Introduction

The bacterium Xylella fastidiosa (Xf) is a plant endophyte native to the Americas, which develops in up to 300 plant species including ornamental and agricultural plants1. Xf is transmitted between plants by xylem-feeding insects belonging to several families of Hemiptera (Aphrophoridae, Cercopidae, Cicadellidae, Cicadidae and Clastopteridae)2. Xf causes severe plant pathologies leading to huge economic losses3 e.g., the Pierce’s disease of grapevines PD4, the olive quick decline5, the oak bacterial leaf scorch6, the phony peach disease7, the Citrus variegated chlorosis CVC8 and the almond leaf scorch9. As Xf induces diseases to a large number of economically important plants including vine1, the biology of this pathogen and the mechanisms of vector transmission have been extensively studied to design management strategies10.

On the basis of genetic data obtained with Multilocus Sequence Typing MLST11,12, Xf was subdivided into six subspecies (fastidiosa, morus, multiplex, pauca, sandyi and tashke). Those subspecies were further characterized by different geographic origins, distributions and host preferences in the Americas13,14,15. However, the intraspecific taxonomic boundaries of Xf are still debated16 and only the two subspecies fastidiosa and multiplex are formally considered valid names1,17. Xf subsp. fastidiosa18 occurs in North and Central America, where it causes, among others, the harmful PD and the almond leaf scorch (ALS). Genetic analyses suggest that this subspecies originates from southern parts of Central America19. The subspecies multiplex is widely distributed in North America (from California to western Canada and from Florida to eastern Canada), where it was detected on a wide range of host plants (e.g., oak, elm, maple, almond, sycamore, Prunus sp., etc.) as well as in South America20,21. The subspecies pauca, which causes severe diseases in Citrus (CVC) and coffee (Coffee Leaf Scorch)22 in South and Central America, is speculated to be native to South America23. The subspecies morus recently proposed by Nunney et al.24, occurs in California and eastern USA, where it is associated to mulberry leaf scorch. Xf subsp. sandyi, responsible for oleander leaf scorch, is distributed in California12, while the subspecies tashke was proposed by Randall et al.25 for a strain occurring on Chitalpa tashkentensis in New Mexico and Arizona. Overall, intraspecific entities of Xf display noticeable differences in host range suggesting that the radiation of Xf into multiple subspecies and strains is primarily associated to host specialization26.

Xylella fastidiosa is now of worldwide concern. In 2013, the CoDIRO strain (subsp. pauca) was detected on olive trees in southern part of the Apulia territory (Italy). Genetic analyses suggest that this strain was accidentally introduced in Italy from Costa Rica or Honduras via infected ornamental coffee plants5. Since then, Xf subsp. pauca has spread northward and killed millions of olive trees in the Apulia territory, causing unprecedented socio-economic issues. During the period 2015–2017 several subspecies and strains were detected on ca. 30 different host plants in Southern France (PACA region) and Corsica27. According to national surveys performed in France, the vast majority of plant samples were contaminated by two strains of Xf subsp. multiplex, and two strains were identified (hereafter referred to as the French ST6 and ST7 strains)27. These strains are closely related to the Californian strains Dixon (ST6) and Griffin (ST7), belonging to the “almond group”26 and that were detected on numerous plant species, though without evident specialization. To a lesser extent, other strains occur in Southern France i.e., the strain ST53 (Xf subsp. pauca) was detected on Polygala myrtifolia in Côte d’Azur (Menton) and on Quercus ilex in Corsica27 and the recombinants strains (ST76, ST79 or not yet fully characterized) were detected in a few plant samples. In 2016, Xf subsp. fastidiosa was detected on rosemary and oleander plants overwintering in a nursery in Germany28. In 2017, Spanish plant biosecurity agencies officially confirmed the detection of Xf strains belonging to the subspecies multiplex, pauca and fastidiosa on almond trees, grapevine, cherry and plums in western parts of the Iberian Peninsula and Balearic islands29. Outside Europe, the detection of Xf was officially confirmed in Iran on almond trees and grapevines30, in Turkey31 as well as in Taiwan on grapevines32.

