1 EXAMINING THE DISTRIBUTION OF FRANCISELLA TULARENSIS THE CAUSATIVE AGENT OF TULAREMIA, IN UKRAINE USING ECOLOGICAL NICHE MODELING By JAKE MICHAEL HIGHTOWER A THESIS PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE UNIVERSITY OF FLORIDA 2012
2 2012 Jake Michael Hightower
3 To my mother and father for their unrelenting support in all facets of my life
4 ACKNOWLEDGMENTS This project was funded by the United States Defense Threat Reduction Agency (DTRA) through the Cooperative Biological Engagement Program under the Cooperative Biological Research Project UP 2. F unding was administered by the Joint Universi ty Partnership managed by the University of New Mexico. I thank my family for their compassion and encouragement throughout my entire life, especially throughout my entire academic career I am indebted to the individuals of the Spatial Epidemiology and Ec ology Research Lab at the University of Florida, including Ian Kracalik, Jocelyn Mullins, and Lillian Morris who helped guide me simultaneously as great friends and professionals. I am truly grateful for the members of the Geography Department at the Unive rsity of Florida, including Mike Falkner for helping me develop as a young professional and providing invaluable friendship during the entirety of my graduate school process. I thank my advisor Dr. Jason Blackburn for his commitment and dedication since th e very inception of my graduate career. He has provided me access to a world of science I would have otherwise never encountered and pushed me to strive for excellence at a degree by which others might call a dream. I also thank the members of my committee Dr. Pete Waylen and Dr. Liang Mao for their persistent support and care in the development of my thesis. I would also like to thank all of my colleagues encountered as part of the research in Ukraine The laboratory team at CSES lead by Dr. Nataliya Vyday ko co developed the historical data set used in this thesis and provide d insights into the disease system. The field team of Dr s Gregory Glass, Doug Goodin, Jason Richardson, Mikeljon Nikolich and Ivan Rusev provided much insight into this thesis and trained me on the
5 field sampling for this project Dr. Rusev also provided much of the background information on the history of Tularemia in Ukraine. Furthermore, I thank Mary Gutt ieri and Nataliya Mykhaylovska o f Southern Research Institute the staff at Black & Veatch Kyiv, who made many of the trips to Ukraine not only immensely enjoyable, but also possible.
6 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ .. 4 LIST OF TABLES ................................ ................................ ................................ ............ 8 LIST OF FIGURES ................................ ................................ ................................ .......... 9 LIST OF ABBREV IATIONS ................................ ................................ ........................... 10 ABSTRACT ................................ ................................ ................................ ................... 11 CHAPTER 1 OVERVIEW ................................ ................................ ................................ ................ 13 Tularemia ................................ ................................ ................................ ................ 14 Ukraine ................................ ................................ ................................ ................... 16 Ecological Niche Mod eling ................................ ................................ ...................... 17 Maxent ................................ ................................ ................................ .................... 20 Thesis Objectives ................................ ................................ ................................ ... 21 2 SPATIO TEMPORAL CLUSTER ANALYSIS OF FRANCISELLA TULARENSIS ISOLATES IN UKRAI NE ................................ ................................ ......................... 29 Introduction ................................ ................................ ................................ ............. 29 Materials and Methods ................................ ................................ ............................ 34 Francisella tularensis Data Set Development ................................ ................... 34 Function Development ................................ ................................ .... 34 Jacquez K Nearest Neighbors Analysis Development ................................ ..... 35 Spatial Scan Statistic Development ................................ ................................ .. 36 Results ................................ ................................ ................................ .................... 37 Function ................................ ................................ .......................... 37 Jacquez k Nearest Neighbor ................................ ................................ ............ 38 Spatial Scan Statistic ................................ ................................ ........................ 38 Discussion ................................ ................................ ................................ .............. 39 3 PREDICTING THE POTENTIAL DISTRIBUTION OF FRANCISELLA TULARENSIS IN UKRAINE USING ECOLOGICAL NICHE MODELING ............... 47 Introduction ................................ ................................ ................................ ............. 47 Materials and Methods ................................ ................................ ............................ 52 Franci sella Tularensis Data Set Development ................................ .................. 52 Rodent Host Distribution ................................ ................................ ................... 53 Tick Vector Distribution ................................ ................................ .................... 54
7 Environmental Layers ................................ ................................ ....................... 55 Maximum Entropy Modeling ................................ ................................ ............. 55 Model Development ................................ ................................ ......................... 57 Results ................................ ................................ ................................ .................... 59 Maxent Models ................................ ................................ ................................ 59 European Proje cted Models ................................ ................................ ............. 60 Discussion ................................ ................................ ................................ .............. 61 4 CONCLUSION AND FUTURE RESEARCH ................................ .............................. 75 LIST OF REFERENCES ................................ ................................ ............................... 78 BIOGRAPHIC AL SKETCH ................................ ................................ ............................ 87
8 LIST OF TABLES Table page 2 1 Results of the Jacquez k nearest neighbor analysis ................................ ........... 43 2 2 SaTScan results of space time permutation analysis of Francisella tularensis .. 46 3 1 A list of environmental variables that were tested ................................ ............... 67
9 LIST OF FIGURES Figure page 1 1 The Triangle of Human Ecology described by Meade and Earickson. ............... 23 1 2 Ukrainian topography map. ................................ ................................ ................ 24 1 3 Historical data set distribution.. ................................ ................................ ........... 25 1 4 Temporal distribution of Francisella tularensis isolates from Ukraine. ................ 26 1 5 Spatial distribution of small mammal and tick species. ................................ ....... 27 1 6 Spatial distribution of Francisella tularensis isolate locations.. ........................... 28 2 1 Francisella tularensis isolates. .......................... 42 2 2 Space time clust ers of all Francisella tularensis isolates. ................................ ... 44 2 3 Space time clusters of all Francisella tularensis isolates. ................................ ... 45 3 1 Comparison of mammal and tick collection geography ................................ ...... 66 3 2 D and I statistic results ................................ ................................ ..... 68 3 3 Results of the identity tests between each of the models ................................ ... 69 3 4 A comparison of global database collections. ................................ ..................... 70 3 5 Side by side comparison of ecological niche models created by Maxent ........... 71 3 6 Side by side comparison of ecological niche models created by Maxent ........... 72 3 7 Side by side comparison of ecological niche models created by Maxent .......... 73 3 8 Side by side comparison of ecological niche models created by Maxent ........... 74
10 LIST OF ABBREVIATION S AUC Area under the ROC curve BioClim Bioclimatic CFU Colony Forming Unit CSES Central Sanitary Epidemiological Station EFSA European Food Safety Authority ENM Ecological Niche Model GADM Database of Global Administrative Areas GBIF Global Biodiversity Information Facility GIS Geographic Information Systems HNH Host Niche Hypothesis JKNN Jacques K Nearest Neighbor Analysis KOMOCHUM Special Commission for the Prevention of and Fight against Plague LPS Lipopolysaccharide Maxent Maximum Entropy MLVA Multilocus VNTR Analysis NDVI Normalized Difference Vegetation Index PNH Pathogen Niche Hypothesis ROC Receiver Operating Characteristic Subsp. Subspecies TALA Trypanosomiasis and Land Use in Africa VNTR Variable Number Tandem Repeat WHO World Health Organization
11 Abstract of Thesis Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Master of Science EXAMINING THE DISTRIBUTION OF FRANCISELLA TULARENSIS THE CAUSATIVE AGENT OF TULAREMIA, I N UKRAINE USING ECOLOGICAL NICHE MODELING B y Jake Michael Hightower December 2012 Chair: Jason Blackburn Major: Geography Francisella tularensis is a gram negative, zoonotic pathogen which causes the infectious disease tularemia. F rancisella tularensis is found over a wide geography throughout most of the Northern hemisphere and in particular Ukraine. The bacteria is a highly infectious, Category A select agent that remains the target of several preventative studies because of the possibility of bioterrorism. This study sought to examine an hist orical dataset of F. tularensis isolates for spatio temporal clustering within Ukraine T he data were also used to create ecological niche models (ENM) of confirmed pathogen isolates in primary vector and host spec ies, with the use of the Maximum Entropy (Maxent) model ing algorithm Environmental layers used in ENM predictions were derived from Trypanosomiasis and Land Use in Africa (TALA) and altitude variables, aggregated to a 0.1 decimal degree resolution across the study region. Both overlapping and non overlapping cluster windows in SaTScan generally indicate d pathogen persistence over decades or longer, with the exception of two clusters that may have been the result of localized short term events, possibly in dicative of epidemics Ten individual Maxent experiments were conducted. A
12 single model of all isolates represented the extent of the pathogen in Ukraine A composite model of the four most abundant species, Mircrotus arvalis, Apodemus agrarius, Dermancen tor reticulatus, Ixodes ricinus, two mammalian hosts and two tick vectors, respectively, and four individual models of each. Each of the four models was Maxent experiments consistently predicted similar distributions across central to western Ukraine for each of the six data partitions. Niche similarity metrics indicated a high degree of niche overlap between each of the four species and the other two models, with the mamm al models having higher degrees of similarity than ticks. These results suggest that F. tularensis has a unique ecological niche and is not the summation of host ranges. This supports th e ism in Ukraine. based on F. tularensis positive samples from Ukraine, an additional four models, one for each species was developed for Europe and projected onto Ukrain e. Generally, each species modeled showed a wider distribution across Ukraine than F. tularnesis. These modeling results provide improved basemaps for understanding areas that support F. tularensis and should inform risk mapping and public health efforts.
13 CHAPTER 1 OVERVIEW Medical geographers often refer to the triangle of human ecology to describe the ure 1 1). At the vertices lie health [ 1 ] Population describes the biological aspects of human beings with traits like their genetics, gender, and age. Each of these traits influences the susceptibility of a human being to disease. While population describes the intrinsic biological properties of a person, their habitat describes the natural, social, and built environment about them. Malaria resistance, for example, can be conferred by the genetic predisposition of sickle cell anemia. Likewise, an individual can attempt to protect themselves from m alaria by changing their built environment, using insecticide treated bed nets to prevent transmission. The last vertex of the triangle is behavior which is governed by a have to be formed with each of these vertices in mind and their potential influence on a mathematical models, we can further define the sho uld not be thought of as independent, but constantly interacting forces. Pathogen mapping and modeling help identify under what conditions the tips of these vertices interact, e.g. does the human built environment overlap with the pathogen environment? Wit h the principles of human ecology taken into account, the first step is to understand the pathogen of interest.
14 Tularemia The disease tularemia maintains a vast geography that includes most of the Northern hemisphere, with regular reporting from the United States, Sweden, Finland, Russia, and much of Eastern Europe [ 2 ] Tularemia is found over a wide geography across the globe, limited to primaril y northern latitudes [ 2 3 ] Tularemia is a highly infectious, bacterial disease caused by the organi sm Francisella tularensis [ 4 ] This zoonotic pathogen has been isolated from a diverse group of mammals including livestock and wildlife [ 5 ] Th e pathogen is vector borne, transmitted by arthropod s including ticks biting flies, and mosquitoes, to a susceptible host [ 6 ] ( Petersen, Mead et al. 2009 ) Few infectious diseases have a comparable breadth of host s and transmission route s [ 7 ] The bacterium also survives in species of fish, invertebrates, amphibians, and surveyed in soil samples [ 8 ] The bacteria primarily engage in enzootic (sylvatic) transmission cycles between arthropod vectors and small mammal hosts. However, incidental h uman infection does occur by arthropod bite, cutaneous infection (e.g. handling infected animal tissue), ingestion of contam inated food or water, and inhalation of aeros olized bacteria [ 9 ] Tularemia has been categorized by the select agents program ( www.selectagents.gov ) as a Category A biological agent, the most dangerous agen rates and has the potential for major public health impact 3) might cause public panic [ 9 10 ] There is limited evidence to support human to human transmission of tularemia and human cases of tularemia can be treated with antibiotic drugs [ 11 ] Research efforts to monitor the bacterium outside of a clinical setting are primari ly focused on sylvatic transmission.
