Combined phylogenetic and geographic data can predict plant–pest interactions with high accuracy

Summary

• Non‐native plant pests can pose major threats to biodiversity, with destructive ecological and economic consequences. The ability to predict future threats would allow limited resources to be concentrated on managing the most serious risks.

• We built a Bayesian model to predict hosts at risk from Agrilus, a beetle genus of over 3000 species including one of the world's worst tree pests, using phylogenetic and geographic relationships between known and potential hosts.

• We assess risk to Quercus (oak), their most common host, by predicting the probability of over 7000 possible oak–Agrilus interactions to identify species at risk and inform future prevention efforts. Our model detects known hosts with 83.6% accuracy under Leave‐One‐Out cross‐validation, and successfully classifies novel hosts of Agrilus species in new areas, indicating strong predictive performance on independent or misclassified data. Geographic proximity is a strong predictor of host sharing, with the likelihood declining rapidly with distance. In general, hosts cluster phylogenetically, with a tendency for closely related oaks to share the same Agrilus species.

• Our approach uses readily available data and could be implemented to assess Agrilus interactions with other plant genera, and extended to additional host–pest systems to help prioritise countermeasures against threats world‐wide.

Schematic overview of the study pipeline.

Introduction

Throughout history, human activity has resulted in the breakdown of biogeographical barriers, prompting new biological invasions (Santini et al., ²⁰¹⁸). In the last 200 yr, the frequency of these invasions has been exacerbated by globalisation and international trade, resulting in an increase in the incidence and severity of damage by plant pests and diseases (Mack et al., ²⁰⁰⁰; Brockerhoff & Liebhold, ²⁰¹⁷; Seebens et al., ²⁰¹⁷; Santini et al., ²⁰¹⁸; Seebens et al., ²⁰²⁵).

Although most newly introduced organisms will fail to survive, establish, and spread, a small fraction become extremely damaging, leading to cascading ecological and economic repercussions (Mack et al., ²⁰⁰⁰; Aukema et al., ²⁰¹⁰; Brockerhoff & Liebhold, ²⁰¹⁷). This disproportionate effect of non‐native species may be explained by encounters with non‐coevolved hosts that lack effective defence responses and the presence of fewer natural enemies in novel ranges (Keane & Crawley, 2002; Gandhi & Herms, ²⁰¹⁰). Furthermore, when hosts are moved into new areas (such as through planting of tree species outside their native ranges), this can also bring them into contact with organisms with which they have not coevolved, providing further opportunity for novel interactions to arise (Pearse et al., ²⁰¹³; Pearse & Altermatt, ²⁰¹³).

Non‐native invaders, which can remain undetected and poorly understood for decades after their introduction in a new habitat (Mack et al., ²⁰⁰⁰; Roy et al., ²⁰¹⁴; Brockerhoff & Liebhold, ²⁰¹⁷; Liebhold et al., ²⁰¹⁷; Bebber et al., ²⁰¹⁹), are difficult and costly to eradicate once established (Liebhold & Tobin, 2007). Consequently, the most efficient way to address this issue is to prevent their arrival (Mack et al., ²⁰⁰⁰). These efforts often involve trade measures such as quarantine rules (Mack et al., ²⁰⁰⁰) and phytosanitary inspections at entry points and other vulnerable locations (Wittenberg & Cock, 2001; Panzavolta et al., ²⁰²¹). However, identifying threats in advance can be challenging (Pearse et al., ²⁰¹³; Pearse & Altermatt, ²⁰¹³) given the vast pool of organisms from which future risks could arise, and as the potential of species to cause harm often remains unrecognised until damage has already occurred (Mack et al., ²⁰⁰⁰; Liebhold et al., ²⁰¹⁷).

