Artificial Intelligence: A Promising Tool in Exploring the Phytomicrobiome in Managing Disease and Promoting Plant Health

There is increasing interest in harnessing the microbiome to improve cropping systems. With the availability of high-throughput and low-cost sequencing technologies, gathering microbiome data is becoming more routine. However, the analysis of microbiome data is challenged by the size and complexity of the data, and the incomplete nature of many microbiome databases. Further, to bring microbiome data value, it often needs to be analyzed in conjunction with other complex data that impact on crop health and disease management, such as plant genotype and environmental factors. Artificial intelligence (AI), boosted through deep learning (DL), has achieved significant breakthroughs and is a powerful tool for managing large complex datasets such as the interplay between the microbiome, crop plants, and their environment. In this review, we aim to provide readers with a brief introduction to AI techniques, and we introduce how AI has been applied to areas of microbiome sequencing taxonomy, the functional annotation for microbiome sequences, associating the microbiome community with host traits, designing synthetic communities, genomic selection, field phenotyping, and disease forecasting. At the end of this review, we proposed further efforts that are required to fully exploit the power of AI in studying phytomicrobiomes.

Uncertainty quantification for epidemiological forecasts of COVID-19 through combinations of model predictions

Scientific advice to the UK government throughout the COVID-19 pandemic has been informed by ensembles of epidemiological models provided by members of the Scientific Pandemic Influenza group on Modelling. Among other applications, the model ensembles have been used to forecast daily incidence, deaths and hospitalizations. The models differ in approach (e.g. deterministic or agent-based) and in assumptions made about the disease and population. These differences capture genuine uncertainty in the understanding of disease dynamics and in the choice of simplifying assumptions underpinning the model. Although analyses of multi-model ensembles can be logistically challenging when time-frames are short, accounting for structural uncertainty can improve accuracy and reduce the risk of over-confidence in predictions. In this study, we compare the performance of various ensemble methods to combine short-term (14-day) COVID-19 forecasts within the context of the pandemic response. We address practical issues around the availability of model predictions and make some initial proposals to address the shortcomings of standard methods in this challenging situation.

Adding Epidemiologically Important Meteorological Data to Peanut Rx, the Risk Assessment Framework for Spotted Wilt of Peanut

Management of disease affecting peanut in the southeastern United States has benefited from extensive field research identifying disease-associated risk factors since the 1990s. An assessment of risk factors associated with tomato spotted wilt (TSW), caused by tomato spotted wilt virus and spread exclusively by thrips, is available to growers through Peanut Rx, a tool developed to inform peanut management decisions. Peanut Rx provides an assessment of relative TSW risk as an index. The assessment provides information about the relative degree to which a field characterized by a specified suite of practices is at risk of crop loss caused by TSW. Loss results when infection occurs, and infection rates are determined, in part, by factors outside a grower's control, primarily the abundance of dispersing, viruliferous thrips. In this study, we incorporated meteorological variables useful for predicting thrips dispersal, increasing the robustness of the Peanut Rx framework in relation to variation in the weather. We used data from field experiments and a large grower survey to estimate the relationships between weather and TSW risk mediated by thrips vectors, and developed an addition to Peanut Rx that proved informative and easy to implement. The expected temporal occurrence of major thrips flights, as a function of heat and precipitation, was translated into the existing risk-point system of Peanut Rx. Results from the grower survey further demonstrated the validity of Peanut Rx for guiding growers' decisions to minimize risk of TSW.

Testing predictability of disease outbreaks with a simple model of pathogen biogeography

Predicting disease emergence and outbreak events is a critical task for public health professionals and epidemiologists. Advances in global disease surveillance are increasingly generating datasets that are worth more than their component parts for prediction-oriented work. Here, we use a trait-free approach which leverages information on the global community of human infectious diseases to predict the biogeography of pathogens through time. Our approach takes pairwise dissimilarities between countries' pathogen communities and pathogens' geographical distributions and uses these to predict country-pathogen associations. We compare the success rates of our model for predicting pathogen outbreak, emergence and re-emergence potential as a function of time (e.g. number of years between training and prediction), pathogen type (e.g. virus) and transmission mode (e.g. vector-borne). With only these simple predictors, our model successfully predicts basic network structure up to a decade into the future. We find that while outbreak and re-emergence potential are especially well captured by our simple model, prediction of emergence events remains more elusive, and sudden global emergences like an influenza pandemic are beyond the predictive capacity of the model. However, these stochastic pandemic events are unlikely to be predictable from such coarse data. Together, our model is able to use the information on the existing country-pathogen network to predict pathogen outbreaks fairly well, suggesting the importance in considering information on co-occurring pathogens in a more global view even to estimate outbreak events in a single location or for a single pathogen.