15 The integration of ecology to the study of infectious agents assists in the prediction of disease occurrence as well as the indirect consequences of an outbreak to local communities [ 12 ] The relationship between a disease and the environment can be a complex system of interactions including temperature, precipitation, altitude, soil pH, nutrient saturation, etc. Tularemia is widely accepted as one of the most dispersed diseases throughout the northern hemisphere with a complicated epizootiolog y at a global scale. Despite the breadth of potential global reservoirs of tularemia, regional disease cycles tend to be limited to far fewer key hosts and vectors [ 5 ] Current research is needed to clarify local, regional, and global ecological influences as well as their respective interactions. Particular emphasis in research is needed in areas of Eastern Europe and Western Asia which continue to display outbreaks of tularemi a. Many of these developing nations cannot afford the surveillance capabilities of their first world counterparts. Research dedicated to the ecological perspective of tularemia in these countries increases the efficiency of disease surveillance and prevent ion systems. With these systems in place, further development of phylogenetic analysis may yield even more insights on geographic dependencies of unique substrains. A recent study in Georgia used a Multi locus Variable Number Tandem Repeat (VNTR) Analysis (MLVA) and a S ingle N ucleotide P olymorphism (SNP) analysis to determine a unique phylogenetic relationship and a novel lineage of F. tularensis subsp. h olarctica [ 13 14 ] Such research furthers the understanding of infectious agents and the environmental dependen cies which may govern their persistence, survivability, and aptitude to become an outbreak.
16 Ukraine The impetus for the current disease monitoring programs originated largely from the need to facilitate medical aid to injured troops, monitoring the health of immigrants, or suppressing the magnitude of epidemics, as was the case in the early stages of the [ 15 ] Yet halfway around the world at the same time in the tsarist Russian Empire, was the beginning of a true epidemiological station : the Special Commission for the Prevention of and Fight against Plague (KOMOCHUM) [ 16 ] This research institute was aimed at the study of plague as caused by Yersinia pestis and included secure laboratories as well as outpost field labs in areas experiencing freque nt epizootics [ 16 ] This KOMOCHUM would begin developing anti plague laboratories, medical observation posts, vaccine productions, and quarantine checkpoints over the next 100 y ears until the 1917 Bolshevik Revolution and ensuing Civil War that decimated the anti plague system [ 17 ] However, under the Soviet regime, the anti plague system would be re kindled and massively grown to include over 100 facilities in over 11 republics of the Soviet Union [ 18 ] This expansion of strategically loca ted disease monitoring facilities not only reacted to plague outbreaks but also other bacterial and viral diseases, increased disease surveillance, and research/ training activities [ 18 ] The Soviet powered anti plague system would continue to grow massively and with the advent of war, would expand to include the most sophisticated biological warfare program in the world [ 19 ] The collapse and dissolution of the Soviet Union in 1991 led to the independence of several previously Soviet states, including Ukraine. Despite the still operation al anti plague stations, operational funding was severely decreased and maintenance of these stations became exceedingly difficult. The operation of epidemiological stations with the
17 capacity of effective disease surveillance and containment are crucial steps for disease response. The landscape of Ukraine is dominated by ferti le plains and rivers with a temperate climate (Figure 1 2) leading to an agricultural sector which is the third largest grain exporter in the world in 2009 2010 [ 20 ] Many Ukrainians maintain rural lifestyles, so the danger of incidental zoonotic disease in humans is ever present. Quantitative statistical analyses are becoming easier to operate and maintain as technical a dvances continue into the 21 st century. Despite the turmoil that the Ukrainian public health infrastructure has experienced, pathogen collections have continued and repositories of long historical records still exist. These historical data sets can be used with current spatio temporal software to examine the occurrence of disease clustering throughout the country (Figure s 1 3 1 4, 1 5 and 1 6 ) Furthermore, with the addition of background environmental variables derived from satellite data and past writte n records, it is possible to generate dynamic models to predict the distribution of specific infectious diseases. Applications of global information systems (GIS) and disease modeling are vital to improving the surveillance capabilities of the current publ ic health administration. Ecological Niche Modeling Several techniques can be used to quantify the relationship between infectious diseases, their hosts, and their environments. This study sought to employ ecological niche models (ENM), which use environme ntal variable data and species presence data in order to predict a species defined niche. The ecological signatures that promote pathogen survival and spread are poorly understood in Ukraine. Modeling the niche of vector and host species may serve as a pro xy to understand the spatial distributional qualities of the bacteria itself. Before we explore different mathematical models that
18 derive ENMs, we must define a niche. The environmental niche was first coined by Grinnell [ 21 ] who stated the dependence of an individual to their various physical structures and adaptations as the ultimate associational niche. The niche continued to gain momentum in the scientific community and was further redefined by Elton (1927), Gause (1934), and Hutchinson (1957). Among thes e fundamental and a realized niche. The fundamental niche describes the range of conditions a species would thrive under without inhibition by either biotic or abiotic factors, while the realized niche takes into consideration environmental pressures, spec ies predation, available resources, and other selective pressures which define the observed species niche [ 22 ] distribution of the species based on the ecological relationships of environmental vari ables In the case of zoonotic diseases, ENMs can be used to predict the potential geographic distribution of pathogens while identifying related ecological conditions and environmental variables [ 23 ] Creating an ENM of host, reservoir, or vector species can provide accurate presence/absence distribution These could form the basis of improved risk maps. On ce we can identify the several key research questions can be explored such as : where do diseased populations exist relative to uninfected, what could be causing these discrepancies, what environments appear to support the pathogen or disease, is there a shift in the dispersion of infected species over time, etc. Maher et al. (2010) also recognized that, in the case of plague transmission, there existed two possible hypotheses that might direct the spatial dependence of disease: a
19 host niche hypothesis (HNH) and a pathogen niche hypothesis (PNH). In short, the HNH proposes that the distribution of a given disease is dependent on the complex interaction of host and/ or vector ranges. The PNH describes the distribution of a disease as dependent on its own respective niche based on ecological parameters. In the Maher et al study, ENMs were generated which predicted the distribution of 8 different mammal species known to harbor Yersinia pestis the causative agent of plague. Their rese arch sought to identify if a HNH or a PNH aptly describes the proposed ecological distribution of the species and disease. For the HNH to be validated, it must be observed that both (1) plague infection in host species are not ecologically distinct from th e overall ecology of the host and (2) each plague infected host distribution has a unique ecological profile. Conversely, the PNH would predict that (1) plague infection and host ecology hold unique ecologies and (2) similar infected host distribution prof iles. The authors concluded that while a HNH clearly establishes a link between host and disease profile, a PNH is much more daunting. If the hosts are not predominantly responsible for the ecological profile of a disease then external factors like vector species and environmental variables need to be examined closely for their impact. In order to begin predicting the ecological niche of a species, it s presence data must be collected and confirmed While the observation of a species implicitly defines pres ence, the non detection of a species does not imply absence [ 24 ] The lack of presence during a sampling technique may be due to seasonal migrations, recent epidemics, anthropomorphic influences, improper sampling, etc. More appropriately, ecologists will refer instead to the detection (presence) and non detection (absence) of
20 a species obtained by sampling Over large spatial scales, data sources (e.g. museum records, field surveys, and historical archives) are often biased toward presence data and may misrepresent the absence of a species [ 25 ] Several modeling techniques have been derived which rely only on presence data and will generate pseudo absence points A variety of methods have been created to generate pseudo absence data, allowing for predictive presence/ absence models [ 26 28 ] Maxe nt The Maximum Entropy (Maxent) modeling of species geographic distributions was developed as a tool to create ecological niche models, especially under the circumstances when species absence data are lacking or completely withheld. A variety of research h as been performed utilizing the ENMs developed by Maxent in order to better understand disease cycles. Larson et al. (2008) used Maxent and the Genetic Algorithm for Rule set Prediction ( GARP ) to generate ENMs of three mosquito species, two of which may ha rbor West Nile Virus (WNV). Maxent models were useful in the comparison of WNV human incidence. Fischer et al (2011) incorporated present climate suitability and future climate scenarios to predict the geographic distribution of phlebotomine sandflies whi ch may transmit leishmaniasis. Recently, Flory et al. (2012) investigated white noise syndrome (WNS), an emerging disease that has been decimating North American bats. WNS is caused by the cold growing fungus Geomyces destructans and the relationship to the high mortality of North American bats is still precarious. The use of Maxent to develop ENMs allowed for a better understanding of which environmental conditions are most important for disease management. Maxent m odeling creat es an ENM which is estimated based on the target distribution of a species with the least variation [ 29 ] To be specific, Maxent models are
21 built to analyze a geographic space as a set of grid cells under which a set of points, representing species locality data, have been recorded. Environmen tal variables, like elevation, temperature, and precipitation are then interpolated into said grid cells to be used for predicting likely distributions of the locality data. A maximum entropy model uses iterative scaling algorithms posed by Darroch and Rat cliff, random field variants of Pietra, and machine learning logistic regression as developed by Collins, Schapire, and Singer [ 30 32 ] The sequential implementation of these different algorithms allows for the creation of a species probability distribution that represents the set of constraints which are derived from the l ocality data, with a probability score between 0 and 1 [ 33 ] The output of this probability distribution is a choropleth map of the defined study region, displaying areas of highest probability in warmer colors. The output ENM of Maxent is versatile enough to be input into other GIS programs (like ArcGIS) for further analysis and data management. To measure the performance of a Maxent derived ENM, the program intrinsically performs a ROC plot and AUC score. Thesis Objectives Central to the goal of medical geography are questions regarding the underlying spatial components of disease. This study is of no exception and seeks to uncover the spatial dependence of F. tularensis to the ecology of different s pecies and the geography that contains the highest probability of containing the pathogen. The first part of the thesis is designed to answer questions of spatial dependence and clustering based on all historical records of F. tularensis. The next section shifts towards the ecology of primary host and vector species in these records. Modeling the niche of infected and uninfected species in Ukraine allows us to understand what impact these species have on limiting the spatial bo undary of the pathogen. This thesis addresses
22 several research questions: 1) D o the historical records of F. tularensis indicate pathogen persistence or clustered distribution ? 2) D o Maxent derived ENMs support a PNH or HNH for the distribution of F. tular ensis in Ukraine ? 3) D oes the distribution of uninfected host and vector species differ from those that are infected, based on a projected, European ENM ?
23 Figure 1 1. The Triangle of Human Ecology described by Meade and Earickson (2005; page 25 )
24 Figure 1 2. Ukrainian topography map. Elevation map of Ukraine with respective bordering countries inset, created with WorldClim ( http://www.worldclim.org/ ). Categorized land use provided by GlobCorine ( ESA 2010 and Universit Catholique de Louvain http://due.esrin.esa.int/globc orine/ ).
25 Figure 1 3 Historical data set distribution. Points represent the t emporal distribution of Francisella tularensis isolates used in this thesis for spatio temporal and ecological niche modeling experiments Western asterisk identifies Ukrainian city Lviv, while the northern asterisk represents the capital city of Kyiv.
26 Figure 1 4 Temporal distribution of F rancisella tularensis isolates from Ukraine adjusted by sample size at each location Colors indicate the timing of isolation and sample size is graduated.