Some of the most prominent forest pests and pathogens involve non‐native species that have caused damage to entire ecosystems (Lowe et al., ²⁰⁰⁰; Aukema et al., ²⁰¹⁰; Brockerhoff & Liebhold, ²⁰¹⁷; Liebhold et al., ²⁰²³). Examples include Dutch elm disease (Karnosky, 1979), ash dieback (Enderle et al., ²⁰¹⁹), the spongy moth (Elkinton & Liebhold, 1990; Tobin et al., ²⁰¹²), and the emerald ash borer (Herms & McCullough, 2014). This last species – Agrilus planipennis (Coleoptera: Buprestidae) – is a beetle native to Asia, where it attacks its natural Fraxinus (ash) hosts only when they are already under stress (Liu et al., ²⁰⁰³; Wei et al., ²⁰⁰⁴). In North America, however, it has killed hundreds of millions of otherwise healthy ash trees following its accidental introduction (McCullough, 2020), causing > 99% mortality in some places (Klooster et al., ²⁰¹⁴) and constituting the most destructive and costly forest insect invasion of the United States to date (Herms & McCullough, 2014; Fei et al., ²⁰¹⁹).

Horizon scanning can help prevent similar outcomes to those noted above by narrowing the pool of taxa on which surveillance and mitigation efforts should focus, for example, by highlighting organisms that warrant targeted monitoring or restrictions on the movement of high‐risk materials (DEFRA, ²⁰²⁶; Peyton et al., ²⁰²⁵; Ribaya et al., ²⁰²⁵). This can, however, be a time‐consuming and labour‐intensive process (Antoniou et al., ²⁰²⁴), often requiring detailed evaluation of individual species (e.g. see the UK Plant Health Risk Register, DEFRA, ²⁰²⁶). Furthermore, these assessments tend to focus heavily on organisms with already documented impacts and are therefore likely to overlook novel threats (Hulme, 2025). Our intention is to address this challenge by providing a modelling‐based workflow that predicts highly likely plant–pest interactions, efficiently identifying pest candidates for in‐depth assessment and helping to direct resources towards those most likely to pose a threat.

Previous research has shown that there is often a phylogenetic signal in the host ranges of plant enemies, meaning that more closely related plants tend to share the same pests or pathogens (Parker & Gilbert, 2004; Pearse & Hipp, ²⁰⁰⁹; Yguel et al., ²⁰¹¹; Gilbert & Parker, ²⁰¹⁶). Consequently, models informed by phylogenetic relationships between known and potential hosts have been developed in efforts to predict novel biotic interactions to facilitate the assessment of risks to plant health (e.g. Gilbert & Parker, ²⁰¹⁶; Robles‐Fernández & Lira‐Noriega, ²⁰¹⁷). However, this single predictor may not always be sufficient to provide accurate results, particularly at finer phylogenetic scales (Cavender‐Bares et al., ²⁰⁰⁶) due, for example, to phenotypic convergence reducing the strength of the signal (Becerra, 1997; Pearse et al., ²⁰¹³). Therefore, models integrating additional variables, such as geographic, trait, and environmental data, have also been used to predict novel interactions in a range of taxonomic groups (Novotny et al., ²⁰¹⁰; Gilbert et al., ²⁰¹²; Pearse & Altermatt, ²⁰¹³; Albery et al., ²⁰²⁰; Farrell et al., ²⁰²²; Gougherty et al., ²⁰²⁴), or to estimate the probability of infection or damage from known pests and pathogens (Barrow et al., ²⁰¹⁹; Mech et al., ²⁰¹⁹). However, approaches incorporating multiple predictors can pose challenges in scalability and complexity, and information such as trait data can be incomplete or hard to compile (Albery et al., ²⁰²⁰). For example, Mech et al. (2019) excluded certain traits from their modelling of risks from herbivorous insects where data were unavailable for the majority of host species, and Barrow et al. (2019) had to infer trait status from related taxa in some cases when modelling parasite infections in birds. Moreover, Barrow et al. (2019) found species‐level predictors of host susceptibility, including environmental variables such as temperature and aridity, differed in importance depending on the genus of parasites under consideration. These examples emphasise the challenge of identifying predictors to supplement the phylogenetic signal that can be broadly applied.