The severity of Xf-induced diseases has recently increased possibly due to global warming33. Indeed, it has been demonstrated that cold winter temperatures might affect the survival of Xf in xylem vessels and allow plants to partly recover from Xf-induced diseases (‘cold curing phenomenon’)34,35. For instance, Purcell34 showed that grapevines with symptoms of PD recovered after multiple exposures to temperatures below −8 °C during several hours. Further, Anas et al.36 suggested that areas experiencing more than 2 to 3 days with minimal temperature below −12.2 °C (or alternatively 4 to 5 days below −9.9 °C) should be considered at low risk for PD incidence, although these thresholds were considered too conservative by Lieth et al.37. Several studies aimed to forecast the potential distribution of Xf in Europe38 and/or all over the world39. For instance, Hoddle et al.39 used the CLIMEX algorithm to forecast the worldwide potential severity of PD. Their model suggested that most of Mediterranean areas are suitable for PD even though cold in winter would presumably hamper Xf range expansion into several of the most economically-important wine-producing regions of France and central and northern parts of Spain and Italy. Bosso et al.38 fitted a Maxent model to forecast the potential distribution of Xf subsp. pauca under current and future climate conditions, and concluded that climate change would not affect the future distribution of Xf. Here, we analyze the potential distribution of three subspecies of Xylella fastidiosa using datasets describing native and newly established ranges and a set of four different species distribution models.

Material and Methods

Distribution data

We collected occurrence data for subspecies fastidiosa, multiplex and pauca from the scientific literature, field surveys and public databases (Fig. 1). These datasets comprised both occurrences from native area (the Americas) and recently invaded regions (south Italy, France and Spain - Fig. 1). The occurrences located in France were collected in 2015–2019 and stored in the French national database managed by the French Agency for Food, Environmental and Occupational Health & Safety (ANSES) (Fig. 1C). For each subspecies, we randomly generated 10,000 background points within a wide area to properly depict the background environment in the Maxent calibration (see below)40,41 (Supplementary Fig. S1). Pseudo-absences were randomly generated within regions located in the native area where we could confidently consider that the studied subspecies is absent. We restricted these areas to cold regions because low winter temperatures constitute a well-known factor constraining the distribution of Xf 35 (Supplementary Fig. S1).

Figure 1
figure 1

Occurrences of three Xylella fastidiosa subspecies used in the study. (A) Xylella fastidiosa fastidiosa, (B) Xylella fastidiosa multiplex in its native range and (C) in Europe, (D) Xylella fastidiosa pauca in its native range and (E) in Europe. European occurrences of Xylella fastidiosa fastidiosa are sparse and not shown.

Full size image

Bioclimatic descriptors

We used a set of bioclimatic descriptors hosted in the Worldclim database42. We retained raster layers of 2.5-minute spatial resolution, which corresponds to about 4.5 km at the equator. The data represent the average climate conditions for the period 1970–2000.

Models

The potential distribution of Xf subsp. fastidiosa, pauca and multiplex were assessed using species distribution modeling. Such approach establishes mathematical species-environment relationships using occurrence/absence records and environmental descriptors in order to assess the potential distribution of species43. Different modeling techniques exist and their ability to predict species distribution / habitat suitability is known to vary substantially according to the algorithm used, the occurrence dataset, the environmental descriptors and the model calibration parameters44,45. As a consequence, there is no single ‘best’ modeling technique46. Such variation has led to the development of ensemble forecasting approach, which aims at building more robust forecasts by combining individual models in a consensus model47. However, averaging models’ outputs in the form of probabilities might raise issues. Indeed, models’ response to species prevalence could differ and yield non-comparable probabilities. A solution consists to use a committee averaging approach44,48.

In the present study, we adopted a two-step modeling strategy46. In the first step of the analysis we tested a set of four algorithms known for their good performance in species distribution modeling (Table 1). The test consisted in (i) fitting models using the occurrences available in the native areas only and (ii) evaluating the predictive power of those models into the recently colonized areas in Europe. It is known that bioclimatic descriptors used to calibrate the models can strongly impact performance and transferability. However, choosing proper descriptors is not easy. Consequently, we constituted seven subsets of bioclimatic descriptors by associating different variables that we a priori considered as ecologically meaningful when working on Xf (Table 1). We intentionally used a limited number of climate descriptors comprised between two to four to avoid model over-parameterization, which is a recommended practice, particularly when assessing invasion risk49. We used two descriptors of the low temperatures during the coldest periods of the year (bio6: minimum temperature of the coldest month; bio11: mean temperature of the coldest quarter). We also used a variable reflecting high temperatures during the warmest period of the year (bio10) and a variable reflecting the rainfall seasonality (bio15) that recently proved to be a good predictor of the spatial distribution of Xf in Corsica50. For each subspecies of Xf, algorithms were calibrated using each climate data subsets and the occurrences available in the Americas. The predictive power of each model was evaluated using the area under the curve of the receiver operating curve (AUC) statistic and the true skill statistic (TSS)51,52. This procedure aimed to identify the combinations of algorithms × climate datasets with weak predictive performances and to discard them from further analysis46.