Variation in the Internal Transcribed Spacer Region of Phakopsora pachyrhizi and Implications for Molecular Diagnostic Assays

Phakopsora pachyrhizi, the causal agent of soybean rust (SBR), is a global threat to soybean production. Since the discovery of SBR in the continental United States, quantitative polymerase chain reaction assays based on the internal transcribed spacer (ITS) ribosomal DNA locus were established for its rapid detection. However, insufficient data were initially available to test assays against factors that could give rise to misidentification. This study aimed to reevaluate current assays for (i) the potential for false-positive detection caused by nontarget Phakopsora species and (ii) the potential for false-negative detection caused by intraspecific variation within the ITS locus of P. pachyrhizi. A large amount of intraspecific and intragenomic variation in ITS was detected, including the presence of polymorphic ITS copies within single leaf samples and within single rust sori. The diagnostic assays were not affected by polymorphisms in the ITS region; however, current assays are at risk of false positives when screened against other species of Phakopsora. This study raises caveats to the use of multicopy genes (e.g., ITS) in single-gene detection assays and discusses the pitfalls of inferences concerning the aerobiological pathways of disease spread made in the absence of an evaluation of intragenomic ITS heterogeneity.

Fine-grained dengue forecasting using telephone triage services

Thousands of lives are lost every year in developing countries for failing to detect epidemics early because of the lack of real-time disease surveillance data. We present results from a large-scale deployment of a telephone triage service as a basis for dengue forecasting in Pakistan. Our system uses statistical analysis of dengue-related phone calls to accurately forecast suspected dengue cases 2 to 3 weeks ahead of time at a subcity level (correlation of up to 0.93). Our system has been operational at scale in Pakistan for the past 3 years and has received more than 300,000 phone calls. The predictions from our system are widely disseminated to public health officials and form a critical part of active government strategies for dengue containment. Our work is the first to demonstrate, with significant empirical evidence, that an accurate, location-specific disease forecasting system can be built using analysis of call volume data from a public health hotline.

FLIRT-ing with Zika: A Web Application to Predict the Movement of Infected Travelers Validated Against the Current Zika Virus Epidemic

Introduction: Beginning in 2015, Zika virus rapidly spread throughout the Americas and has been linked to neurological and autoimmune diseases in adults and babies. Developing accurate tools to anticipate Zika spread is one of the first steps to mitigate further spread of the disease. When combined, air traffic data and network simulations can be used to create tools to predict where infectious disease may spread to and aid in the prevention of infectious diseases. Specific goals were to: 1) predict where travelers infected with the Zika Virus would arrive in the U.S.; and, 2) analyze and validate the open access web application’s (i.e., FLIRT) predictions using data collected after the prediction was made.

Method: FLIRT was built to predict the flow and likely destinations of infected travelers through the air travel network. FLIRT uses a database of flight schedules from over 800 airlines, and can display direct flight traffic and perform passenger simulations between selected airports. FLIRT was used to analyze flights departing from five selected airports in locations where sustained Zika Virus transmission was occurring. FLIRT’s predictions were validated against Zika cases arriving in the U.S. from selected airports during the selected time periods.  Kendall’s τ and Generalized Linear Models were computed for all permutations of FLIRT and case data to test the accuracy of FLIRT’s predictions.

Results: FLIRT was found to be predictive of the final destinations of infected travelers in the U.S. from areas with ongoing transmission of Zika in the Americas from 01 February 2016 - 01 to April 2016, and 11 January 2016 to 11 March 2016 time periods. MIA-FLL, JFK-EWR-LGA, and IAH were top ranked at-risk metro areas, and Florida, Texas and New York were top ranked states at-risk for the future time period analyzed (11 March 2016 - 11 June 2016). For the 11 January 2016 to 11 March 2016 time period, the region-aggregated model indicated 7.24 (95% CI 6.85 – 7.62) imported Zika cases per 100,000 passengers, and the state-aggregated model suggested 11.33 (95% CI 10.80 – 11.90) imported Zika cases per 100,000 passengers.