27 Figure 1 5 Spatial distribution of small mammal and tick species with confirmed F rancisella tularensis used to construct spatio temporal and ecological niche modeling experiments
28 Figure 1 6 Spatial distribution of F rancisella tularensis isolate locations All isolates (color and gray symbols) were included in a single experiment representing all sources. Color symbols were excluded from those ecological niche model ing experiments designed the test the Maher et al. (2010) hypothesis that the pathogen is
29 CHAPTER 2 SPATIO TEMPORAL CLUSTER ANALYSIS OF FRANCISELLA TULARENSIS ISOLATES IN UKRAINE Introduction Francisella tularensis is a zoonotic, gram negative bacterium and the causative agent of the infectious disease tularemia [ 8 ] Human exposure to t he pathogen can occur through several pathways including arthropod bites, ingestion of contaminated food products or liquids, inhalation of aerosolized bacteri a and h andling infected animals [ 34 ] As a vector borne pathogen F. tularensis may be transmitted through many arthropod vectors including fleas, lice, midges, bed bugs, ticks, mosquitoes, and flies with ticks among the most common vectors [ 35 ] Tularemia is a debilitating disease which may initially present with nonspecific flu like symptoms, often leading to clinical misdiagnosis [ 36 ] There are four re cognized subspecies of F. tularensis including holarctica mediasiatica novicida and tularensis each with spatial variation in their global distribution [ 37 ] Indivi duals with no prior immunity to tularemia may become infected by F. tularensis subsp. tularensis by inhaling as few as 10 cfu, making it comparab l e to other infectious pathogen, such as Mycobacterium tuberculosis [ 2 37 ] Tularemia is classified as a Category A infectious agent the highest priority designation of the select agent list by the United States Select Agent program ( www.selectagents.gov ). Generally, tularemia remains a rare disease among humans [ 10 ] though this has not been true throughout its history, particularly in parts of the former Soviet Union ( Pollitzer 1963 ) Like many vector borne diseases, human beings act as incidental hosts in the transmission cycle of tularemia, often infected when they enter the environmental reservoir for the host species or encounter bacteremic vectors in these areas [ 23 ] Many countries throughout the Northern Hemisphere have reported
30 the isolation of the pathogen F. tularensi s [ 37 ] Despite a broad geographic distribution across the much of the Northern Hemisphere, there is need to understand the geographic extent of the pathogen at regional scales to improve surveillance and better delineate zones of risk for human infection. However, di sease surveillance and modeling of zoonotic pathogens can be prohibitively time consuming and expensive [ 38 ] To improve upon these efforts, it is imp ortant to understand the spatial properties of a pathogen to increase the efficiency of surveillance efforts. A useful method for describing the spatial structure of a dataset is to create distribution maps and spatial structure functions [ 39 ] Mapping the geography of disease at local, regional, and global scales are important for describing the ecology of disease. Geographic information systems (GIS) and related mapping tools are useful in describing the geography of disease S uch computer based systems can integrate spatially diverse data and analyze relevant spatial properties. GIS applied studies of disease have the benefit of interacting spatially referenced, pathogen data with the geographical distribution of interactive d ata layers (e.g. population density, climate, etc.) Several studies, for example, have used GIS to examine the geospatial properties of such infectious diseases as Lyme disease [ 40 43 ] tick borne encephalitis (TBE) [ 44 46 ] and malaria [ 47 49 ] The WHO epidemic and pandemic alert and response team recognizes the importance of tularemia surveillance and the collection of reliable case data. At the far edge of eastern Europe, Ukraine is in particular need of improved tularemia disease surveillance and management. Natural tularemia foci have been defined in Ukrainian historical surveillance efforts since the 1940s [ 50 ] After the collapse of the Soviet
31 Union, Ukraine and other former Soviet countries found themselves face d with an underequipped public health infrastructure. This included a lack of funding and/ or personnel for pathogen surveillance. Currently tularemia still causes significant health concern and is monitored by the Ukrainian Central San itary Epidemiological Station. Identifying spatial patterns of a pathogen improves our understanding of influential ecological processes. The distribution of a zoonotic pathogen is a complex system involving ecological and geographic factors. Observed spa tial patterns are the result of interactions between biological, climatic, topographic, and other types of variables. A simple but important spatial pattern is the presence of clustering or dispersion. The K function is useful here as a second ord er spatial point pattern analysis that can detect global spatial clustering and the scale of spatial dependence using point level [ 51 ] The K function has been used repeatedly to verify the presence of cl usters at a global level with infectious disease [ 52 54 ] Lentz et al (2011) study which sought to examine the spat ial patterns of a white band disease (WBD) outbreak in the U.S. Virgin Islands. White band disease is a poorly understood disease that infects coral reef systems. the researcher s were able to identify significant local clustering of WBD in the study area, which c ould then be used to inform sampling efforts to create improved disease spatial rela tionships of a pathogen dataset. An important component to historical data sets is the temporality by which they occur. Examining pathogen clusters in a spatiotemporal framework requires a function which interpolates distances between points in both space and time. The Jacquez k
32 nearest neighbor analysis (JKNN) creates two such test statistics J k and k These statistics test for space time clusterin g under the variable changes of nearest neighbors k [ 55 ] By defining a nearest neighbor value k we remove the necessity of user defined critical space and/or time thresholds that may introduce subjectivity [ 56 ] The test statistics are not based on linear models and should be sensitive to plausible non linear relationships between the dependency of adjacen t events in either space or time [ 5 6 ] The JKNN allows users to determine if the probability of pathogen presence inside each k nearest neighbor search window is unique from that of the entire study region as a whole [ 57 ] Defining clusters as actual intervals of time and space, rather than a function of nearest neighbors requires a different s pace time scan statistic such as spatial scan statistic [ 58 ] Briefly that analyzes the study region in both space and time in relation to other nearby points, examining for potential clustering This statistic has been employed in previous zoo notic infectious disease studies, to examine potential disease clusters in space time. Brownstein et al. (2002) examined the human risk of West Nile virus (WNV) in New York City. West Nile virus is transmitted by arthropod vector and five such mosquito spe cies are vector competent in New York City. Using the SaTScan spatial scan statistic, the study found significant clustering of WNV cases. Having non normally distributed case data informs public health surveillance efforts to concentrate on areas which ap pear to cluster. Such analyses can also lead to the formation of risk models, which examine to what degree does the probability of disease occurrence vary within a study region.
33 Since the collapse of the USSR, initiatives such as the US Russian Cooperatio n on Dangerous Pathogens have come together to encourage legitimate research and activities on dangerous pathogens [ 59 ] These cooperative efforts aid in improving the understanding of pathogens and thei r respective public health threat while reinforcing the ability to diagnose, treat, and prevent infectious disease. Currently tularemia still causes significant health concern and is monitored by the Ukrainian Central Sanitary Epidemiological Station. This study incorporates a historical data set of confirmed F. tularensis isolates collected in Ukraine from 1941 to 2008 in the Central Sanitary Epidemiological Station (CSES) bacterial archive. Isolate collection includes samples from mammals, ticks, humans, and water source. Both spatial and spatio temporal analyses can be performed on historical data sets to discover pathogen clustering in relation to the geography of the data [ 60 ] The re were three objective s to this study : 1) to perform a spatial analysis of F. tularensis isolate records in Ukraine in order to examine the potential extent of spatial clustering, 2) to create a space time analysis of F. tularensis isolates using a JKNN approach, for the purpose of determining at what k defined scale the data cluster, and 3) to use the spatial scan statistic to examine precisely over what spatial area and for what period of time the pathogen isolates cluster. By understanding to what degree spatially and spatiotemporally these historical datasets cluster, we are informed on the potentiality of pathogen persist ence as well as potential pathogen sampling bias. These analyses help us evaluate the utility of the data set and a preliminary understanding of the ecology surrounding F. tularensis in Ukraine.
34 Materials and Methods Francisella t ularensis Data Set Develop ment An occurrence data set of 3 393 samples and 1284 confirmed Francisella tularensis isolates were collected from 1941 2008 and provided by the CSES (Central Sanitary Epidemiological Station) bacterial archives in Kyiv Ukraine (Figure 1 2) These data i solates where then mapped to the nearest village location and processed as spatially unique points at 0.1 degree resolution using ArcGIS, ArcMap software version 10. This dataset describes the distribution of the F. tularensis from records that include pathogen positive hosts and vectors environmental samples, and human case data. Function Development The spatial distribution of Francisella tularensis isolates in Ukraine was analyzed function (ClusterSeer 2.3.26, BioMedware Ann Arbor, MI, USA). function examines the distribution of point data to a Monte Carlo simulation produced via Poisson point process. A function k(h) is formed based on the number of cas es located within a d istance h of each individual case. function is defined as R represents a region of interest in square kilometers n is the number of cases in region R while i and j represent two separate localities. The distance between an i and j locality is represented by d ij I h (d ij ) is a binary indicator function with a value of 1 when d ij h or 0 at all other values, and w ij is an edge correction factor which provides cond itio nal weights to any point i dependent on the proportion of its radius which truly resides within the study region R To determine if case events are clustering, the
35 estimated distribution of K (h) is compared to another function L(h) which represents a homogenous Poisson point process which is equal to the square root of K Since this pathogen dataset was aggregated to the nearest village a n important ould represent clustering of the pathogen or merely clustering at the village level. To overcome this issue, the Lentz et al (2011) approach was taken. In this effort, K values were defined for not only the pathogen isolates, but also for a distributional data of all Ukrainian villages. The Ukrainian village K values were then subtracted from the pathogen K values, to create the difference function is that it will identify the spatial scale by which clustering and/ or dispersion occurs in a given study region, but it does not reveal where in the landscape it occurs [ 61 ] Despite such a limitation, very useful information can be gleamed from understanding the spatial scale of pathogen clustering. Jacquez K Nearest Neighbors Analysis Development To examine the historical dataset in both space and time, t he JKNN test was employed with the use of Biomedware program ClusterSeer version 2.3.26 ( http://www.biomedware.com/?module=Page&sID=clusterseer ). The JKNN test is a measure for global clustering which formulates two test statistics; the cumulative test statistic J k and the k specific test statistic k The cumulative test statistic counts the number of case pairs that are within k defined nearest neighbors of one another in both space and time. W hen there exists space time interaction between case pairs, J k will be large. The k specific test measures the change in the number of space time neighbors as k increases by 1. The cumulative test statistic in ClusterSeer is defined as:
36 In this equation s is the spatial neare st neighbor component and t is the temporal nearest neighbor component. Both the spatial and temporal nearest neighbor measures are binary variables, 1 if case j is k nearest neighbor of case i i n space or time respectively and a zero in all other circumstances. The k speci fic test statistic measures space time interaction about k observed for the k 1 nearest neighbor s; k is assigned values that range from 1 to 10 [ 55 ] The null hypothesis under the nearest neighbor test is true only when the adjacency of events in nearest neighbor space is independent of its relationship in nearest neighbor time at the proposed significance levels. Using this statistic, we learn whether clustering occurs at all k nearest neighbors, or if a difference occurs between space time scales. Spatial Scan Statistic Development This dataset was further analyzed for the presence of clusters in space and time wi th the Ku lldorff spatial scan statistic [ 62 ] employed with the use of SaTScan TM version 9.0 ( www.satscan.org ) The space time scan statistic creates varying sized cylinders around each case, where the circular base examines the spatial component while the vertical cylinder height is used to evaluate temporal clustering. This cylindrical window is then varied in space and time throughout the study region so that by the end of the an alysis, all data has been analyzed as a potential cluster in both space and time A retrospective space time statistic was calculated for the Ukrainian historical F. tularensis isolate data which was sorted by year of isolate collection. The study period w as set in
37 year format, from 1941 to 2008. SaTScan was performed as a space time permutation probability model and time aggregated to 67 years, scanning for areas with high rates. One of the factors that had to be considered from SaTScan derived models, was the effect of parameter settings, such as the significance of allowing cluster s to overlap or not Chen et al. [ 63 ] suggest ed difficulty in determining optimal SaTScan settings To examine this experiments were run with both overlapping and non overlapping cluster radii The first experiment represents a spatial scan that does not allow for the cluster radius of any individual window to overlap with another. Experiment two allowed for window overlap so long as the centroid of any individual cluster was not included in the cluster radius of another window. The spatial scan statistic defines primary and secondary clusters. A primary cluster will have the highest likelihood ratio test statistic while the secondary clusters are ordered according to their respective ratios. The p value of the likelihood functions are produced via Monte Carlo simulations, comparing rates inside the cylinder to outside the cylinder Results Function A total of 999 Monte Carlo randomizations were performed s K function, where distance h was internally defined as 383.195 with 1084 total points examined. Figure 2 1 K function itself. K function are strongly indicative of global clustering at all interpoint distances Results indicate significant clustering above the confidence interval and dispersion of pathogen isolates below. At lower K values the data indicat es clustering while at higher K values the data supports dispe rsion.
38 Jacquez k Nearest Neighbor Jacquez k n earest n eighbor spatio temporal analysis was set to run from k = 1 to k =10 neighbors and simulated with a total of 999 Monte Carlo randomizations. The p value from the M on te Carlo simulations was 0.58 4 and t able 2 1 shows upper tail significance values for the change in J(k) The Jacquez k nearest neighbor analysis as well as the reported Bonferroni and Simes statistic fail to reject the null hypothesis that there is no evidence to support the claim of global spatial and temporal clustering in Ukraine at 1 to 10 k nearest neighbors Spatial Scan Statistic Both SaTScan space time analyses were very similar in the production of primary and secondary clusters throughout Ukraine. Figure 2 2 illustrates the output of experiment 1 which classified clusters without overlap, while Figure 2.3 is the graphical representation of experiment 2 which classified clusters with overlap. Experiment 1 displayed 4 secondary clusters with a mean radius of 85.45 km, while experiment 2 displayed 6 secondary clusters with a mean radius of 163.59 km. Both experiments exhibited identical primary clusters as well as three secondary clusters. All SaTScan runs and cluster types were significant with a P value of 0.001. Table 2 2 lists primary and secondary cluster types for both SaTScan runs. In the transition from non overlapping to overlapping cluster definitions, the centroid of cluster 5 in run 1 is nearly identical to cluster center 6 of run 2, however with a drastically in creased cluster radius. Clusters 4 and 5 of run 2 appear to be wholly unique respective to run 1.