Table 1 Range of the evaluation metrics for species distribution models calibrated with different climate datasets for Xylella fastidiosa fastidiosa, X. fastidiosa multiplex and X. fastidiosa pauca.
Full size table

In a second step of the analysis we used the best combinations of algorithms × climate datasets to calibrate models with all available subspecies occurrences to predict habitat suitability in Europe (referred to as full models). For each combination of model and climate datasets we calibrated the models using a randomly selected subset of eighty percent of the occurrence data (native and invaded ranges) and used the remaining twenty percent for model evaluation using AUC and TSS metrics. Five replications were done for each full model. Evaluations were done using a set of 100 randomly generated pseudo-absences.

We removed autocorrelated data to improve model predictive performances53,54. To do so, we used the first two axes of a Principal Component Analysis performed on the bioclimatic variables55 recorded at each occurrence points. The first axis was divided into 100 bins. The bins of the second axis were fixed to have the same amplitude as for axis 1. The occurrence points were projected onto the resulting grid. When a grid cell contained more than one point, a random selection was used to retain only one point for further model calibration. This procedure was repeated for each Xf subspecies and for each climate dataset.

Model outputs were transformed into binary projections using the threshold that optimized the TSS statistics on the testing data56. The resulting projections were averaged to compute the committee (consensus) averaging that shows the likelihood of the presence of a species given the available data. The consensus model ranges from 0 (all the models predict absence) to 100 (all the models predict presence)44,56. We removed the individual models that did not reach the minimum quality threshold of TSS >0.6 and AUC >0.8556 before computing the consensus (Table 1).

Our set of four algorithms comprised approaches belonging to the three main functional groups of species distribution models57. First, we selected the envelope model Bioclim58,59 that relies only on presence data and as such makes no assumption about the absence of the organism under study60. Second we employed the maximum entropy algorithm Maxent that discriminates presences with background data61. Third, we used modeling techniques that rely on presence and absence or pseudo-absences namely the generalized linear model (GLM)62 and the artificial neural network56,63. Maxent was fitted using 10,000 background points while GLM and artificial neural network were fitted using 200 pseudo-absences (Fig. S1).

Caution is usually warranted when interpreting models projected into new areas with climate conditions different from the calibration area64,65,66. Thus, we assessed the similarity of climate conditions between the calibration dataset and the projected area (i.e., Europe) by computing the MESS index41.

The following R67 packages were used to perform analyses and generate graphical outputs: biomod268, cowplot69, dismo70, ecospat71, ggplot272, raster73 and rmaxent74.

Results

Algorithm and climate datasets selection

The best combinations of algorithms × climate datasets for Xf multiplex and Xf pauca were evaluated with a dataset comprising all European occurrences. As there were too few occurrences in Europe for Xf fastidiosa, a random subset of 20% of the native range occurrences was used. Models were selected on the basis of an arbitrary threshold of the TSS fixed to 0.656. Only the following combination did not get sufficient statistical support to be retained in future analysis of the potential distribution of Xf pauca: bioclim × dataset 2, 3, 6 and 7; GLM × dataset 2, 3, 6 and 7 and Ann × dataset 6 (Table 1). The remaining combinations were used to calibrate the full models based on the complete set of occurrences.

The accuracy of the resulting full models was examined and we only retained the full models associated to TSS >0.6 and AUC >0.85 for the computation of the consensus model. This quality check resulted in discarding seven, four and four models for Xf fastidiosa, Xf multiplex and Xf pauca respectively. Table 1 shows the range of the evaluation metrics for the models used to compute the consensus models (Fig. 2).

Figure 2
figure 2

Potential distribution of three subspecies of Xylella fastidiosa: (A) Xf subspecies fastidiosa (B) Xf subspecies multiplex (C) Xf subspecies pauca. Maps depict the ensemble forecast derived from committee averaging based on lowest presence thresholding (see methods section for details). The index varies from 100 when all models predict presence to 0 if all the models predict absence of the subspecies.

