OTA-Grapes: A Mechanistic Model to Predict Ochratoxin A Risk in Grapes, a Step beyond the Systems Approach

Ochratoxin A (OTA) is a fungal metabolite dangerous for human and animal health due to its nephrotoxic, immunotoxic, mutagenic, teratogenic and carcinogenic effects, classified by the International Agency for Research on Cancer in group 2B, possible human carcinogen. This toxin has been stated as a wine contaminant since 1996. The aim of this study was to develop a conceptual model for the dynamic simulation of the A. carbonarius life cycle in grapes along the growing season, including OTA production in berries. Functions describing the role of weather parameters in each step of the infection cycle were developed and organized in a prototype model called OTA-grapes. Modelling the influence of temperature on OTA production, it emerged that fungal strains can be shared in two different clusters, based on the dynamic of OTA production and according to the optimal temperature. Therefore, two functions were developed, and based on statistical data analysis, it was assumed that the two types of strains contribute equally to the population. Model validation was not possible because of poor OTA contamination data, but relevant differences in OTA-I, the output index of the model, were noticed between low and high risk areas. To our knowledge, this is the first attempt to assess/model A. carbonarius in order to predict the risk of OTA contamination in grapes.


Prototype Predictive Model OTA-Grapes
The relational diagram, slightly modified with respect to the first published version [3], is shown in Figure 1. The inoculum source, mainly mycelium overwintering in the soil [39], not quantified in the model (Oi), is the opening state variable; spores produced from this source are air-borne to bunches. The amount of spores disseminated to berries (SoB) is regulated by a dispersal rate (DisR), depending on the temperature (T) and rain (R) regimes. The germination rate (GeR) regulates the amount of spores that can germinate on berries (GSoB), and the growth rate (GR) regulates the growth of germinated spores (GB). Then, the mycelium can follow two infection pathways: it can (i) stay on berries and colonize their skin (CB); or (ii) penetrate and infect berry flesh (IB). The pathway depends on the berries' status (BS), mainly (i) presence/absence of openings in the skin and (ii) grape growth stage [37]. Many factors can determine skin damages, but R during ripening and pest-disease (P&D) presence play a major role. CB and IB are considered for the OTA production index (OTA-I) through a toxin production rate (ToxR). The influence of grape variety has been removed in this version of the relation diagram because of the lack of useful data to develop a proper function; this information is now accounted for by GS.

Overwintering Inoculum (Oi)
Spore production depends on the amount of fungal inoculum in soil and ecological conditions. The inoculum was assumed to be naturally available, based on many surveys managed in different vine growing areas signaling black aspergilli on berries at setting [40][41][42].
Sporulation of A. carbonarius was reported to take place with a combination of T and relative humidity (RH) equal or higher than 25 °C and 85%, respectively, referring to quantitative data collected for A. niger sporulation on onion [43,44]. No quantitative data are available in the literature regarding the sporulation of black aspergilli on grapes in different ecological conditions, but based on the experience with the AFLA-maize model, developed to predict the behavior of A. flavus in maize [45], we assumed that spores are available with a combination of T and RH equal to or higher than 10 °C and 80%, respectively. This was included as a binomial variable (yes/no answer) in the model.

Spore Dispersal (DisR)
Dispersal of A. carbonarius spores was confirmed to be wind borne by Kazi, Emmett, Nancarow and Clarke [39], but they did not publish quantitative data. Therefore, spore dispersal was included as a binomial variable, too, and it was assumed as not possible on rainy days or with RH ≥ 80%, similarly to the statement used in the AFLA-maize model for A. flavus [45].

Germination Rate (GeR)
Spore germination is mainly influenced by T and aw. Information available in the literature about this step of the infection cycle [47,48] till recently was poor, but suitable quantitative data on A. carbonarius germination on grape flesh and skin and their modelling were recently published [46]. Briefly, conidia germination occurred more rapidly on grape flesh (6 h), followed by grape skin (24 h), under the optimal conditions of 30-35°C and 100% RH. The combined effect of T and RH/aw on the germination of A. carbonarius conidia in grape berries (both on skin and flesh) was modelled with a Bete equation (T) and a polynomial equation (aw); these functions were used for GeR in the OTA-grapes model developed here (Table 1).
where Teq is the equivalent of T, computed as: Tmin and Tmax, the cardinal temperatures, are 5 °C and 45 °C, respectively, in this case. Estimated parameters are provided in Table 1.
Regarding T (Figure 2a), Aspergillus carbonarius growth starts from 10 °C; GR increases up to 25-30 °C (optimal conditions), decreasing at higher T values; growth was never observed at 45 °C.
Regarding aw (Figure 2b), the best fit of data was obtained with a linear equation. Water activity = 0.88 is the lowest limit for growth, while aw = 0.98-0.99 is the optimal condition. Table 2. Symbols included in Figure 2 and Figure 3 for the references, including quantitative data used in this study to run regression analysis and implement equation parameters for growth rate (GR) and ochratoxin production rate (ToxR) reported in Table 1.