39 Discussion Preliminary analysis of the substantial historical data set provided by the Ukrainian CSES was to establish the presence of spatial and/ or spatio temporal clustering. The function confirms the expectation that the data is indeed spatially clustered up to 40 km but with over 60 years of historical data, the next question was whether or not there existed space time clusters. The presence of spatially and temporally related clustering could be the result of many factors, but among them are reporting biases which this study aimed to avoid. Analyzing the isolate data through a JKNN reported no evidence to support space time clustering up to K =10 nearest neighbors at a p value of 0.05. However, the data set covers not only a large geographic area but also a long period of time ( 67 years of isolate recovery ) When the SaTScan analysis was set to prevent cluster membership overlap, SaTScan repo rted 1 primary cluster and 4 secondary clusters. The primary cluster in the northern region of Ukraine spans 7 years in over a 170 km radius and retains our primary concern of whether or not sampling bias occurred in a given area. Similarly, secondary clus ters 2 and 5 are possibly the result of an outbreak among small mammals. The last two secondary clusters 3 and 4 contain cumulatively 16 isolates, each over nearly a 12 year span. The spatial scan statistic was also run under the parameter that scanning wi ndows and cluster boundaries may overlap, which yielded a total of 7 clusters. Four of these clusters including the primary cluster, appear identical to the non overlapping SaTScan results. However, cluster number 5 ( non overlapping analysis ) and ( 6 overla pping analysis) have very similar cluster centers but a n apparently large increase in respective cluster radius. The primary cluster located in northern Ukraine for both SaTScan runs may also be the product of increased reporting effort, due to the proxim
40 Kyiv Runs 1 and 2 are dominated by clusters that report disease over the span of decades rather than short periods of time. Two clusters in run 1 take place within a 2 year time span, which is likely indicative of local epid emics. However, when overlapping space time windows are allowed, one of these outbreaks disappears and becomes a much larger cluster in both space and time. Also, there exist a large proportion and area of data that are non clustered in both runs. These re sults together likely indicate the persistence of F. tularensis throughout the environment, lacking intermittent or resurgent epidemics which would have likely been characterized by increased isolate collection Also, it is furthermore possible that some of the clusters and/ or aggregations of non clustered data may be the result of an anti plague station nearby. At the peak of plague system boasted over 100 facilities in 11 republics, with the capacity to react to react to several bacterial and viral diseases, as well as hosting research and training [ 18 ] Continued research on the distribution of individual host and vector species distribution would greatly increase our understanding of F. tularensis in Ukraine. A r ecent study in Hungary (a bordering country of Ukraine) used a historical dataset consisting of F. tularensis records to examine vector and host relationships in the spread of tularemia [ 64 ] At a genetic level, spatial modeling of F. tularensis substrains might yield important and distinct geographic patterns. Several studies have capitalized on this effort by focusing on the geographic distribution of subspecies and clade specific lineages of F. tularensis [ 65 67 ] These studies further the distinction between strain specificity which may aid in explaining variations in both local and regional disease
41 distributions. Isolation and molecular identification of F. tularensis subspecies in Ukraine would increase the accur acy of these space time models.
42 Figure 2 1 Difference Francisella tularensis isolates. K values were first defined for village level distribution data for Ukraine and subtracted from observed k values from pathogen datas et. Resulting observed values for all villages were subtracted from observed F rancisella tularensis isolate locations and the difference was plotted (Y axis) against distance steps (X axis; each step equals 20 kilometers) The r ed line represents the diffe rence in observed value s Dashed grey lines indicate the upper and lower bounds of the 90 % confidence interval (CI) for the overall Ukrainian village dataset The blue line is the line of expectation. When the red line is above the CI, isolate locations ar e significantly clustered. Below the CI locations are dispersed and within the CI isolate locations are spatially random.
43 Table 2 1 Results of the Jacquez k nearest neighbor analysis, where k represents the number of nearest neighbors, J(k) is the count of number of case pairs that are k nearest neighbors in space and time, and DJ(k) measures the change J(k) by increasing k by 1. Both P(k) and DP(k) represent p values of J(k) and DJ(k) respectively. Upper Tail Upper Tail k J(k) P(k) DJ(k) DP(k) 1 1 0.444 1 0.444 2 9 0.079 8 0.054 3 14 0.149 5 0.407 4 25 0.098 11 0.140 5 25 0.388 0 1.000 6 43 0.196 18 0.103 7 47 0.404 4 0.893 8 54 0.600 7 0.850 9 62 0.760 8 0.825 10 78 0.754 16 0.537
44 Figure 2 2 Space time clusters of all F rancisella tularensis isolates using the SaTScan space time permutation with a predefined maximum spatial cluster size set at 50% of the population at risk while also preventing geographic overlap of cluster boundaries. Circles represent the spatial extent of a given cluster with all cluster members colored identically. Primary cluster center is denoted by a star, while all other secondary clusters are represented by squares. Time periods indicate the duration of the cluster in y ears while the distance in km represents the radius of the cluster. Western asterisk identifies Ukrainian city Lviv, while the northern asterisk represents the capital city of Kyiv.
45 Figure 2 3 Space time clusters of all Francisella tularensis isolates using the SaTScan space time permutation with a predefined maximum spatial cluster size set at 50% of the population at risk while allowing geographic overlap of cluster boundaries. Circles represent the spatial extent of a given cluste r with all cluster members colored identically. Primary cluster center is denoted by a star, while all other secondary clusters are represented by squares. Time periods indicate the duration of the cluster in years while the distance in km represents the r adius of the cluster. Western asterisk identifies Ukrainian city Lviv, while the northern asterisk represents the capital city of Kyiv.
46 Table 2 2. SaTScan results of space time permutation analysis of Francisella tularensis isolates in Ukrai ne. Cluster Isolates in Cluster Radius Duration O/E P Value No Overlap P 109 171.88km 2001 2008 6.64 0.001 S 42 85.13km 1990 1991 12.62 0.001 S 6 55.48km 1962 1975 1.75 0.001 S 10 127.64km 1941 1953 57.31 0.001 S 13 73.55km 1998 1998 121.82 0.001 With Overlap P 109 171.88km 2001 2008 6.64 0.001 S 42 85.13km 1990 1991 12.62 0.001 S 6 55.48km 1962 1973 1.75 0.001 S 85 157.51km 1951 1961 6.54 0.001 S 22 211.77km 1962 1973 1.75 0.001 S 71 344.01km 1963 1973 1.65 0.001 S 10 127.64km 1941 1953 57.31 0.001
47 CHAPTER 3 PREDICTING THE POTEN TIAL DISTRIBUTION OF FRANCISELLA TULARENS IS IN UKRAINE USING ECOLOG ICAL NICHE MODELING Introduction Zoonotic diseases represent a significant public and veterinary health burden globally. Zoonoses are those diseases that primarily infect animals which may spillover into humans or vice versa [ 23 ] As examples, Crimean Congo hemorrhagic fever (CCHF), Lyme disease, and tick borne encephalitis (TBE) represent several common zoonoses. Management of these diseases is imperative and can be improved with na tional disease surveillance. However, disease surveillance is expensive and often difficult to execute [ 68 ] Due to the prohibitive costs of broadly implemented surveillance analytical techniques can be adopted to help target surveillance At the same time, public awareness can be informe d by better defining areas at risk and targeting education and surveillance campaigns where zoonotic risk is greatest [ 68 ] Specific to disease surveillance, many zoonotic pathogens are carried by animal hosts and associated arthropod vectors which maintain their own geography. Geospatial analyses allow investigators to examine the potential geography of hosts and vectors to better delineate potential risk zones [ 23 ] and understand t he ecology of disease transmission [ 69 ] Tularemia is an important bacterial zoonosis across much of the northern hemisphere caused by the intracellular organism Francisella tularensis [ 4 ] The disease is widespread across northern latitudes and frequently reported across Europe. Recent tularemia outbreaks have been reported in five countries (Bulgaria, France, Germany, Kosovo, and Sweden) in Euro pe from 2000 to 2006 [ 70 ] This wide geographic range also includes a diverse spectrum of mammalian hosts and associated vector species
48 that can act to maintain pathogen endemicity. Human to human transmission of tularemia has not been reported. Because of this, transmission studies should focus on the role of infectious animals, arthropod bites, aerosols, etc [ 71 ] This also suggests that surveillance or spatial modeling should focus on the distribution of hosts and vectors. Currently tularemia causes significant health concern and is monitored by the Ukrainian Central Sanitary Epidemiological Station. Ukraine is in particular need of improved tularemia disease surveillance and management. Ukrainian historical surveillance efforts have defined naturally occurring tularemia foci across Ukraine s ince the 1940s [ 50 ] otics followed by human epidemics [ 72 ] To curtail these infections, the Soviet government issued mandatory immunoprophylaxis from 1957 to 1989 [ 73 ] After the collapse of the Soviet Union, Ukraine and other former Soviet countries found themselves faced with an underequipped public health infra structure. This included a lack of funding for pathogen surveillance. Such initiatives as the US Russian Cooperation on Dangerous Pathogens have since come together to encourage research and activities on dangerous pathogens [ 59 ] These cooperative efforts aid in improving the understanding of pathogens and their respective public health threat while reinforcing the ability to diagnose, treat, and prevent infectious disease. It has long been established that tularemia undergoes a sylvatic transmission cycle, with arthropod tick vectors and small mammalian hosts [ 72 ] There is also potential for pathogen survival in water. While these sources serve as known reservoirs, characterizing tularemia transmission dynamics is difficult, but patterns are emerging [ 64 74 ] An important question is whether or not the landscape described from disease
49 report s represents the extent of pathogen persistence Landscape ecology often attempts to answer this question by int erpreting the environmental parameters that govern the spatial extent of a disease [ 75 ] The relationship between the environment and disease transmission ca n then be summar ized on a mapped surface [ 69 76 ] The goal of these maps is to identify relevant factors that influence the spatial limits of a pathogen Efforts have been initiated to better understand the relationship between the epidemiology of tularemia and the ecology that surrounds it. For example, Pikula et al. (2004) predicted the potential distribution of tularemia in the Czech Republic by examini ng 6 different environmental factors in relation to the pathogen. This research led to the discovery of two favorable territories in Southern Moravia an d Central Bohemia for tularemia It is also possible to distinguish the ecology associated with hosts of a disease. In order to identify areas of elevated tularemia risk in Missouri, Brown et al [ 77 ] collected tick specimens to create a vector distribution map. Using environmental data and tick distribution data, the researchers were able to develop a probability map of a vector for tularemia in the United States. Studies that locate areas of high risk for hosts and vectors of pathogens increase the efficiency of prevention initiatives and pathogen control efforts Ecological niche modeling approaches have been evaluated in similar scenarios and provided important epidemiological information which can help stage public health intervention strategies [ 78 81 ] ENMs employ pattern matching algorithms (e.g. Genetic Algorithm for Rule Set Prediction; GARP) [ 27 ] or probabilistic models, such as Maximum Entropy (Maxent) [ 82 83 ] or discriminant function analysis [ 84 ] to relate non