Here, we develop a statistical model to predict potential novel host plant associations of Agrilus, in order to anticipate future threats from this major pest‐containing genus. Our aim is to develop a tractable approach that relies only on information that is readily available for many plant taxa – host phylogenetic relatedness and geographic overlap. We hypothesise that incorporating geographic overlap between known and potential novel hosts of Agrilus beetles into phylogenetically informed statistical models will improve their power to predict novel interactions because host jumps are more likely between plants occurring in closer physical proximity (Parker & Gilbert, 2004). Although host geographic overlap has been found to be predictive of pest and pathogen sharing between mammal species (Albery et al., ²⁰²⁰; Brennan et al., ²⁰²⁴), to the best of our knowledge, it has not been used to aid prediction of novel interactions between plants and their pests.

We use Agrilus as our test case since multiple Agrilus species, including the emerald ash borer, are of significant concern to plant health (Nielsen et al., ²⁰¹¹; Herms & McCullough, ²⁰¹⁴; Lopez et al., ²⁰¹⁴; Volkovitsh et al., ²⁰²⁰). With > 3000 described species world‐wide (Jendek & Grebennikov, 2023), the number of potential pest species that could emerge from within this genus alone is substantial. We focus on Agrilus species that exploit oaks (Quercus spp.), their most common host genus (Jendek & Poláková, 2014), and one of major ecological, cultural, and economic importance (Tantray et al., ²⁰¹⁷). We test for phylogenetic signal in Agrilus host use, and evaluate whether combining phylogenetic relatedness with geographic proximity yields accurate predictions for Quercus–Agrilus interactions. Our approach provides a straightforward framework for identifying potential novel hosts of Agrilus species and offers a starting point for targeted early‐detection efforts and for informing policies designed to reduce the risk to plant health. Moreover, it can be used for predicting novel interactions both at the establishment phase in the scenario of pests invading new areas, and where potential hosts have been introduced into the range of pests. While developed for Agrilus, our framework could be applied in other systems where trait data are sparse but phylogenetic and geographic information is accessible, allowing large numbers of species to be analysed simultaneously and enabling identification and prioritisation of potential pests for more in‐depth assessment.

Materials and Methods

See Fig. 1 for an overview of the methods.

Fig. 1.

Schematic overview of the study pipeline. We retrieved a Quercus phylogenetic tree (Hipp et al., ²⁰²⁰), produced a list of larval hosts of Agrilus beetles that utilise oaks (Vansteenkiste et al., ²⁰⁰⁴; Nielsen et al., ²⁰¹¹; Brown et al., ²⁰¹⁵; Cipollini, ²⁰¹⁵; Jendek, ²⁰¹⁶; Peterson & Cipollini, ²⁰¹⁷; Prokhorov & Vasilyeva, ²⁰¹⁷; Jendek & Nakládal, ^2019a,b,c; Volkovitsh et al., ²⁰²⁰) and downloaded plant geolocation information from the Global Biodiversity Information Facility (GBIF, ²⁰²²). We cleaned the data and then tested for the presence of a phylogenetic signal in our data using Fritz and Purvis' D statistic (Fritz & Purvis, 2010). Next, we computed several geographic distance metrics, compared them and subsequently built multiple binary logistic models (Bürkner, 2017) using different combinations of several explanatory variables to assess which have a significant effect on Agrilus host status. We employed the Leave‐One‐Out cross‐validation Information Criterion (LOOIC) (Vehtari et al., ²⁰¹⁷) for model selection, and used the chosen model to predict Agrilus host status for all Quercus species – Agrilus species pairs. Finally, we performed Leave‐One‐Out (LOO) analyses to assess the ability of our model to correctly identify known hosts. Min., minimum; No., number.