Full size image

Potential distribution of Xylella fastidiosa

Figure 2 shows the committee averaging for the three subspecies of Xf. Each map shows the proportion of models predicting the presence of Xf in Western Europe. The potential distribution of Xf subsp. fastidiosa includes large regions of Spain, France, Italy, Croatia, Greece and Turkey as well as the coastal regions of North Africa (Fig. 2A). The proportion of models indicating favorable climate conditions in these areas was close or equal to 100%. The agreement between algorithms × climate datasets was somewhat lower in the northern part of France, the British Islands, Belgium and Netherlands where a lower proportion of models predicted the presence of Xf subsp. fastidiosa. In the case of Xf subsp. fastidiosa (Fig. 2A), the MESS index indicated that climate conditions encountered within Western Europe do not differ from the climate conditions that characterize the native area (Supplementary Fig. S2).

The predicted potential distribution of Xf subsp. multiplex is depicted in Fig. 2B. European climate appears favorable in a very large area covering Spain, France, the British Isles, Italy, the Adriatic coast, Greece, Turkey and some coastal areas of the Black Sea. North Africa and Mediterranean coast of near East countries are also climatically suitable. The MESS index indicated a good match between the range of climate conditions that prevail in Europe and in the American range of Xf subsp. multiplex (Supplementary Fig. S3).

The predicted extent of climatically suitable conditions for Xf subsp. pauca is limited to the Mediterranean coastal regions with the exception of south Portugal and Spain Atlantic coasts. The MESS index indicated a mismatch between the minimum temperatures (bio6 and bio11) in northern Europe and the native area (Supplementary Fig. S4).

Discussion

Geographical distribution and possible impacts in Europe

In a rapidly changing world, the design of pest control strategies (e.g., early detection surveys and planning of phytosanitary measures) should ideally rely on accurate estimates of the potential distribution and/or impact of pest species as well as their responses to climate change75. In the present study, bioclimatic models predicted that a large part of the Mediterranean lowlands and Atlantic coastal areas of Europe are characterized by climatically suitable conditions for Xf subsp. fastidiosa, multiplex and pauca (Fig. 2). To a lower extent, favorable climate suitability is also observed in northern and eastern regions of Europe (North-eastern France, Belgium, the Netherlands, Germany, etc.).

Our models displayed good evaluation measures and predicted high climatic suitability in all European areas where symptomatic plants are currently infected by the subspecies fastidiosa, multiplex or pauca (e.g., Balearic Islands, lowlands of Corsica island, south-eastern France and the Apulia region). This suggests that climate suitability maps provided in the present study are reliable for the design of further sampling strategies, including ‘sentinel insects’ survey76,77. They may also be helpful to anticipate the spread of the different subspecies and provide guidance on which areas should be targeted for an analysis of local communities of potential vectors and host plants and to design further management strategies and research projects.

The results show that the different subspecies of Xf studied here might significantly expand in the near future, irrespective of climate change. For example, the subsp. multiplex known from Corsica and southern France have a large potential for expansion in Europe (Fig. 2B), which is not surprising since this subspecies has a wide distribution, ranging from Florida to Canada78. Actually, its expansion probably depends more on plant exchanges, vector spatio-temporal patterns and disease management than on climate suitability per se. Xf multiplex is associated to economically important plants such as almonds and olives26 but may also colonize multiple ornamental plants or forest species27. Its potential distribution in Europe extends far beyond areas where the subspecies has been reported which suggests that new outbreaks may occur that could result in important economic losses.

The subspecies fastidiosa, which has been currently reported from a limited number of localities in Europe, could encounter favorable climate conditions in various areas (Fig. 2A). Notably, the models predict climate suitability in strategic wine-growing areas in different countries. The herein estimates of the potential distribution of the subsp. fastidiosa are consistent with the risk maps provided by Hoddle et al.39 and Purcell (available in Anas et al.36).

The case of subsp. pauca is somewhat different (Fig. 2C). Most of the European occurrences come from southern Italy and the Balearic Islands and the potential distribution of this subspecies appears limited. Nevertheless, southern Spain and France, Portugal, Corsica, Sardinia, Sicilia and North Africa that are areas where growing olive trees is multisecular offer suitable conditions, which potentially implies huge socio-economic impacts.

One factor that proved to be critical for some insect-borne plant diseases is the distribution/availability of vectors and hosts. Here, none of these factors is limiting in most of Europe since Xf is capable of colonizing a vast array of plants present in Europe and Philaenus spumarius, the putatively most efficient European vector so far79,80, occurs across most of the continent77.