50 random bioclimatic conditions (e.g temperature, soil conditions, etc) to occurrence data on the target organism (here the host, vector or pathogen [ 68 ] For a larger review of niche modeling, readers are directed to Alexander et al. [ 23 85 ] In the case of zoonotic diseases, ENMs can be used to predict the potential geographic distribution of pathogens while identifying related ecological conditions and environmental variables [ 86 ] [ 23 ] As an example, Costa et al (2002) developed GARP based ENMs to predict the distribution of Triatoma brasiliensis (an important vector for Chagas disease) in northeastern Brazil. This study confirmed four unique and ecologically distinct populations of T. brasiliensis In a separate study, Nakazawa et al. (2010) developed ENMs to evaluate the geographic distribution of F. tularensis genetic l ineages in the United States. The results indicated both spatially unique regions and areas of overlap for the potential distribution of the A1 and A2 F. tularensis sub lineages. One of the advantages of niche modeling is the capacity to project onto areas with no occurrence data, either landscapes with no occurrence data in a contemporaneous period or future climate scenarios for the same landscape modeled [ 68 ] These study examples support the idea that ENMs provide valuable ecological information on both the pathogen and reservoir speci es. In order to aptly describe the interaction between a zoonotic pathogen and its respective host and vector species, this study adopted the Maher et al. (2010) framework. Maher and his colleagues proposed two mutually exclusive hypotheses that maintain t defines the distribution of pathogen isolates as directly mediated by the distribution of the host species (Fig 3.1). In such a circumstance, the actual geographic range would
51 be depend ent on the composition of several host ranges that comprise the entire the distribution of pathogen isolates is independent of host ranges. While overlap will occur betw een host species, the PNH supports the idea that the host distributions overlap an underlying pathogen distribution. The study by Maher et al also examined a zoonosis, plague, caused by the bacterium Yersinia pestis Comparable to F. tularensis Y. pestis is transmitted by the bite of small flea arthropods and commonly infects small mammal hosts, like the prairie dogs ( Cynomys spp. ) in the United States [ 87 ] Since the distribution of a pathogen infected host or vector may represent only a subset of the ibution, the study developed ENMs for both infected and non infected hosts. Predicted models of plague distribution in the United States predominantly supported a PNH as there was consistent similarity in plague infected host distributions but not overall host ranges. There were two main goals of this study. The first was to develop a series of ENMs with Maxent, to examine the relationship of multiple F. tularensis infected reservoir species with their environment in Ukraine. These ENMs, along with the nic he overlap studies, allow us to examine whether the distribution of tularemia infected hosts and/ or vectors support a HNH or PNH. The second goal of this study was to analyze whether or not the ENMs which predict tularemia infected species represent a sub set of a wider uninfected niche range. This study will potentially increase the efficiency of surveillance efforts, create targeted zones for epidemic preparedness, and add to the scientific community on the ecology and biodiversity of F. tularensis
52 Materials and Methods Francisella Tularensis Data Set Development An occurrence data set of 3,393 samples and 1,284 confirmed Francisella tularensis isolates was collected from 1941 2008 and provided by the CSES (Central Sanitary Epidemiological Station) b acterial archives in Kiev, Ukraine. isolates where then mapped to the nearest village location and processed as spatially unique points at 0.1 degree resolution using ArcGIS ArcMap software version 10. Large occurrence datasets can be used for the creation of ENMs. One of the advantages in creating ENMs is the ability to use presence only data to project on to areas were no samples have been collected. With anticipation of testing a PNH vs. HNH, the data set was further subset by the two most prevalent spec ies of both mammalian hosts and tick vectors. The se data were split into six total groups: The first group contained all of the F. tularensis isolates. The second group was limited to four primary hosts/ vectors of tularemia in the database, as determined made up of samples from two mammal species Apodemus agrarius and Microtus arvalis and two tick vectors Dermacentor reticulatus and Ixodes ricinus Both have been previously reported as significant in the transmis sion of tularemia in endemic foci [ 88 ] The remaining four data partitions repres ented each individual species that make up the composite model. These host and tick species were chosen solely on the basis of sample size in the historical database. This study does not analyze or compare the species in relation to vector competence, host success as a reservoir for F. tularensis or relationships between hosts and vectors. The next step in data set development was to delete any duplicate accounts of isolates recorded in the same geographic locations Ecological niche models are based
53 presen ce/absence and multiple occurrences per cell do not improve model performance. As point datasets are randomly divided into training and testing subsets, it is important to avoid selecting both a training and test point from the same cell, as this can over inflate accuracy metrics performed posthoc [ 89 ] Gross inspection was performed to develop a spatially unique occurrence data set for each data partition at the native 8 km resolution of the environmental data set (see below). Pathogen isolates were provided by the Central Sanitary Epidemiological Stat ion (CSES). This historical data set also included samples that were obtained from water sources, farm produce, and a few human cases, all of which could easily complicate a fully comprehensive predicted distribution model. Therefore the se data w ere examin ed for the largest sample sizes of isolates collected from mammals (hosts) and ticks (vectors) that represent the greatest proportion of the isolate collection. These primary hosts and vectors include rodent species Microtus arvalis (common vole) and Apode mus agrarius (striped field mouse), as well as tick species Dermacentor reticulatus (ornate cow tick) and Ixodes ricinus (castor bean tick). Each of these individual species maintains a diverse ecology, with a wide geographic range throughout Eurasia. Rode nt Host Distribution The rodent A. agrarius has a n extensive distribution and can be seen throughout much of Europe and Asia. The mouse can reach a length of 126 mm, with relatively small ears and eyes [ 90 ] A podemus agrarius has been implicated as a common host for both hantavirus and tularemia [ 74 91 ] Like A. agrarius the commo n vole M. arvalis is an abundant species that can be found in open grasslands over much of Eastern Europe and parts of West Asia [ 92 ] M icrotus arvalis also is a host for disease including hantavirus and tularemia [ 93 94 ] For the formation of a Maxent derived European
54 model which would later be projected onto Ukraine, a much larger rodent distribution was nee ded for both primary host species M arvalis and A agrarius Mammal records were acquired from the Global Biodiversity Information Facility (GBIF) via the GBIF data portal [ 95 ] The search for a large European data set for A. agrarius yielded 6,635 records with a total of 4,168 that were georeferen ced and used for analysis. In the case of host species M. arvalis a total of 12,276 records were collected and 5,991 were also applicable for analysis. These rodent distributions were then reformatted to be analyzed by Maxent. Tick Vector Distribution Also important to the transmission of tularemia are the tick species D. reticulatus and I. ricinus [ 72 ] Both of these tick species have been noted as important carriers to F. tularensis and often prey on small rodents, including both M. arvalis and A. agrarius [ 74 ] These tick species have been noted as important vectors of viral, bacterial, and parasitic diseases including tick borne encephalitis, Lyme borreliosis, tick borne lymphadenopathy, tularemia, bab esiosis microti, etc [ 96 ] I xodes ricinus can be found throughout most of Europe, as far north as parts of Sweden and to the South in regions of Spain, Portugal, Greece, and Turkey [ 97 ] Tick species D. reticulatus is also reported throughout Eu rope, with particularly high presence in France, Germany, and Hungary [ 97 ] As a method of reference to the entire data set, these four chose and vectors were also compiled into a composite model so that a comparison might be made between these individuals and the entire isolate data collection. Both primary host vectors D reticulatus and I ricinus were also used in a Europea n model with Maxent. The derivation of the vector data points was taken from the European Food Safety
55 Authority (EFSA) Panel on Animal Health and Welfare (AHAW) which discusses and reveals geographic distributions of tick borne infections and their vectors in Europe, as well as other regions of the Mediterranean Basin [ 97 ] In total, 1201 records were georeferenced for tick species analysis, with 1105 for species I. ricinus and 97 for D. reticulatus The se data w ere collected from the journal and formatted for Maxent data analysis. Environmental Layers Environmental covera ge sets were constructed largely from the 0.1 degree climatic data set (e.g. temperature and mean normalized difference vegetation index [NDVI]) from the Trypanosomiasis and Land Use in Africa (TALA) research group [ 98 ] Elevation was also evaluated by the use of the Global Land One km Base Elevation Project (GLOBE) database at 8km. [ 99 ] Data coverages were refined by the use of a moment correlation coefficient matrix to reveal any environmental data that were significantly correlated. Only one coverage was chosen if two or more coverages were significantly correlated. Maximum Entropy Modeling The Maximum Entropy (Maxent) modeling of species geographic distributions was utilized to create ecological niche models to predict areas where species absent data is lacking or completely withheld. The modeling process analyzes data with maximum entropy, which is to say an ENM is estimated based on the target distribution of a species with the least variation [ 29 ] To be specific, Maxent models are built analyzing a geographic space as a set of grid cells under which a set of points, representing species locality data, have been reco rded. Environmental variables, like elevation, temperature, and precipitation are then interpolated into the grid cells to be used for predicting likely
56 distributions of the locality data. A maximum entropy model employs iterative scaling algorithms posed by Darroch and Ratcliff in1972, random field variants of Pietra, and machine learning logistic regression as developed by Collins, Schapire, and Singer [ 30 32 ] The sequential implementation of these different algorithms allows for the creation of a species probability distribution that represents the set of constraints which are derived from the locality data, with a probability score between 0 and 1 [ 33 ] To measure the performance of a Maxent derived ENM, the program intrinsically performs a ROC plot and AUC score. Maxent version 3.3.3e, was used for all Maxent model performances in this analysis. For the first set of models, the historical records of all confirmed F. tularensis ere created that described the probability distribution of 1) all F. tularensis isolates, 2) a composite of four reservoir species ( A. agrarius M. arvalis D. reticulatus and I. ricinus ), and 3) of each reservoir species individually. In these models, en vironmental layers were clipped around the Ukrainian national border, provided by the database of global administrative areas (GADM, www.gadm.org ). Each of these ENMs examined the probability distribution of infected spe cies and as such, may represent a subset of the actual host or vector distribution. To examine this possibility, European ENMs were created to predict the distribution of each of the four reservoir species and then projected onto Ukraine. Data sets used he re were collected from EFSA and GBIF for tick and rodent collections respectively. European models also required environmental coverages that were clipped to a much larger geography, which included the spatial extent of the geolocated data points. This ove rlap allowed for the comparison of infected hosts and
57 vector species with an uninfected counterpart. All Maxent models were run with a Logistic output format and visualized with ArcGIS 10. Model Development A national boundary shapefile of Ukraine was acc essed and downloaded from the database of Global Administrative Areas (GADM version 1). A GIS database of confirmed F. tularensis isolates throughout Ukraine was refined for distributional analysis. Environmental layers were then fit to the Ukrainian admin istrative boundary in a rasterized dataset scaled to the 8km (~0.1 degree) resolution. Occurrence data sets were split into training and test data, where the training data w ere used to develop ENMs and the test data w ere used to assess the accuracy of the models. Maxent models were developed in an 80/20 split of occurrence data for training and external validation respectively. A maximum number of iterations was set at 500. Using the predetermined environmental variables, the first six models were created w hich examined the probability distribution of all F. tularensis isolates, a composite model of all infected reservoir species, host A. agrarius model, host M. arvalis model, tick vector D. reticulatus model, and tick vector I. ricinus model. Afterward, environmental coverages were clipped to Europe and the EFSA and GBIF occurrence data were used. Maxent settings on extrapolation and clamping were turned off and models were developed and projected onto Ukraine from Europe. Model accuracy for each ENM was assessed using the area under the curve (AUC) metric of the receiver operator characteristic (ROC) curve. AUC scores range from 0 to 1 and indicate the accuracy between omission and commission. A score of 0.5 is expected of a perfectly random prediction while a score of 1 indicates not only perfect predictive ability, but a model with zero omission.