Discussion: The results from 01 February 2016 to 01 April 2016 and 11 January 2016 to 11 March 2016 time periods support that modeling air travel and passenger movement can be a powerful tool in predicting where infectious diseases will spread next. As FLIRT was shown to significantly predict distribution of Zika Virus cases in the past, there should be heightened biosurveillance and educational campaigns to medical service providers and the general public in these states, especially in the large metropolitan areas.  

Rift Valley Fever: An Emerging Mosquito-Borne Disease

Rift Valley fever (RVF), an emerging mosquito-borne zoonotic infectious viral disease caused by the RVF virus (RVFV) (Bunyaviridae: Phlebovirus), presents significant threats to global public health and agriculture in Africa and the Middle East. RVFV is listed as a select agent with significant potential for international spread and use in bioterrorism. RVFV has caused large, devastating periodic epizootics and epidemics in Africa over the past ∼60 years, with severe economic and nutritional impacts on humans from illness and livestock loss. In the past 15 years alone, RVFV caused tens of thousands of human cases, hundreds of human deaths, and more than 100,000 domestic animal deaths. Cattle, sheep, goats, and camels are particularly susceptible to RVF and serve as amplifying hosts for the virus. This review highlights recent research on RVF, focusing on vectors and their ecology, transmission dynamics, and use of environmental and climate data to predict disease outbreaks. Important directions for future research are also discussed.

Rodent reservoirs of future zoonotic diseases

The increasing frequency of zoonotic disease events underscores a need to develop forecasting tools toward a more preemptive approach to outbreak investigation. We apply machine learning to data describing the traits and zoonotic pathogen diversity of the most speciose group of mammals, the rodents, which also comprise a disproportionate number of zoonotic disease reservoirs. Our models predict reservoir status in this group with over 90% accuracy, identifying species with high probabilities of harboring undiscovered zoonotic pathogens based on trait profiles that may serve as rules of thumb to distinguish reservoirs from nonreservoir species. Key predictors of zoonotic reservoirs include biogeographical properties, such as range size, as well as intrinsic host traits associated with lifetime reproductive output. Predicted hotspots of novel rodent reservoir diversity occur in the Middle East and Central Asia and the Midwestern United States.

Spatial Big Data Analytics of Influenza Epidemic in Vellore, India

The study objective is to develop a big spatial data model to predict the epidemiological impact of influenza in Vellore, India. Large repositories of geospatial and health data provide vital statistics on surveillance and epidemiological metrics, and valuable insight into the spatiotemporal determinants of disease and health. The integration of these big data sources and analytics to assess risk factors and geospatial vulnerability can assist to develop effective prevention and control strategies for influenza epidemics and optimize allocation of limited public health resources. We used the spatial epidemiology data of the HIN1 epidemic collected at the National Informatics Center during 2009-2010 in Vellore. We developed an ecological niche model based on geographically weighted regression for predicting influenza epidemics in Vellore, India during 2013-2014. Data on rainfall, temperature, wind speed, humidity and population are included in the geographically weighted regression analysis. We inferred positive correlations for H1N1 influenza prevalence with rainfall and wind speed, and negative correlations for H1N1 influenza prevalence with temperature and humidity. We evaluated the results of the geographically weighted regression model in predicting the spatial distribution of the influenza epidemic during 2013-2014.

Predicting Fusarium head blight epidemics with boosted regression trees

Predicting major Fusarium head blight (FHB) epidemics allows for the judicious use of fungicides in suppressing disease development. Our objectives were to investigate the utility of boosted regression trees (BRTs) for predictive modeling of FHB epidemics in the United States, and to compare the predictive performances of the BRT models with those of logistic regression models we had developed previously. The data included 527 FHB observations from 15 states over 26 years. BRTs were fit to a training data set of 369 FHB observations, in which FHB epidemics were classified as either major (severity ≥ 10%) or non-major (severity < 10%), linked to a predictor matrix consisting of 350 weather-based variables and categorical variables for wheat type (spring or winter), presence or absence of corn residue, and cultivar resistance. Predictive performance was estimated on a test (holdout) data set consisting of the remaining 158 observations. BRTs had a misclassification rate of 0.23 on the test data, which was 31% lower than the average misclassification rate over 15 logistic regression models we had presented earlier. The strongest predictors were generally one of mean daily relative humidity, mean daily temperature, and the number of hours in which the temperature was between 9 and 30°C and relative humidity ≥ 90% simultaneously. Moreover, the predicted risk of major epidemics increased substantially when mean daily relative humidity rose above 70%, which is a lower threshold than previously modeled for most plant pathosystems. BRTs led to novel insights into the weather-epidemic relationship.