Assembly of data

Information on world‐wide Agrilus Curtis, 1825 host plants was extracted from an extensive critical review of the literature (Jendek & Poláková, 2014). These data were supplemented by records from subsequent publications (Vansteenkiste et al., ²⁰⁰⁴; Nielsen et al., ²⁰¹¹; Brown et al., ²⁰¹⁵; Cipollini, ²⁰¹⁵; Jendek, ²⁰¹⁶; Peterson & Cipollini, ²⁰¹⁷; Prokhorov & Vasilyeva, ²⁰¹⁷; Jendek & Nakládal, ^2019a,b,c; Volkovitsh et al., ²⁰²⁰) and formatted as a dataframe, with information available for a total of 241 Agrilus species. Plants were considered hosts only if there were records of beetle larvae exploiting them, as (1) larval host data are more reliable (since adults could be present on trees they do not exploit), and (2) it is the larvae that inflict significant host damage. For records from Jendek & Poláková (2014), we only retained those with a confidence index of 3 (highest confidence), at least one larval reference (marked as ‘!’), and no remarks indicating the data might be unreliable. All plant taxon names were standardised against the World Checklist of Vascular Plants (WCVP) v.3 (Govaerts et al., ²⁰²¹). For subsequent analyses, this dataset was restricted to Agrilus species that utilise Quercus L., retaining all confirmed larval host species reported within their native ranges, including non‐native hosts (Supporting Information Table S1). In the case of Agrilus auroguttatus, Agrilus bilineatus and Agrilus sulcicollis, all of which have been introduced outside their native range, any non‐native hosts only recorded in these new areas were removed. Our final dataset contained information for 32 Agrilus species that are known to use, in total, 51 Quercus species. Continent‐level native distribution data in Table S1 were inferred from Jendek & Grebennikov (2023) and GBIF (2025) for Agrilus species, and from WCVP v.10 distribution data (Govaerts et al., ²⁰²¹) for host taxa.

A time‐calibrated phylogenetic tree for Quercus (Hipp et al., ²⁰²⁰) was retrieved from https://github.com/andrew‐hipp/global‐oaks‐2019 (file tr.singletons.correlated.1.taxaGrepCrown.tre). It represents c. 250 Quercus species (> 50% of all recognised oak species in WCVP v.3), and was inferred using restriction‐site associated DNA sequencing and with fossil calibration (Hipp et al., ²⁰²⁰); we chose to use this tree as it was the most comprehensive time‐calibrated species‐level phylogenetic tree for Quercus available at the time of our analyses. Although Hipp et al. (2020) also included an equivalent tree using stem calibration, phylogenetically informed models of pest host use are robust to the precise choice of host phylogenetic tree (Pearse & Hipp, 2014; Uden et al., ²⁰²²). Species names in this tree were standardised against the WCVP v.3; intraspecific taxa were reduced to species‐level, and one tip per species was retained. One known Agrilus Quercus host (Q. gambelii Nutt.) is missing from this phylogenetic tree and was therefore removed from our dataset. The final modified tree contained 238 Quercus species, including two new, uncharacterised species, which were retained for the phylogenetic signal testing but excluded from the statistical modelling analyses as they lack geographic information.

Global geographic distribution records for Quercus species and any other known host taxa of the 32 Agrilus species were obtained from the Global Biodiversity Information Facility on June 19, 2022 (GBIF, ²⁰²²), using the rgbif R package v3.7.3 (Chamberlain et al., ²⁰⁰³). This dataset was cleaned, following recommendations from the CoordinateCleaner R documentation (Zizka et al., ²⁰¹⁹): we removed entries with unreliable coordinate information, flagged as ‘COORDINATE_INVALID’, ‘COORDINATE_OUT_OF_RANGE’, ‘COORDINATE_REPROJECTION_SUSPICIOUS’, ‘COORDINATE_UNCERTAINTY_METERS_INVALID’, ‘COUNTRY_COORDINATE_MISMATCH’, ‘PRESUMED_SWAPPED_COORDINATE’, ‘PRESUMED_NEGATED_LONGITUDE’, ‘PRESUMED_NEGATED_LATITUDE’, or ‘ZERO_COORDINATE’. We also excluded records with unusually high individual counts (> 10), from fossils, or predating 1950, since these may less accurately reflect current species occurrences (see ‘Improving data quality’ recommendations in https://cran.r‐project.org/web/packages/CoordinateCleaner/vignettes/Cleaning_GBIF_data_with_CoordinateCleaner.html, Zizka et al., ²⁰¹⁹). The set was further filtered using the clean_coordinates() function from the CoordinateCleaner R package v.2.0‐0 (Zizka et al., ²⁰¹⁹) (testing for ‘capitals’, ‘centroids’, ‘duplicates’, ‘equal’, ‘outliers’, ‘gbif’, ‘institutions’, ‘zeros’ and ‘seas’, and with outliers_method=‘distance’, capitals_rad=1000, inst_rad=1000, outliers_size=50, zeros_rad=1 and all other parameters set default). Due to memory constraints, species with > 100 000 georeferenced entries following cleaning were randomly subsampled to 100 000. Finally, a single manual coordinate entry was added for oak species in the phylogeny lacking georeferenced information (i.e. Q. sagrana Nutt. and Q. yiwuensis Y.C.Hsu & H.Wei Jen); their native ranges were checked via https://powo.science.kew.org, and the coordinates for their region centroid according to CoordinateCleaner were extracted.