58 After ENMs have been developed, many studies have been dedicated to examine the niche inter relatedne [ 100 103 ] While models were developed with the use of Maxent, the software ENMTools v1.3 was used as a comprehensive u tility for measuring niche overlap and hypothesis testing via identity functions [ 104 ] A measure of niche overlap can be seen as a comparison of inte rspecific niche separation with intraspecific niche breadth [ 105 ] This is why it is important set a limit to what defines niche overlap, as opposed to statistically identical D and I statistic is one method of measuring a nd comparing niche overlap, quantified into values that range from 0 (models contain zero overlap) to 1 (models are identical) [ 104 ] D and I statistic were employed to examine D statistic is defined as: D is a measure of similarity between Maxent generated ENMs. The ENM suitability scores are normalized to that they sum to 1. In this equation, the p values are probabilities assigned by the ENM for both species ( x or y ) to cell i A similarity statistic I was developed by Warren et al in 2007 as a more applicable per mutation of the Hellinger distance H [ 104 ]
59 While the D statistic is the same as 1 I statistic is 1 D, the I statistic ranges from zero (no overlap of predicted ENMs) to one (perfect overlap of ENMs). Both the D and I statistic are measured from the suitability scores ENMs. Next the identity test was performed as a comparative hypothesis test that examines whether or not two (or more) populations are statistically identical. The niche identity test creates a series o f user defined pseudoreplicate data points where each pseudoreplicate represents a randomized re distribution of georeferenced data points for each population, The niche identity hypothesis is rejected when the observed values for D or I is significantly l ess than the expected values from the pseudoreplicated data [ 106 ] ENMTools version 1.3 was employed to perform niche similarity tests between unique model runs. Results Maxent Models Environmental coverages used for all Maxent models are summarized in Table 3.1. All six Maxent models consistently predicted areas in northern and northwestern Ukraine. The geography predicted for Dermacentor reticulatus was the most limiting model AUC scores for all Francisella tularensis isolates was 0.823, while the composite model reached an AUC of 0.864. Individual models reached AUC scores of 0.858 for Apodemus agrarius 0.849 for Microtus arvalis 0.948 for Dermacentor reticulatus and 0.883 for Ixodes ricinus The geography of each isolate collection is displayed in figure 3.2. Each of the models failed to predict isolate presence along the Carpathian
60 mountains to the southwestern edge of Ukraine. The mideastern to eastern edge of Ukraine wa s very rarely predicted by any of the Maxent models. The Crimean south of Ukraine received patchy predictions, mostly about the south and southeastern regions. D score for each host and vector species in relation to a ll F. tularensis isolates or the composite model. All models scored a D score above 0.5. The I statistic scores for each species when compared to all F. tularensis isolates are also shown. The I statistic scores were consistently high for each of the four species. The composite model held slightly higher I F. tularensis M. arvalis Results of the identity tests between each of the models are shown in Figure 3.5. The recorded similarity score was 0 .85 between all F. tularensis isolates and the composite model. Apodemus agrarius and the composite model received a similarity score of 0.74. Comparison between Microtus arvalis and the composite model demonstrated a similarity score of 0.77. The highest similarity score of an individual species versus the composite model was Dermacentor reticulatus which had a score of 0.80. The similarity score for Ixodes ricinus and the composite model was 0.74. European Projected Models European Maxent models were deri ved from sets of host and vector data whose individual geography can be seen in Figure 3.5. Projected modes varied dependent on the host or vector species. The comparison between Ukrainian and European projected Maxent models for A. agrarius are shown in F igure 3.6. Model predictions differed in their designation of absent or present by 22.22%. European projected and Ukrainian models for M. arvalis where the most different in terms of raw percentage difference (35.52%) in landscape designation (Figure 3.7). The projected distribution of M. arvalis
61 based of European collections predicted a small 0.26% of Ukraine as present for the host. Figure 3.8 displays the model outputs for both European projected and Ukrainian models of D. reticulatus These models diffe red by 35.26%, where the projected ENM predicted a much broader host distribution about the middle to western region of Ukraine, as well as parts of Crimea. Figure 3.9 compares European projected and Ukrainian models of I. ricinus with a raw difference in predicted landscape of 30.08%. Areas in Crimea, the southwest, and the northeast were predicted in larger quantities by the projected models. Discussion Examination of the spatial relationships of tularemia in Ukraine is useful for the improvement of di sease surveillance efforts. Mapping pathogen isolates and creating predictive ecological niche models allows for a better understanding of the ecology that drives the pathogen. In the case of zoonotic diseases, the roles of both vectors and hosts of diseas e that may lead to spillover in humans becomes an important question as well. To answer this question ecological niche models were created which describe the environmental parameters and associated factors that impact the presence of a pathogen in the envi ronment. Using the Maher et al (2010) framework, this study sought to answer whether or not the distribution of pathogen isolates was driven by a predominantly PNH or a HNH. y of pathogen samples dependent on the entirety of host and vector distributions, rather than a subset or limited region. The pathogen niche hypothesis would however describe the underlying pathogen niche as independent of the overall host ranges. While th e two most certainly overlap, a pathogen niche is only a subset of an entire host distribution.
62 Our Maxent models demonstrated a large similarity between all F. tularensis isolates and the composite model which was limited to only four total species (two t ick D and the I statistic indicate the composite model as a sufficient indicator to the broader geography of F. tularensis isolates. The tick vector D. reticulatus consistently reached the lowest similarity scores to the composite model, which was our first indicator that the distribution of pathogen isolates may follow a PNH rather than an HNH. Results of the identity test however indicate that despite the differen ces in geography between individual species and the composite model, there was not a significant difference in population composition. The ENMs of Ukrainian pathogen isolates represent predicted distribution maps of infected species. While modeling a dise ased host and vector implicitly provides also represent a subset of the greater, uninfected species distribution. In order to fully capture the host and vector ecology with respect to the distribu tion of F. tularensis an uninfected species range was developed. The European models were created by Maxent from large global databases of our mammal and tick species, without relation to disease. These Maxent models were then projected onto Ukraine and c ompared with our original models. There was evidence for niche similarity between infected and uninfected species of A. agrarius based on both Ukrainian and European models. Unfortunately, previous research shows that niche overlap studies which compare S D or the I statistic are unreliable [ 106 ] Comparisons between European projected and Ukrainian derived models are completed through raw cumulative predictions. The mode ls for A. agrarius for example, differed only by 22.27% of predicted presence and absence localities. The
63 ranges of their predictions from an observational perspective are similar, except that the European projection model predicted more presence localiti es into the western region of Ukraine Unlike all other models, the European projected model predicted a very small fraction of Ukraine as present for M. arvalis (0.25%). European models for D. reticulatus predicted much more of the middle, southern, and s outhwestern portion of Ukraine as present. The European model predicted nearly double the proportion of Ukraine as present for the tick species, than did the Ukrainian model. Similar to D. reticulatus the European model for I. ricinus also predicted a lar ger geography than its Ukrainian model counterpart. The European model predicted nearly five times as much of the landscape as did the Ukrainian model. These included a much greater predicted presence in Crimea and the southwest as well as an overall incr ease in areas already predicted in the Ukrainian model. There were several limitations in this study which include the lack of absence data in Ukraine. While ENMs created by Maxent certainly aim to overcome the challenges of presence only data, absence da ta would certainly improve the accuracy of models. The limitation of presence only data does truncate our ability to interpret the ecological drivers in ENMs to only those which are linked to presence. The Ukrainian presence data w ere also limited in that it described only tularemia positive samples. In order to overcome the lack of pathogen free host and vector samples, the European collections were obtained, modeled, and projected on to Ukraine. Because these European samples did not occur in Ukraine, the re are intrinsic limitations to the accuracy of projection models. Projected models suffer penalties if they are projected well outside the geography of the original background data [ 107 ] Also, during the formation of the
64 identity test, suitability scores derived from Maxent models had to be thresholded in order to obtain binary presence/ absence values. The interpretation of such thresholded v alues requires caution as does any model based on binary predictions [ 106 ] Despite these limitations, valuable data w ere recorded and we were able to compare host and vector niches in relation to the pathogen. Based on the niche overlap tests and the raw overlap percentages of Ukrainian models and the European projected models, we support the PNH taken from Maher et al (201 0) that F. tularensis retains its own ecological niche independent of vector or host ranges. This knowledge is important for disease prevention systems because any modeling of these host or vector species will support F. tularensis surveillance since both the tick and rodent species are Once these environmental relationships between epidemiology and ecology of reservoir species has been established, further pathogen modeling at a molecular le vel can increase the accuracy of these predictive maps. Farlow et al. (2005) examined the spatial distribution of F. tularensis isolates based on subspecies and subpopulation. Not only did the team find spatially unique regions that describe A.I and A.II s ubpopulations of F. tularensis subs. tularensis but they also noted a positive association between tick vectors and isolate subpopulations. In the central and eastern United States (and California) the A.I isolates occurred primarily along with tick vecto rs A. americanum and D. variabilis The natural foci of the A.II isolates in the U.S. were closely associated with tick vectors D. andersoni and C. discalis The study also noted a relationship in the geography of different rabbit hosts with respect to F. tularensis isolates
65 The goal of future research should be to increase the predictive accuracy of these models and work to establish what enables F. tularensis to maintain its own distinct ecological niche. Local museum records with presence/ absence data recorded of these primary hosts and vectors would greatly improve this effort.
66 Figure 3 1 Comparison of mammal and tick collection geography. The figure has been split into four components which represent the isolate collection of A) Apodemus agrari us B) Microtus arvalis C) Dermacentor reticulatus and D) Ixodes ricinus
67 Table 3 1. Correlation, to be included in Maxent model building Variables with asterisk had a high moment o f correlation and were not included in model formation. A ratio and therefore has no units; pvs= proportion of variance in original signal; NDVI= normalized difference vegetation index; LST= land surface temperature Variable Variable description Units Source Included in models alt Elevation (m) Hijmans et al. (2005) X wd1014mx Maximum NDVI Hay et al. (2006) X wd1014mx Minimum NDVI X wd1014a0 Mean NDVI X wd1014a1 NDVI annual amplitude X wd1014a2 NDVI bi annual amplitude X wd1014a3 NDVI tri annual amplitude X wd1014da NDVI pvs (%) wd1014d1 NDVI by annual cycle (months) wd1014d2 NDVI pvs bi annual cycle (%) X wd1014d3 NDVI pvs tri annual cycle (%) X wd1014p1 NDVI phase of annual cycle (months) X wd1014p2 NDVI phase of bi annual cycle (months) X wd1014p3 NDVI phase of tri annual cycle (months) X wd1014vr NDVI variance (%) wd1007mx Maximum LST (C) X wd1007mn Minimum LST (C) X wd1007a0 Mean LST (C) wd1007a1 LST annual amplitude (C) X wd1007a2 LST bi annual amplitude (C) X wd1007a3 LST tri annual amplitude (C) X wd1007da LST pvs (%) wd1007d1 LST pvs annual cycle (%) X wd1007d2 LST pvs bi annual cycle (%) X wd1007d3 LST pvs tri annual cycle (%) X wd1007p1 LST phase of annual cycle (months) X wd1007p2 LST phase of bi annual cycle (months) X wd1007p3 LST phase of tri annual cycle (months) X wd1007vr LST variance (%)
68 Figure 3 D and I statistic results. Each of the four primary vector and host species were evaluated for a D and I statistic score in relation to all Francisella tularensis isolates and a composite model. The composite model is made up of all four primary vector isolate distribution.
69 Figure 3 3 Results of the identity tests between each of the models when compared to the composite model. The identity test measures the similarity score between ENMs with background distributions of similarity scores drawn from randomly pooled occurrences of the two species. Similarity scores for each species are shown as a black arrow and are A) All Francisella tularensis isolates 0.85, B) Apodemus agrarius 0.74 C) Microtus arvalis 0.77, D) Dermacentor reticul atus 0.80, and E) Ixodes ricinus 0.74. Dashed blue line represents the normal distribution.
70 Figure 3 4 A comparison of global database collections for each of the mammal and arthropod species. A) A podemus agrarius B) M icrotus arvalis C) D ermacentor reticulatus and D) I xodes ricinus Mammal collections were provided from the GBIF portal. Tick collections were summarized in the EFSA panel on tick borne infections.
71 Figure 3 5 Side by side comparison of ecological niche models created by Maxent of primary host Apodemus agrarius Model (A) represents ENMs created from Ukrainian isolate data while model (B) represents the European projected model. Maxent models outputs were limite d to the extent of the area predicted by the minimum training presence for the purposes of comparison. Grey indicates no probability of presence while red indicates a positive probability of presence. Model (C) is an overlay comparison of both Ukrainian a nd European models.
72 Figure 3 6 Side by side comparison of ecological niche models created by Maxent of primary host Microtus arvalis Model (A) represents ENMs created from Ukrainian isolate data while model (B) represents the European project ed model. Maxent models outputs were limited to the extent of the area predicted by the minimum training presence for the purposes of comparison. Grey indicates no probability of presence while red indicates a positive probability of presence. Model (C) i s an overlay comparison of both Ukrainian and European models.
73 Figure 3 7 Side by side comparison of ecological niche models created by Maxent of primary vector Dermacentor reticulatus Model (A) represents ENMs created from Ukrainian isolate data while model (B) represents the European projected model. Maxent models outputs were limited to the extent of the area predicted by the minimum training presence as a threshold for the purposes of comparison. Grey indicates no probability of presence while red indicates a positive probability of presence. Model (C) is an overlay comparison of both Ukrainian and European models.
74 Figure 3 8 Side by side comparison of ecological niche models created by Maxent of primary vector Ixodes ricinus Model (A) represents ENMs created from Ukrainian isolate data while model (B) represents the European projected model. Maxent models outputs were limited to the extent of the area predicted by the minimum training presence as a threshold for the purposes of comparison. Grey indicates no probability of presence while red indicates a positive probability of presence. Model (C) is an overlay comparison of both Ukrainian and European models.