Adaptation of an Apple Sooty Blotch and Flyspeck Warning System for the Upper Midwest United States

A warning system for sooty blotch and flyspeck (SBFS) of apple, developed in the southeastern United States, uses cumulative hours of leaf wetness duration (LWD) to predict the timing of the first appearance of signs. In the Upper Midwest United States, however, this warning system has resulted in sporadic disease control failures. The purpose of the present study was to determine whether the warning system's algorithm could be modified to provide more reliable assessment of SBFS risk. Hourly LWD, rainfall, relative humidity (RH), and temperature data were collected from orchards in Iowa, North Carolina, and Wisconsin in 2005 and 2006. Timing of the first appearance of SBFS signs was determined by weekly scouting. Preliminary analysis using scatterplots and boxplots suggested that cumulative hours of RH ≥ 97% could be a useful predictor of SBFS appearance. Receiver operating characteristic curve analysis was used to compare the predictive performance of cumulative LWD and cumulative hours of RH ≥ 97%. Cumulative hours of RH ≥ 97% was a more conservative and accurate predictor than cumulative LWD for 15 site years in the Upper Midwest, but not for four site years in North Carolina. Performance of the SBFS warning system in the Upper Midwest and climatically similar regions may be improved if cumulative hours of RH ≥ 97% were substituted for cumulative LWD to predict the first appearance of SBFS.

Evaluation of a Disease Warning System for Phomopsis Cane and Leaf Spot of Grape: A Field Study

A field evaluation of a warning system for Phomopsis cane and leaf spot of grape (Vitis spp.), caused by Phomopsis viticola, was conducted in Ohio over 3 years (2002 to 2004) by applying fungicides and fungicide-adjuvant combinations based on predicted infection events. Three different criteria for risk-light, moderate, and high-were evaluated with the warning system. The warning system is based on measured weather conditions (temperature and wetness duration following rain) and a model for risk of leaf and internode infection. Vines were sprayed with fungicides based on either the warning system or a calendar-based 7-day protectant program, from 2.5-cm shoot growth (Eichhorn-Lorenz [E-L] stage 7) to the end of the broom (E-L stage 27). Fungicides were tested with or without an adjuvant (JMS Stylet-Oil or Regulaid). In the controls, the mean percentage of leaves and internodes with infections ranged from 36 to 100%, the number of lesions per leaf ranged from 1 to 28, and percentage of internodes covered by lesions ranged from 1 to 12%. Both the calendar-based protectant treatment (based on use of mancozeb) and the warning system treatment based on spraying in response to light or moderate predicted infection events (especially with mancozeb + Regulaid) resulted in significantly less disease incidence and severity compared with the controls. The mean percent control (relative difference in disease between a treatment and the control) was higher for the protectant schedule (˜55% and ˜80% for incidence and severity, respectively, based on application of mancozeb) than for the warning system (˜36% and ˜60% for incidence and severity, respectively, based on application of mancozeb + Regulaid), but there were two to three times more fungicide applications with the protectant schedule than with the warning system.

Development of MILIONCAST, an Improved Model for Predicting Downy Mildew Sporulation on Onions

The effects of temperature and relative humidity on Peronospora destructor sporulation on onion (Allium cepa) leaves were studied under controlled environmental conditions. Sporangia were produced most rapidly at 8 to 12°C after 5 h of high humidity during dark periods. The greatest number of sporangia was produced at 100% relative humidity (RH), and sporulation decreased to almost nil when humidity decreased to 93% RH. A model, named MILIONCAST (an acronym for MILdew on onION foreCAST), was developed based on the data from these controlled environment studies to predict the rate of sporulation in relation to temperature and relative humidity. The accuracy of prediction of sporulation was evaluated by comparing predictions with observations of sporulation on infected plants in pots outdoors. The accuracy of MILIONCAST was compared with the accuracy of existing models based on DOWNCAST. MILIONCAST gave more correct predictions of sporulation than the DOWNCAST models and a random model. All models based on DOWNCAST were more accurate than the random model when compared on the basis of all predictions (including positive and negative predictions), but they gave fewer correct predictions of sporulation than the random model. De Visser's DOWNCAST and ONIMIL improved their accuracy of prediction of sporulation events when the threshold humidity for sporulation was reduced to 92% RH. The temporal pattern of predicted sporulation by MILIONCAST generally corresponded well to the pattern of sporulation observed on the outdoor potted plants at Wellesbourne, UK.