The potential distribution of the three subspecies of Xf studied here appeared to be limited by minimum winter temperatures with Xf subsp. pauca being much more sensible than the others. Because “cold curing” appears to be the main regulating mechanism, is it very likely that climate change would alter the distribution of suitable areas for Xf in Europe as the minimum winter temperatures might increase34,35,81. Furthermore, the potential dynamics of Xf in areas experiencing extremely high temperatures in summer (e.g., southern and central Spain) remain largely uncertain as the impact of extreme heat on Xf is poorly known82. Although warm spring and summer temperatures enhance multiplication of Xf in plants, it has been showed that Xf populations decrease in grapevines exposed to temperatures above 37 °C35. As southern and central Spain frequently experience temperatures above 40 °C in summer, additional data would be helpful to better understand the potential distribution and impact of Xf in these regions. Another point requiring clarification is the effect of rain and moisture. As Xf bacteria live in the xylem of plants they are subject to stress whenever their host is itself under water-stress. Precipitation or moisture may also have indirect impacts on Xf through insect vectors whose activity or behavior could be altered by water stress83. Although the three subspecies of Xf seem to have different tolerances to cold, it is, however, unclear whether realized niche divergence among subspecies reflects inherent differences in thermal tolerances or rather host-pathogen interactions as it was observed for Ralstonia solanacearum84. Additional investigations would allow a better understanding of the effect of temperatures on the different strains of Xf. It is noteworthy that potential distributions show large areas of potential co-occurrence. This may have important implications as it may increase the risk of intersubspecific homologous recombination11.

Limits and opportunities for risk assessment

Maps of habitat suitability should be cautiously interpreted as they are derived from correlative tools that depict the realized niche of species i.e., a subset of the fundamental environmental tolerances constrained by biotic interactions, landscape structure and dispersal limits85. In addition, time-periods associated to occurrences and climate descriptors dataset do not perfectly overlap. The models were fitted with climate descriptors that represent average climate conditions for the 1970–2000 period, while some presence records were collected after 2000 in a period characterized by milder winter temperatures. Moreover, we deliberately fitted the models using a few climate descriptors to avoid model over-parameterization and/or extrapolation and enhance model transferability. Consequently, we cannot exclude that bioclimatic models presented here did not fully capture the entire range of environmental tolerances and did not fully depict the complexity of the climatic niche of Xf as well as potential interactions between climate descriptors. Better models and hence, better risk assessment could be obtained by collecting additional occurrence data as well as reliable absence data. The possible adaptation of Xf to new environmental constraints in its invaded range (e.g., by genetic recombination) is another important source of uncertainty. Finally, it is worth noting that bioclimatic models predict climatic suitability of a geographic region for Xf rather than a proper risk of Xf-induced disease incidence. To predict the proper severity of Xf-induced diseases in a given locality, statistical models should account for many additional factors playing a role in Xf epidemiology, including e.g., microclimate conditions, inter-annual climate variability, host-plant sensitivity, host-pathogen interactions, landscape structure and the spatio-temporal structure of the community of potential vectors. Although recent entomological studies identified the meadow spittlebug P. spumarius as the main vector of Xf in Italy79,80, a better knowledge of all European vectors capable of transmitting Xf to plants as well as their ecological characteristics (geographic range, efficiency in Xf transmission, demography, overwintering stage, intra-specific diversity, etc.) is needed86. In this context, habitat suitability maps could allow to design cost-efficient vector surveys, with priority given to geographic regions predicted as highly climatically suitable for Xf. The study by Cruaud et al.77 provides a good insight into how species distribution modeling and DNA sequencing approaches may be combined for an accurate monitoring of the range of Xf and its vectors in Europe. We believe that SDMs are valuable tools to help in designing research experiments, control strategies as well as political decisions at the European scale.

Conclusions/highlights

Species distribution models all indicate that the currently reported geographical range of Xf in Europe is small compared to the large extent of climatically suitable areas. This is true for all studied subspecies of Xf although the subspecies pauca appears to have a smaller potential range. Xf has a certain potential to adapt to climate and biotic conditions (hosts, vectors) encountered in Europe. The magnitude of this adaptive potential remains largely unknown but could nevertheless lead to a substantial spread of this plant pathogen across Europe. A further important research effort is thus needed to decipher the potential host plants – insect vectors – bacterium interactions in the (sub)natural ecosystems as well as agro-ecosystems at risk55. Only in this way could we develop an appropriate and efficient strategy to control Xf in the future.