75 CHAPTER 4 CONCLUSION AND FUTUR E RESEARCH The goal of this t hesis was to examine the spatial clustering of Francisella tularensis isolates from an archive of historical data at the Central Sanitary and Epidemiological Station (CSES) in Kyiv Ukraine. Francisella tularensis is the causative agent of tularemia, a zo onosis of small animals that can greatly impact humans. To better understand the ecology of the disease in Ukraine, we buil t ENMs of primary host and vector species from this collection to test the h yp othe sis of Maher et al. [ 108 ] that a pathogen either has its ecological niche or is a composite of the niche conditions of the hosts and/or vectors The spatial and spatio temporal analyses performed in chapter 2 indicate pathogen persistence throughout Ukraine, over the 67 year dataset. The er k values, the pathogen isolates were indeed clustered, but at larger values were significantly dispersed. This variation in clustering and dispersion of data may have been tied into the wide range of F. tularensis positive vector and host distributions. T he JKNN reported no significant space time clusters. In contrast, the spatial scan statistic identified both primary and secondary clusters in space and time. However, the majority of these clusters usually spanned decades. For clusters that occurred within a scanning window of one year, it is possible that these were due to increased case reporting during that year that may indicate large epidemics, which would have prompted increased surveillance in the rodent pop ulations One limitation to these data is the lack of absence reporting. Without knowing for sure how many samples were taken in a given year, it is hard infer increased disease reports or specific epidemics However, F. tularensis isolates did appear to p ersist across the Ukrainian landscape over the study period The persistence
76 of large spatial clu sters over relatively long time periods indicates that F. tularensis was isolated with frequency and consistency, not just in response to isolated epidemic typ e events This suggests the dataset is not biased to epidemic events and useful for studying the broader ecology of the pathogen in Ukraine. Ecological niche model ing experiments constructed in Maxent accurately predicted the potential distribution of F. tularensis using either vectors or hosts. These model ing results supported The pathogen niche was consistently independent of host ranges and not merely a subset of their distribution, which is consistent w ith the PNH. In the absence of local pathogen free host and vector data, a Maxent derived ENM was built on a European dataset and projected onto Ukraine. Generally, predictions of host and vector distributions projected from European experiments were broad er than pathogen positive ranges using the CSES database. In other words, each host or vector had a wider distribution than F. tular en sis. Coupled with the measures of niche similarity these projections further support the PNH. There are many potential avenues of future research, which first and foremost are concerned with the acquisition of novel datasets. Local museum records would greatly increase the modeling potential and accuracy of F. tularensis with verified presence and absence d ata of uninfected mammals and ticks. Farlow et al. (2005) also recognized the utility of modeling F. tularensis subsp. tularensis into two distinct subpopulations A.I and A.II. The geographic distribution of these subpopulations was unique within the Unite d States and correlated in different ways to local host and vector populations. A similar study could be approached in Ukraine that strain typed F. tularensis subs. holarctica
77 the dominant strain throughout Eastern Europe, in a geospatial analysis. As Meade and Earickson (2005) describe in the triangle of human ecology, these studies have predominantly focused on the natural habitat of pathogen reservoirs. Future work could also include other aspects of the triangle, including the built environment. Landscape classification and remote sensing could be used to create finer resolution predictive models. Human incidence of tularemia would also be very useful for risk mapping. These studies are useful for the improvement of disease surveillance on a public health front, as well as increasing our knowledge of disease ecology associated with F. tularensis While the USSR maintained one of the largest disease surveillance programs ever recorded, their dissolution has set back many Eastern European countries in their ability to properly analyze and manage disease systems. Technologies like GIS, remote sensing, and modeling algorithms are becoming increa singly prevalent in these regions and serve to improve upon an already recovering system. This thesis has begun the initial steps of examining F. tularensis in Ukraine and set the foundation for future research in Maxen t modeling of zoonotic diseases
78 LIST OF REFERENCES 1. Meade, M.S. and R.J. Earickson, Medical Geography, 2ne Ed. 2005. 2. Oyston, P.C.F., A. Sjstedt, and R.W. Titball, Tularaemia: bioterrorism defence renews interest in Francisella tularensis. Nature Reviews Microbiology, 2004. 2 (12): p. 967 978. 3. Johansson, A., et al., Worldwide genetic relationships among Francisella tularensis isolates determined by multiple locus variable number tandem repeat analysis. Journal of bacteriology, 2004. 186 (17): p. 5808 5818. 4. Dennis, D.T., et al., Tularemia as a biological weapon. JAMA: The Journal of the American Medical Association, 2001. 285 (21): p. 2763. 5. Petersen, J.M. and M.E. Schriefer, Tularemia: emergence/re emergence. Veterinary research, 2005. 36 (3): p. 455 467. 6. Cox, K.B., Tularemia and deer flies in the environs of Utah Lake, Utah. Great Basin Nat, 1965. 25 : p. 13 29. 7. Beran, G.W., Handbook of Zoonoses Second Edition ed. Section B: Viral1994: CRC Press. 608. 8. Keim, P., A. Johansson, and D.M. Wagner, Molecular epidemiology, evolution, and ecology of Francisella. ANNALS OF THE NEW YORK ACADEMY OF SCIENCES, 2007. 1105 (1): p. 30 66. 9. Engleberg, N.C.D., Victor; Dermody, Terence S., Schaechter's Mechanisms of Microbial Disease Fourth Edition ed2007: Lippincott Wi lliams & Wilkins. 762. 10. Charette, J.D., CDC CATEGORY EXPLANATION (A, B, C) OVERVIEW. Toxico terrorism: emergency response and clinical approach to chemical, biological, and radiological agents, 2007. 755 : p. 329. 11. Enderlin, G., et al., Streptomycin and alternative agents for the treatment of tularemia: review of the literature. Clinical infectious diseases, 1994. 19 (1): p. 42 47. 12. Ostfeld, R.S., F. Keesing, and V.T. Eviner, Infectious disease ecology: effects of ecosystems on disease and of dise ase on ecosystems 2010: Princeton University Press. 13. Gvantsa, C., et al., Phylogeography of Francisella tularensis subspecies holarctica from the country of Georgia. BMC Microbiology. 11
79 14. Chanturia, G., et al., Phylogeography of Francisella tularen sis subspecies holarctica from the country of Georgia. BMC microbiology, 2011. 11 (1): p. 139. 15. Rosen, G., A history of public health 1993: Johns Hopkins University Press. 16. Melikishvili, A., Genesis of the anti plague system: The tsarist period. Crit ical reviews in microbiology, 2006. 32 (1): p. 19 31. 17. Henze, C., Disease, Health Care and Government in Late Imperial Russia Vol. 72. 2011: Routledge. 18. Ouagrham Gormley, S.B., Growth of the Anti plague System during the Soviet Period. Critical rev iews in microbiology, 2006. 32 (1): p. 33 46. 19. Zilinskas, R.A., The anti plague system and the Soviet biological warfare program. Critical reviews in microbiology, 2006. 32 (1): p. 47 64. 20. UMG, U.M.G., Chamber Country Profile Ukraine 2012, The American Chamber of Commerce in Ukraine. p. 116. 21. Grinnell, J., The niche relationships of the California Thrasher. The Auk, 1917. 34 (4): p. 427 433. 22. Hutchinson, Concluding Remarks. Cold Spring Harb Symp Quant Biol, 1957. 22 : p. 415 427. 23. Alex ander, K.A., et al., Modeling of Wildlife Associated Zoonoses: Applications and Caveats. Vector Borne and Zoonotic Diseases, 2012. 24. MacKenzie, D.I., et al., Estimating site occupancy rates when detection probabilities are less than one. Ecology, 2002. 83 (8): p. 2248 2255. 25. Anderson, R.P., Real vs. artefactual absences in species distributions: tests for Oryzomys albigularis (Rodentia: Muridae) in Venezuela. Journal of Biogeography, 2003. 30 (4): p. 591 605. 26. Zaniewski, A.E., A. Lehmann, and J.M.C. Overton, Predicting species spatial distributions using presence only data: a case study of native New Zealand ferns. Ecological Modelling, 2002. 157 (2): p. 261 280. 27. Stockwell, D. and D. Peters, The GARP modelli ng system: problems and solutions to automated spatial prediction. International Journal of Geographical Information Science, 1999. 13 (2): p. 143 158.
80 28. Engler, R., A. Guisan, and L. Rechsteiner, An improved approach for predicting the distribution of r are and endangered species from occurrence and pseudo absence data. Journal of Applied Ecology, 2004. 41 (2): p. 263 274. 29. Phillips, S.J., M. Dudk, and R.E. Schapire. A maximum entropy approach to species distribution modeling 2004. ACM. 30. Darroch, J.N. and D. Ratcliff, Generalized iterative scaling for log linear models. The annals of mathematical statistics, 1972. 43 (5): p. 1470 1480. 31. Della Pietra, S., V. Della Pietra, and J. Lafferty, Inducing features of random fields. Pattern Analy sis and Machine Intelligence, IEEE Transactions on, 1997. 19 (4): p. 380 393. 32. Collins, M., R.E. Schapire, and Y. Singer, Logistic regression, AdaBoost and Bregman distances. Machine Learning, 2002. 48 (1): p. 253 285. 33. Phillips, S.J. and M. Dudk, M odeling of species distributions with Maxent: new extensions and a comprehensive evaluation. Ecography, 2008. 31 (2): p. 161 175. 34. Staples, J.E., et al., Epidemiologic and molecular analysis of human tularemia, United States, 1964 2004. Emerg Infect Dis 2006. 12 (7): p. 1113 1118. 35. Petersen, J.M., P.S. Mead, and M.E. Schriefer, Francisella tularensis: an arthropod borne pathogen. Veterinary research, 2009. 40 (2). 36. Switaj, K., et al., Tularaemia after tick exposure typical presentation of rare dis ease misdiagnosed as atypical presentation of common diseases: a case report. Cases journal, 2009. 2 37. Sjostedt, A., Tularemia: History, epidemiology, pathogen physiology, and clinical manifestations. Annals of the New York Academy of Sciences, 2007. 1 105 (Journal Article): p. 1 29. 38. Murphy, F.A., Emerging zoonoses. Emerging Infectious Diseases, 1998. 4 (3): p. 429. 39. Legendre, P. and M.J. Fortin, Spatial pattern and ecological analysis. Plant Ecology, 1989. 80 (2): p. 107 138. 40. Dister, S.W., et al., Landscape characterization of peridomestic risk for Lyme disease using satellite imagery. The American journal of tropical medicine and hygiene, 1997. 57 (6): p. 687.
81 41. Glass, G.E., et al., Environmental risk factors for Lyme disease identified with geographic information systems. American Journal of Public Health, 1995. 85 (7): p. 944 948. 42. Kitron, U., J.K. Bouseman, and C.J. Jones, Use of the ARC/INFO GIS to study the distribution of Lyme disease ticks in an Illinois count y. Preventive Veterinary Medicine, 1991. 11 (3): p. 243 248. 43. Kitron, U. and J.J. Kazmierczak, Spatial analysis of the distribution of Lyme disease in Wisconsin. American Journal of Epidemiology, 1997. 145 (6): p. 558 566. 44. Daniel, M., et al., Predic tive map of Ixodes ricinus high incidence habitats and a tick borne encephalitis risk assessment using satellite data. Experimental and Applied Acarology, 1998. 22 (7): p. 417 433. 45. Furlanello, C., et al., GIS and the Random Forest Predictor: Integratio n in R for Tick Borne Disease Risk Assessment 2003. 46. A tick borne encephalitis ceiling in Central Europe has moved upwards during the last 30 years: possible impact of global warming? International Journal of Medical Microbiolo gy Supplements, 2004. 293 : p. 48 54. 47. Beck, L.R., et al., Remote sensing as a landscape epidemiologic tool to identify villages at high risk for malaria transmission. The American journal of tropical medicine and hygiene, 1994. 51 (3): p. 271. 48. Omum bo, J., et al., Mapping malaria transmission intensity using geographical information systems GIS: an example from Kenya. Annals of Tropical Medicine and Parasitology, 1998. 92 (1): p. 7 21. 49. Srivastava, A., et al., GIS based malaria information managem ent system for urban malaria scheme in India. Computer methods and programs in biomedicine, 2003. 71 (1): p. 63 75. 50. Stupnitskaya, V.M.M., M.P.; Litvinenko, Ye. F.; Slesarenko, V.V.; Slesarenko, A.S.; Khizhinskaya, O.P.; Stepanova, I.A.; Buyalo, S.G., N atural Tularemia Foci On the Territory of the Ukrainian SSR. Journal of Microbiology, Epidemiology, and Immunology, 1964. 10 : p. 94 98. 51. Ripley, B.D., Modelling spatial patterns. Journal of the Royal Statistical Society. Series B (methodological), 1977 : p. 172 212. 52. Bastin, L., et al., Spatial aspects of MRSA epidemiology: a case study using stochastic simulation, kernel estimation and SaTScan. International Journal of Geographical Information Science, 2007. 21 (7): p. 811 836.