Phylogenetic signal

We tested for the presence of a phylogenetic signal in host range using Fritz and Purvis' D statistic (Fritz & Purvis, 2010) – which can be used for binary traits – implemented in the R package caper v.1.0.1 (Orme et al., ²⁰²³), using 1000 permutations. To visualise this signal and verify the pattern, we looked for clades in the oak phylogeny with an over‐representation of Agrilus hosts (host‐rich clades, Fig. 2), using an approach originally developed for analyses of community structure (Webb et al., ²⁰⁰⁸) but since used to identify clades with overrepresented traits (see Saslis‐Lagoudakis et al., ²⁰¹² and Holzmeyer et al., ²⁰²⁰). To identify host‐rich clades, we used the nodesig function from the phylocom v.4.2 software (Webb et al., ²⁰⁰⁸), with 1 million randomisations and the default swap method (method 2). The input files were the Quercus phylogenetic tree (file tr.singletons.correlated.1.taxaGrepCrown.tre, noted above, but modified to remove bootstrap support values), a traits file with the type of trait set to binary (type 0), and a sample file with the first column indicating host status and all abundances set at 1.

Fig. 2.

Phylogenetic patterns of clustering of the Agrilus hosts across the Quercus phylogeny. Red and blue dots indicate nodes identified as having significantly more or less descendants that are hosts than expected by chance, respectively, based on the nodesig function from the phylocom software. Grey clades indicate no over‐ or under‐representation of hosts. We identified five main clades with an over‐representation (i.e. a significant degree of clustering) of Agrilus hosts (host‐rich clades, in red). One of these clades represented seven additional nested host‐rich clades (in section Quercus). Non‐hosts present in these clades could be potential future hosts at risk. There are also three host‐poor clades (in blue) where hosts are underrepresented. Non‐hosts present in these clades may be at a lower risk of future host jumps by Agrilus. However, host‐poor clades for Agrilus as a genus might still be host‐rich for particular Agrilus species (e.g. A. auroguttatus utilises three Quercus species present in three nested host‐poor clades). Dark grey barplots indicate the number of known Agrilus species hosted per Quercus species, and light grey circles indicate the number of cleaned GBIF entries (see the ‘Materials and Methods’ section) available for each. Stars indicate geographic regions for major Quercus clades according to (Kremer & Hipp, 2020), and numbers indicate sections (see key). It should be noted that this clustering analysis is scale‐dependent.

Statistical modelling

We built generalised linear mixed‐effects models (or multilevel models, in the terminology of the brms R package, Bürkner, ²⁰¹⁷) using binary logistic regressions to assess whether host status (i.e. a binary trait indicating whether an oak species hosts a given Agrilus species, where 0 s indicate oak species not recorded as a larval host of that particular Agrilus species in our dataset, and 1 s indicate oak species that are confirmed larval hosts; see the ‘Assembly of data’ in the Materials and Methods section) can be explained by the Quercus phylogeny and the geographic proximity to known Agrilus host plants.

