculture to detect the virus through targeted risk-based surveillance activities at live bird markets and high-risk farms in a context of massive vaccination efforts that could potentially mask the clinical expression of the disease within a large population of immunized birds.
Although the epidemiology of HPAIV H5N1 in China does not seem to present radically different features compared with neighboring countries also affected by the disease, it remains unique in terms of the abundance of reservoir species both domestic or wild, providing ample opportunities for a sustained and rapid evolution of the virus and requiring intensive virus monitoring for pandemic preparedness matters. While revisiting the concept of epicenter for pandemic strains of avian origin, the results of this study represent major improvements over previous efforts in mapping the risk of HPAI H5N1 in two main aspects.
First, it allows identifying several risk factors of animal, environmental and anthropogenic nature, with clear biological and epidemiological interpretation. Second, the bootstrapped statistical modeling allows us to robustly estimate the predictive power of our model, but also to map the uncertainty that goes with our predictions (Figure S1 in Text S1
), which is useful information for an applied use of these maps.
Combining innovative modeling techniques with data of improved quality and integrating measures of infection persistence, our results have broad fundamental implications in a country where understanding of the ecology of influenza viruses, although of utmost importance for pandemic preparedness purposes, has remained until now mostly speculative.
Finally, the analyses presented in this study may be improved in the future by several complementary approaches. First, the potential transmission through trade patterns and bird migration should be more comprehensively assessed. An increasing amount of data are being collected on both aspects, and this will ultimately contribute to better understanding of how areas of high potential for HPAIV H5N1 persistence may be connected to each others. Second, the results could be further integrated into an Asia-wide improved understanding of HPAIV H5N1 distribution models, benefiting from several studies that have been undertaken in neighboring countries. Third, information on true negatives obtained through the national surveillance programme would reduce the risk of including false-negatives in the analyses, and provide higher resolution estimates of the relative importance of risk factors.