82 53. Kroeger, A., et al., Effective control of dengue vectors with curtains and water container covers treated with insecticide in Mexico and Venezuela: cluster randomised trials. Bmj, 2006. 332 (7552): p. 1247 1252. 54. O'Brien, D.J., et al., Spatial and tempo ral distribution of selected canine cancers in Michigan, USA, 1964 1994. Preventive veterinary medicine, 1999. 42 (1): p. 1 15. 55. Jacquez, G.M., A k nearest neighbour test for space time interaction. Statistics in medicine, 1996. 15 (18): p. 1935 1949. 5 6. Ward, M.P. and T.E. Carpenter, Analysis of time space clustering in veterinary epidemiology. Preventive veterinary medicine, 2000. 43 (4): p. 225 237. 57. Meliker, J.R., et al., Spatial cluster analysis of early stage breast cancer: a method for public health practice using cancer registry data. Cancer Causes and Control, 2009. 20 (7): p. 1061 1069. 58. Kulldorff, M., et al., A space time permutation scan statistic for disease outbreak detection. PLoS medicine, 2005. 2 (3): p. e59. 59. Howson, C.P., Cont rolling dangerous pathogens: A blueprint for US Russian cooperation. Conversion of Former BTW Facilities, 1998. 21 : p. 135. 60. Marshall, R.J., A review of methods for the statistical analysis of spatial patterns of disease. Journal of the Royal Statistic al Society. Series A (Statistics in Society), 1991: p. 421 441. 61. Gatrell, A.C., et al., Spatial point pattern analysis and its application in geographical epidemiology. Transactions of the Institute of British Geographers, 1996: p. 256 274. 62. Kulldo rff, M., A spatial scan statistic. Communications in Statistics: Theory and Methods, 1997. 26 : p. 1487 1496. 63. Chen, J., et al., Geovisual analytics to enhance spatial scan statistic interpretation: an analysis of US cervical cancer mortality. Internati onal Journal of Health Geographics, 2008. 7 (1): p. 57. 64. Gyuranecz, M., et al., Factors Influencing Emergence of Tularemia, Hungary, 1984 2010. Emerging Infectious Diseases, 2012. 18 (8): p. 1379. 65. Farlow, J., et al., Francisella tularensis in the United States. Emerg Infect Dis, 2005. 11 (12): p. 1835 1841.
83 66. Chanturia, G., et al., Phylogeography of Francisella tularensis subspecies holarctica from the country of Georgia. BMC Microbiology, 2011. 11 (1): p. 139. 67. Kugeler, K.J., et al., Molecular epidemiology of Francisella tularensis in the United States. Clinical Infectious Diseases, 2009. 48 (7): p. 863 870. 68. Blackburn, J., Integrating geographic information systems and ecological niche modeling into disease ecology: a case study of Bacillus anthracis in the United States and Mexico in Emerging and Endemic Pathogens: Advances in Surveillance, Detection, and Identification E.W.S. K.P. O'Connell, A. Sulakvelidze, L. Bakanidze, Editor 2010, Springer. p. 59 88. 69. Peterson, A.T., Eco logic niche modeling and spatial patterns of disease transmission. Emerging Infectious Diseases, 2006. 12 (12): p. 1822. 70. Vorou, R., V. Papavassiliou, and S. Tsiodras, Emerging zoonoses and vector borne infections affecting humans in Europe. Epidemiolog y and infection, 2007. 135 (8): p. 1231 1247. 71. Tarnvik, A.B., T.; Forsman, M.; Peterson, J.; Sjostedt, A.; Pedro, A.; Pearson, A.; Johansson, A.; Grunow, R.; Kosoy, M.; Krause, G.; Chu, M.; Elkins, K.; Nano, F.; Titbal, R.;, WHO Guidelines on Tularaemia A. Tarnvik, Editor 2007, WHO Press. 72. Pollitzer, R., A Review of Selected Problems of Tularemia in the Soviet Union D.o. Defense, Editor 1963. p. 115. 73. Shuliarenko, S.V., [An evaluation of the tularemia control work of epidemiological health stat ions]. Vrach Delo, 1991. 2 : p. 111 3. 74. Gyuranecz, M., et al., Investigation of the ecology of Francisella tularensis during an inter epizootic period. Vector Borne and Zoonotic Diseases, 2011. 11 (8): p. 1031 1035. 75. Blackburn, J.K., Integrating geog raphic information systems and ecological niche modeling into disease ecology: a case study of Bacillus anthracis in the United States and Mexico. Emerging and Endemic Pathogens, 2010: p. 59 88. 76. Waller, L.A., et al., Hierarchical spatio temporal mapping of disease rates. Journal of the American Statistical Association, 1997: p. 607 617. 77. Brown, H.E., et al., An acarologic survey and Amblyomma americanum distribution map with implications for tularemia risk in Missouri. The American journal of tropical medicine and hygiene, 2011. 84 (3): p. 411 419. 78. Nakazawa, Y., et al., Climate change effects on plague and tularemia in the United States. Vector Borne and Zoonotic Diseases, 2007. 7 (4): p. 529 540.
84 79. Blackburn, J.K., et al., Modeling the ge ographic distribution of Bacillus anthracis, the causative agent of anthrax disease, for the contiguous United States using predictive ecologic niche modeling. The American journal of tropical medicine and hygiene, 2007. 77 (6): p. 1103 1110. 80. Joyner, T .A., et al., Modeling the potential distribution of Bacillus anthracis under multiple climate change scenarios for Kazakhstan. PLoS One, 2010. 5 (3): p. e9596. 81. Fischer, D., S. Thomas, and C. Beierkuhnlein, Modelling climatic suitability and dispersal f or disease vectors: the example of a phlebotomine sandfly in Europe. Procedia Environmental Sciences, 2011. 7 : p. 164 169. 82. Phillips, S., R. Anderson, and R. Schapire, Maximum entropy modeling of species geographic distributions. Ecological Modelling, 2006. 190 (3 4): p. 231 259. 83. Phillips, S., M. Dudk, and R. Schapire. A maximum entropy approach to species distribution modeling 2004. ACM. 84. Rogers, D., Models for vectors and vector borne diseases. Advances in parasitology 2006. 62 : p. 1 35. 85. Alexander, K.A., J.K. Blackburn, and E.A. Frimpong, Trails and Tribulations Buffalo and Maslow's hammer. Frontiers in Ecology and the Environment, 2011. 9 (5): p. 302 303. 86. Blackburn, J., Evaluating the spatial ecology of anthr ax in North America: Examining epidemiological components across multiple geographic scales using a GIS based approach in Department of Geography and Anthropology 2006, Louisiana State University: Baton Rouge. p. 141. 87. Eskey, C.R. and V. Haas, Plague in the western part of the United States. Public Health Bulletin, 1940(254). 88. Hublek, Z. and J. Halouzka, Mosquitoes (Diptera: Culicidae), in contrast to ticks (Acari: Ixodidae), do not carry Francisella tularensis in a natural focus of tularemia in t he Czech Republic. Journal of medical entomology, 1997. 34 (6): p. 660 663. 89. Joyner, T., et al., Modeling the potential distribution of bacillus anthracis under multiple climate change scenarios for Kazakhstan. PloS one, 2010. 5 (3): p. e9596.
85 90. Karli k, F.A. Interactive Agricultural Ecologcal Atlas of Russia and Neighboring Countries. Economic Plants and their Diseases, Pests and Weeds 2008 9/24/12]; Available from: http://ww w.agroatlas.ru/en/content/pests/Apodemus_agrarius/ 91. Schmaljohn, C. and B. Hjelle, Hantaviruses: a global disease problem. Emerging Infectious Diseases, 1997. 3 (2): p. 95. 92. Haynes, S., M. Jaarola, and J.B. Searle, Phylogeography of the common vole (Microtus arvalis) with particular emphasis on the colonization of the Orkney archipelago. Molecular Ecology, 2003. 12 (4): p. 951 956. 93. Plyusnin, A., et al., Tula virus: a newly detected hantavirus carried by European common voles. Journal of virology, 1994. 68 (12): p. 7833 7839. 94. Gurycova, D., et al., Importance of surveillance of tularemia natural foci in the known endemic area of Central Europe, 1991 1997. Wiener klinische Wochenschrift, 2001. 113 (11 12): p. 433. 95. GBIF: p. Biodiversity occurr ence data published by: American Museum of Natural History, Arctos MVZ Mammal Catalog, Australian Museum provider for OZCAM, Banc de dades de biodiversitat de Catalunya, BeBIF Provider, California Academy of Sciences, CSIRO Australian National Wildlife Col lection, Direccin General de Medio Natural y Poltica Foresta, Estacion Biologica Donana, Facultad de Ciencias Naturales y Museo, Field Museum Mammals Collections, Finnish Museum of Natural History, Harvard University Museum of Comparative Zoology, James R. Slater Museum of Natural History, Japan National Science Museum, Jyvaskyla University Museum, Kokkola Nature Collection, Mammalia Collection at the Staatssammlung fr Anthropologie und Paloanatomie Mnchen, Michigan State University Museum, Museo Nacio nal de Ciencias Naturales, Museu de Cincies Naturals de Barcelona, Musum d'Histoire Naturelle de Bourges, Museum fr Naturkunde Berlin, National Museum of Natural History, National Museum of Natural History, Natural History Museum Rotterdam, Netherlands Biodiversity Information Facility, Polish Academy of Sciences, Royal Belgian Institute of Natural Science, Senckenberg Collection Mammalia SMF, Swedish Museum of Natural History, SysTax, Texas Cooperative Wildlife Collection, UCLA Dickey Collection, Univer sity Museum of Zoology Cambridge, University of Alaska Museum of the North, University of Alberta Museums, and University of Tartu Natural History Museum (Accessed through GBIF Data Portal, data.gbif.org, 2011 7 12). 96. Srter, T., Z. Szll, and I. Varga, Spatial distribution of Dermacentor reticulatus and Ixodes ricinus in Hungary: evidence for change? Veterinary parasitology, 2005. 128 (3): p. 347 351.
86 97. EFSA, European Food Safety Agency (EFSA) Panel on Animal Health and Welfare (AHAW); Scientific Opinion on Geographic Distribution of Tick borne Infections and their Vectors in Europe and the other Regions of the Mediterranean Basin. EFSA J, 2010. 8 : p. 1723. 98. Hay, S.I., et al., Global environmental data for mapping infectious disease distribution. Advances in parasitology, 2006. 62 : p. 37 77. 99. Hastings, D.A. and P. Dunbar, Development & assessment of the Global Land One km Base Elevation digital elevation model (GLOBE). Group, 1998. 4 : p. 6. 100. Knouft, J.H., et al., Phylogenetic analysis of the evolution of the niche in lizards of the Anolis sagrei group. Ecology, 2006. 87 (sp7): p. 29 38. 101. Wiens, J.J. and C.H. Graham, Niche conservatism: integrating evolution, ecology, and conservation biology. Annual rev iew of ecology, evolution, and systematics, 2005: p. 519 539. 102. Bagchi, S., S. Goyal, and K. Sankar, Niche relationships of an ungulate assemblage in a dry tropical forest. Journal of Mammalogy, 2003. 84 (3): p. 981 988. 103. Albrecht, M. and N. Gotell i, Spatial and temporal niche partitioning in grassland ants. Oecologia, 2001. 126 (1): p. 134 141. 104. Warren, D.L., R.E. Glor, and M. Turelli, Environmental niche equivalency versus conservatism: quantitative approaches to niche evolution. Evolution, 20 08. 62 (11): p. 2868 2883. 105. May, R.M., On the theory of niche overlap. Theoretical Population Biology, 1974. 5 (3): p. 297 332. 106. Warren, D.L., R.E. Glor, and M. Turelli, ENMTools: a toolbox for comparative studies of environmental niche models. Eco graphy, 2010. 33 (3): p. 607 611. 107. Elith, J., et al., A statistical explanation of MaxEnt for ecologists. Diversity and Distributions, 2011. 108. Maher, S.P., et al., Range wide determinants of plague distribution in North America. The American journa l of tropical medicine and hygiene, 2010. 83 (4): p. 736 742.
87 BIOGRAPHICAL SKETCH Jake Michael Hightower was born in 1987 in Oahu, Hawaii. He spent most of his life in Orlando, Florida and graduated from Trinity Preparatory School in May 2005. Hightower then attended the University of Tampa in Fall of 2005 where he was a member of several extracurricular activities, including UT Emergency Medical Services. spital for Children and served in many leadership roles, including Vice President of Theta Chi Fraternity, Director of Field Training in UT EMS, and student representative in Greek Standards Board. In April of 2009, Hightower graduated with a Bachelor of S cience in Biology. For the next year, he worked at Universal Studios, Orlando as a CPR/ AED instructor. In the summer of 2010 he began working in the Spatial Epidemiology and Ecology Research (SEER) Laboratory through the Department of Geography at the Uni versity of Florida. Hightower was then accepted into the Department of Geography at UF to pursue a the SEER lab in conjunction with the Emerging Pathogens Institute (EPI) Through research programs with the SEER laboratory, he collaborated on several projects including pathogen research in Ukraine. He received his received his Master of Science from the University of Florida in fall 2012.