Since no standard measure of geographic proximity between species exists for these purposes, we evaluated six metrics of spatial overlap between oak species and Agrilus hosts (see Notes S1 for their functional definitions), calculated from the GBIF geolocation records (GBIF, ²⁰²²). For our first two metrics, for each possible Quercus species–Agrilus species pair, we calculated the mean and median of the minimum distances (in km) between each record of the given oak species and any known plant host (i.e. including non‐Quercus hosts) of the given Agrilus species. If the focal Quercus species was a known host, the distance was computed to any other host species, and if it was the only known host, it was set to a large value of 30 000 km to represent the absence of alternative hosts. Distances were calculated using the geodist() function from the R package geodist v.0.0.7 (Padgham, 2021), with ‘measure’ set to ‘cheap’ due to time and memory constraints. For the next two metrics, we weighted distances within a plausible dispersal range (Webb et al., ²⁰²¹; Schrader et al., ²⁰²⁰) most heavily. Specifically, we transformed the mean or the median of the minimum distances to the corresponding normal probability density (μ = 0 km, σ = 50 km), such that shorter distances received higher weights. We then multiplied this value by −1, so that species close in space had lower scores, and rescaled the resulting value to range from 0 to 1 for ease of interpretation. For the final two metrics, the original distances were instead log‐transformed and then rescaled to between 0 and 1. We then built separate models using each of these six distance metrics in turn (see Notes S2) and compared them using a Leave‐One‐Out (LOO) analysis (Vehtari et al., ²⁰¹⁷), using the loo() function from the brms v.2.18.0 package, with moment_match = TRUE and reloo = TRUE to deal with high Pareto k diagnostic values (we note that only two out of 7552 observations had a high diagnostic value). The two normal‐transformed distances consistently outperformed the others (Table S2) and were selected as the best‐performing metrics for use in the subsequent analyses.

Additionally, we computed three phylogenetic metrics for each Quercus–Agrilus interaction using the time‐calibrated Quercus phylogenetic tree (see Notes S3 for their functional definitions). These were (1) the mean phylogenetic distance (where the distance to the root = 0.5) of the given oak species to (other) Quercus hosts of the given Agrilus beetle, capturing the overall phylogenetic breadth of its known hosts; (2) the minimum phylogenetic distance of the given oak species to any (other) hosts, capturing the proximity to the closest known host; or (3) the sum of both. If the current oak species was the only known host, distances were set to twice the distance to the root.

We then fitted binary logistic models with brms (Bürkner, 2017), using the brms() function with family=‘bernoulli’, iter=7000, cores=4, control=list(adapt_delta=0.995, max_treedepth=15), and default settings for all other parameters. We modelled host status as a binary response (0 for non‐hosts, 1 for hosts) and compared 132 models comprising all possible combinations of explanatory variables that included: (1) one of the two best‐performing geographic overlap metrics (fixed effect), (2) one of the three computed specific phylogenetic metrics (fixed effect), (3) a proxy for sampling effort (fixed effect), that is, number of GBIF records per oak species, as GBIF data are known to reflect this bias (Beck et al., ²⁰¹⁴; Maldonado et al., ²⁰¹⁵), (4) a general Quercus phylogenetic random effect (with one level per oak species, so that oaks closely related to species that host Agrilus have a greater chance of being hosts) and (5) an Agrilus species random effect. The simplest model was an intercept‐only null model (Table S3). For those models with a phylogenetic random effect, this was included as a varying intercept over oak species, which were correlated as specified by a phylogenetic covariance matrix (see https://cran.r‐project.org/web/packages/brms/vignettes/brms_phylogenetics.html, Bürkner, 2017). To test whether the influence of host relatedness varies among Agrilus species, models including both a fixed phylogenetic variable and a random Agrilus species effect were fitted in two ways (Table S3): (1) keeping the original Agrilus species random effect (intercept‐only), and (2) transforming it into a random‐slope model to allow the strength of the phylogenetic effect to vary among Agrilus species. This reflects the hypothesis that generalist species are less constrained by host‐relatedness than specialists.