Infection Rate (IR)
Fungal infection strictly depends on berry status. The infective capacity of eight strains of A. carbonarius on damaged/undamaged grape berries was considered by Belli et al. [56]. Damaged berries resulted in being more susceptible to the infection compared to intact ones. Relative humidity ≤ 80% resulted in being not conducive for infection; in fact, with RH = 80%, 25% of the studied strains did not infect damaged berries. On un-damaged berries, no infection occurred with RH between 75% and 80%.
Conducive conditions for the infection are computed by IR, taking into account berries status (with/without lesions) and grape growth stage. In fact, A. carbonarius infection is promoted when: (i) berries have lesions; (ii) they are close to harvest (late ripening); and (iii) air temperature is around 30 °C and RH ≥ 85% [37,56].
The Bete equation was chosen to fit all data as a function of T, to compute ToxRT, but it resulted in a poor data fitting (R 2 = 0.54; Table 1, Figure 3a). Strains with significant difference in optimal T were clearly distinguishable in published papers; therefore, data were shared in two groups based on fungal strain behaviors in different T regimes. The Bete equation was run for each of the two derived datasets, and R 2 improved significantly for both functions developed (R 2 > 0.75; Table 1, Figure 3b-c).
Ochratoxin A production rate in different aw regimes (ToxRaw) was successfully fitted by a linear function (Table 1). No OTA production was described at aw regimes ≤ 0.87 and ToxRaw increased from 0.90-0.99, the optimum aw for OTA production. (b,c) data shared based on the dynamic of OTA production and according to the optimal temperature of different strains (20 and 25 °C, respectively). Symbols refer to data from the literature, and the references are reported in Table 2.

Effect of Weather Data on ToxRT
In the fungal population naturally present in vineyards, the prevalence of strains belonging to the two groups previously described for toxin production rate (ToxRT1 and ToxRT2) is unknown. Then, data from 48 weather stations (see Materials and Methods for details) were used to compute ToxRT, for the two strain types, giving them a different role. The two functions developed (ToxRT1 and ToxRT2) were considered to complementarily contribute between 0% and 100% (100% as the global contribution; Table 3). The overall mean for ToxRT1 and ToxRT2 values (output dataset, computed including 48 weather stations and 11 combinations of the two strain types) was 638 and the range of variation was limited, both between years and areas of weather data collection, with a minimum and maximum in 2010 (±0.2-±3.1) and in 2011 (±15.8-±27.6), respectively.
One-way ANOVA was applied to the output dataset. Significant differences (p ≤ 0.01) were noticed regarding ToxRT intra-growing areas and intra-years (p ≤ 0.05) ( Table 3). The index obtained with equal contribution of both strain type (50%/50%) was significantly different only from the 0%/100% and 100%/0% combinations. Therefore, it was assumed as a correct approach to give the same role in the predictive model to the two population types, therefore to the two functions, it being considered impossible to have only one strain type in a grapevine growing area. Table 3. Mean values of ochratoxin A production rate (ToxRT) obtained running the two functions', ToxRT1 and ToxRT2 (see Table 1), generated clustering data, based on fungal strain behavior in different T regimes, in different complementary proportions ranging from 0%-100%.

Model Output (OTA-I)
OTA-grapes was run using as input meteorological data collected in years/georeferenced locations where OTA contamination in grapes was quantified at harvest in previous research studies. In particular, two meteorological stations from the Apulia region, Southern Italy, in 1999 and 2000 [8], and two in the Emilia Romagna region, Northern Italy, in 2012-2013 were selected (Battilani et al. [57]). OTA above the EU legal limit was detected in 1999 in Apulia, while it was absent in 2000 [8]. OTA was never detected in Emilia Romagna. OTA-I was high in Apulia, in both years, higher in 1999 compared to 2000, 3819 versus 2942, while very low levels were obtained running OTA-grapes in Emilia Romagna, six and eight respectively in 2012 and 2013.