Relation of Temperature and Rainfall to Development of Xanthomonas and Pantoea Leaf Blights of Onion in Colorado

During the 1996 to 1999 growing seasons, some areas of Colorado's onion-growing regions experienced a complex of bacterial diseases including leaf blights caused by Xanthomonas campestris and Pantoea ananatis. Crop losses varied depending on adverse weather (associated with rain, storm, and temperature patterns) and stage of onion plant development. Environmental conditions during vegetative development had no significant association with the initial appearance or subsequent intensity of disease. Both pathogens were active at average high temperatures that ranged from 28 to 35°C during bulbing. Multiple regression models were developed to predict the initial appearance (growing degree day [GDD]) and subsequent Xanthomonas leaf blight intensity (final proportion of disease [FPD]) using macroclimatic meteorological conditions, including July average daily high temperature (T_jmax), August cumulative rainfall (P_a), and cumulative rainfall in July and August (P_ja). Initial disease appearance and disease intensity were described by GDD₁₀ = -6,153.43 + 215.50T_jmax - 0.92P_a and FPD = 222.79 - 6.92T_jmax + 0.52P_ja, respectively. Pantoea leaf blight initial appearance was strongly associated with July average daily temperatures (T_j) and was described by GDD₁₀ = -5,930.43 + 289.07T_j. Results are discussed in relation to an integrated pest management strategy in Colorado.

Evaluation of the Burkard Cyclonic Spore Sampler for Collection Efficiency of Ascospores

The commercially available Burkard cyclonic sampler takes in air volumetrically, creates a vortex in an aluminum cylinder, and deposits particulates in an Eppendorf vial. Retention within the cylinder was determined for ascospores released by cultures of Sordaria fimicola and Gibberella zeae. Manufacturer's recommended amperage level and reduced power were tested, and a surfactant was applied to the cylinder wall. Under recommended power, an average of 78% of the S. fimicola ascospores were collected in the vial, while 22% lodged inside the cylinder. Conversely, only 25% of the G. zeae ascospores were collected in the vial, while 75% remained lodged inside the cylinder. Application of a surfactant to reduce the adherence of ascospores on the cylinder wall, instead resulted in 83% of the S. fimicola ascospores and 99.7% of the G. zeae ascospores deposited on the cylinder wall. When the power supply was decreased from 200 mA to 140 mA, the ratio of ascospores of G. zeae retained by the sampler remained nearly the same, indicating that the retention error was not a function of airflow rate within the tested power range. However, the total number of ascospores collected was significantly less under reduced power. A 90-A/h (12 V) battery supplied greater than 150 mA when connected to the Burkard for 4 days at temperatures between 22 and 28°C and could maintain current above 190 mA for 7 days with a 21-W solar panel. A 21-W solar panel charging a 90-A/h battery should maintain the amperage needed for the Burkard to maintain proper flow rate in most environments. The aluminum cylinder should be rinsed out thoroughly when collecting samples of fungal ascospores or any other particulates with the propensity to adhere to the cylinder wall.

Evaluation of a Disease Warning System for Downy Mildew of Grapes

An electronic warning system for grape downy mildew- based on models for the infection of leaves of Vitis lambrusca, production of sporangia by Plasmopara viticola in lesions, and sporangial survival-was tested over 7 years in Ohio. Grapevines were sprayed with metalaxyl plus mancozeb (Ridomil MZ58) when the warning system indicated that environmental conditions were favorable for sporulation and subsequent infection. Over the 7 years, plots were sprayed from one to four times according to the warning system, and from four to 10 times according to the standard calendar-based schedule (depending on the date of the initiation of the experiment). The warning system resulted in yearly reductions of one to six sprays (with median of three sprays). Disease incidence (i.e., proportion of leaves with symptoms) in unsprayed plots at the end of the season ranged from 0 to 86%, with a median of 68%. Incidence generally was very similar for the warning-system and standard-schedule treatments (median of 7% of the leaves with symptoms), and both of these incidence values were significantly lower (P < 0.05) than that found for the unsprayed control, based on a generalized-linear-model analysis. Simplifications of the disease warning system, where sprays were applied based only on the infection or sporulation components of the system, were also effective in controlling the disease, although more fungicide applications sometimes were applied. Effective control of downy mildew, therefore, can be achieved with the use of the warning system with fewer sprays than a with a standard schedule.