Discussion
Ochratoxin A occurrence in grapes and wine has been largely reported after 1996 in several European countries [3]. Surveys agree on the decreasing gradient of OTA incidence and severity from red, rosè and white wines [8,[58][59][60]. Higher contamination is reported for dessert wines, due to the over-maturation/drying of grapes before the wine making process, in agreement with high contamination reported in raisin and currant. In general, grape derivatives produced in South Europe and North Africa are more severely affected by OTA than those from the temperate regions of the middle of Europe [17,61,62].
The proportion of wine in which OTA is detected is high (>50%) in some countries, but the legal limit of 2 µg/L fixed by the EC [63] is rarely exceeded [3]. Therefore, OTA contamination seems a problem actually under control, both in conventional and organic farming, in Europe, as confirmed by Comuzzo et al. [64], who found OTA below the legal limit in 95% of the 971 wine samples analyzed.
The alarm regarding OTA contamination in wine, launched around the end of the 20th century, is now apparently back, thanks to cropping system actions enforced to improve grapevine wellness, mainly preventing pest and disease damages on berry, including the control of black aspergilli, the optimization of postharvest logistic and of wine making [3,18,29,31,65]. Nevertheless, this is true for well-studied and well-managed grape growing areas, but the modelling approach can also support poorly-studied areas, if meteorological data are available, and this can help in grapevine planning and management, more useful with the extreme climatic events frequently faced in recent years. Global warming has markedly shifted the distribution of temperature variability and extremes and the precipitation patterns with significantly higher frequency of exceptionally unfavorable years for several crops [66]. Therefore, it is very important to organize a predictive system for OTA risk in grapes in order to support all of the operators in the grape chain, including policy makers and risk managers, in their strategic and tactic decisions, especially in climate change scenarios, as recently stressed by Battilani and Camardo Leggieri [67] for aflatoxin in maize.
Aspergillus carbonarius is the key fungus for OTA production in grapes; it is commonly present in vineyards [58,68,69], but relevant OTA contamination is quite rare [3]. Therefore, OTA-grapes, the predictive model developed in this study, focused on A. carbonarius to support the prediction of OTA risk in year/area combinations. To our knowledge, this is the first attempt to assess/model the role of A. carbonarius strains in order to optimally represent what happens in reality, which is the principal aim of predictive modelling [70,71]. Empirical models were developed to describe the influence of T, aw and growing medium on specific steps of the infection cycle [21,46,[72][73][74][75][76]. They were a very good support for our modelling; suitable equations were developed to describe rate variables, with R 2 ≥ 0.83 in all of the equations included in OTA-grapes.
In our study, working on the OTA production rate as a function of T, it emerged that fungal strains can be shared in two different clusters, based on the dynamics and the optimum of OTA production as a function of T. This is in agreement with several molecular studies that underlined the great biodiversity of A. carbonarius strains, often unrelated to isolation area and harvest year [21,22,77,78]. The OTA production rate could be considered a potential weakness of the model; in fact, realizing the relevant diversity between strains in OTA production dynamic and it being impossible to know the composition of natural A. carbonarius populations, the statistical approach was used, and an equal contribution of the two strain clusters to OTA synthesis was assumed. Besides, quantitative data available were all obtained in vitro on artificial media, and this could face differences when berries are considered as the growing medium, an issue stressed also for AFLA-maize [45]. Nevertheless, we consider the approach appropriate, the interest being focused on predicting the level of risk, not the amount of OTA content in berries, in agreement with the common opinion that the latter is almost impossible to predict.
The reduction of A. carbonarius population in vineyards could be successfully achieved by control measures, which focus on the application of chemical/biological fungicides [65,79]. Chemical formulates, such as Azoxystrobin, Dinocap, Chorus and Switch, have been tested for their efficacy for preventing OTA accumulation in grapes with successful results [34,79]. In view of avoiding the use of fungicides when unnecessary and to face integrated pest management (IPM), actually mandatory in Europe [80], OTA-grapes increased the patho-system understanding [81], and it could be a good support for stakeholders in the grape chain being able to give a daily prediction on the level of risk in different areas and years.
Model validation, the final mandatory step to release reliable predictions, can be generally summarized as the comparison between model predictions and real observations of a certain phenomenon, e.g., disease outbreak, disease incidence or OTA contamination in grapes, in this case to check their agreement. Model validation requires a consistent set of field data on the modelled phenomenon, possibly shared above and below a reference threshold, the legal limit in this case. Unfortunately, most of the available data reports OTA contamination below the legal limit; then, the definition of a probability function for OTA contamination was not possible. In our study, we observed a certain level of agreement between OTA-grapes predictions/output (OTA-I) and OTA level detected in the field, for the vineyards considered. In fact, as reported in the Results section, OTA-I was greater for OTA-contaminated grapes. This is an encouraging result, even if not yet supported by a consistent validation.
In conclusion, OTA-grapes, even if not yet validated, represents a step beyond the systems approach. Predictive mycology is useful to assess fungal growth and, therefore, mycotoxin contamination in food products [82], and it should receive more attention to improve food safety [75,83]. It must be kept in mind that model predictions, even if reliable, will never be 100% correct. Therefore, supporting tools, like rapid diagnostics applied at harvest [84], paired with risk prediction, should make model support to farmers more robust [67].

Material and Methods
This study followed the approach previously described for AFLA-maize, a predictive model for Aspergillus flavus in maize, another important mycotoxin-producing fungus with a wide range of potential hosts [45]. All quantitative data available in the literature regarding the A. carbonarius-grape patho-system were considered.

Literature Search
Following the protocol developed by Okoli and Schabram [85], published papers that contain quantitative data for the patho-system of interest were retrieved. The papers had to comply with the following criteria, to be used for model development: (i) containing original data on the ecology or epidemiology of A. carbonarius (preferably, but not exclusively, the interaction with the host crop); and (ii) be published in peer-reviewed journals or proceedings of top level congresses. The literature search was carried out in the CAB Abstract database (http://www.cabdirect.org). All papers were reviewed, and in case they met the eligibility criteria, quantitative data reported were collected in a proper database.

Model Development
The relational diagram developed by Battilani and Silva [3] was considered, slightly modified, to develop the prototype predictive model named OTA-grapes ( Figure 1).
Briefly, in the relational diagram [86], the status of the fungus at a certain time is represented by boxes, intended as state variables (Oi, SoB, GSoB, GB, CB, IB and OTA-I); clouds are used when it is not possible to quantify the state variable, commonly the overwintering inoculum (i.e., Oi, Figure 1). The flow from one state to the following is driven by constants/parameters (T, R, RH, aw, P&D) or intermediate variables (BS and GS), essentially derived from weather or crop data. Rate variables are represented by "valves" (DisR, GeR, GR, IB, CR, ToxR). They are described by mathematical functions, and the output is reported on a 0-1 scale, with 1 intended as the rate occurring at optimal conditions; the flow to the next stage is stopped due to limiting conditions when the rate is 0.
The linear or the non-linear regression procedure of the statistical package PASW statistics (Ver. 21, 2012) was used to develop mathematical functions to be used for all rate variables included in the relational diagram. Functions were finally linked in a coherent mathematical framework to calculate the output, i.e., the OTA cumulative index (OTA-I).

Grape Growth Stage Model
The intermediate variable GS (= crop growth stage) is a crucial component of the predictive model. A sub-model able to predict grape growth stages based on normal heat hours (NHH) summation was considered; it was developed for several grape varieties by Mariani et al. [87] and validated in Italy by Cola et al. [88] with good results. NHH are hours weighted (in the scale 0-1) on plant activity T range (Tmin and Tmax) and optimal T (Topt).

OTA-Grapes Data Input and Output
Hourly data on air temperature (T, °C) and relative humidity (RH, %) and rain (R, mm) are the mandatory data input. Data from 1 January are requested for crop phenology modelling; when early veraison is reached, the A. carbonarius predictive model starts. Model running continues till berries are ripe, computing OTA-I on a daily base; it accumulates in time, and this accumulated value is the output of the OTA-grapes predictive model.

Meteorological Data
Hourly data on T, RH and R along grape growing season, available from previous studies, were considered. Eight meteorological stations placed in the Emilia Romagna region (Northern Italy) in 6 years 2006-2011 were used to test the effect of weather data on ToxRT,T1,T2.
Two stations from the Apulia region, Southern Italy, in 1999 and 2000 [8], and 2 in the Emilia Romagna region, Northern Italy, in 2012-2013 were selected due to the availability of data on OTA contamination in grapevines grown close to the stations. They were used to validate the OTA-grapes model.

Effect of Weather Data on ToxRT
Two functions were developed for the rate variable ToxRT in different T regimes to describe 2 groups of A. carbonarius strains based on their behavior in different T regimes (ToxRT1 and ToxRT2 for optimal T at 20 and 25 °C, respectively). The 2 functions were run for all 48 weather stations selected in Emilia Romagna in the period 1 August-15 September, approximately between early veraison and ripening for grapes in the growing area considered, for all 6 years.
The output values obtained for the 2 functions, in each weather station and year, were combined giving to each ToxRT a contribution changing from 0%-100%, with a step of 10%. The aim was to simulate a different role of the 2 types of strains. Eleven ToxRT values were therefore obtained for each station and year, 528 values in total. One-way ANOVA was run with these data using alternatively year and weather station as factor and replicate; the Tukey test was applied to highlight significant differences between means.

Preliminary OTA-Grapes Validation
The model was validated using data from 4 natural epidemics recorded in two grape-growing regions of Italy in four years (1999,2000,2012,2013). Two vineyards were monitored in Apulia in 1999 and 2000, and they included the only data available from previous research activities, regarding OTA contamination in Italy above the EU legal limit. Forty vineyards were sampled in Emilia Romagna in 2012 and 2013; OTA was never detected in these vineyards.