To identify the best‐fit models, we employed a model comparison approach using the LOO Cross‐Validation Information Criterion (LOOIC; Vehtari et al., ²⁰¹⁷), using the brms loo() function as specified above. Due to memory limitations, the models were divided into three groups for initial comparison, and the best‐performing models from each group were then compared against each other to identify the best‐performing model overall (Tables S4–S12). This final model was:

$$\begin{array}{c}{y}_{\mathit{ij}}~\text{Bernoulli}\left(p_{\mathit{ij}}\right)\\ \text{logit}\left(p_{\mathit{ij}}\right)={\alpha }_0+{\alpha }_1{x}_{1\mathit{ij}}+{\alpha }_2{x}_{2\mathit{ij}}+{\mu }_i+e_{\mathit{ij}}\\ {\mu }_i~N\left(0,\Sigma \right);e_{\mathit{ij}}~N\left(0,{\sigma }_e^2\right)\end{array}$$ (Eqn 1)

$y_{\mathit{ij}}ϵ\left\{0,1\right\}$ indicates if Quercus species $i$ can host the Agrilus species $j$. The probability $p_{\mathit{ij}}$ is a function of the following variables. The value ${\alpha }_0$ is the intercept, $x_{1\mathit{ij}}$ is the normal‐transformation of the mean minimum geographic distance from the Quercus species $i$ to any (other) known host of Agrilus species $j$, and ${\alpha }_1$ is the associated coefficient, $x_{2\mathit{ij}}$ is the minimum phylogenetic distance of the Quercus species $i$ to any (other) known oak host of the Agrilus species $j$, and ${\alpha }_2$ is the associated coefficient, ${\mu }_i$ is the phylogenetic random effect with its covariance matrix $\Sigma$ specified by the Quercus phylogeny and an additional parameter ${\sigma }^2$, and $e_{\mathit{ij}}$ is the normally distributed residual error.

Predicting host–Agrilus interactions

We used the model in Eqn 1 to compute predictions for the response variable (i.e. host status for each Quercus species – Agrilus species pair) with the brms fitted() function (using scale=‘linear’ to maintain the log‐odds scale). To classify which Quercus species might be particularly at risk, we also computed a binary threshold to divide predictions into 0 s (i.e. non‐hosts) and 1 s (i.e. hosts). This cut‐off was selected as the point that maximises sensitivity and specificity (i.e. where their functions intercept, see Fig. S1).

To evaluate the model performance, we compared two data limitation scenarios: one in which we falsely assume there is no interaction (misclassified data), the second in which no information is available for the given interaction (cross‐validation using independent data). For the first, we ran a deliberately conservative analysis where, in turn, each known host–Agrilus interaction (i.e. coded as 1 in the original model) was intentionally misclassified as a 0, and the value for this observation was predicted (using fitted() as above). For the second case, we incorporated the absence of knowledge about the interaction using the brms loo_linpred() function (with type = ‘mean’), which draws from the LOO distribution using Pareto‐Smoothed Importance Sampling. This procedure uses the model's posterior distribution as the proposal distribution for importance sampling (Vehtari et al., ²⁰¹⁷). In addition, since k‐fold cross‐validation is a commonly used approach for testing model performance, we also compared the predictions of the second LOO method against those obtained from fivefold cross‐validation. The fivefold analysis was performed using brms functions kfold() (using save_fits=TRUE) and kfold_predict() (using method=‘posterior_epred’). The predictions for the missing observations were then averaged across posterior draws and transformed to the log‐odds scale for comparison.

All analyses were performed in R v.4.2.0.

Results

Our final dataset includes 32 Agrilus species from Africa, Asia, Europe, North and South America, with 40.63% (13) of species recorded from a single continent and 46.88% (15) found in all five (Table S1). Overall, 56.25% (18) of these Agrilus species have Quercus as their only reported hosts (Table S1). Host breadth averages 4.53 plant species (range = 1–15) and 2.03 genera (range = 1–9) per Agrilus species (Table S1). The average number of hosted Agrilus species for the 50 Quercus hosts present in the phylogenetic tree is 2.32 (range = 1–8; Fig. 2; Table S1).

Phylogenetic signal in host status

Agrilus beetle hosts are non‐randomly distributed across the Quercus phylogeny (Hipp et al., ²⁰²⁰), with several clades showing strong clustering according to the phylocom analysis (Fig. 2). On average, for each Agrilus species, there are 3.63 reported Quercus host species (range = 1–10), and 232.38 non‐host Quercus species (range = 226–235) in our dataset. Hosts are overrepresented in section Cerris (with the entire section identified as a host‐rich clade), as well as in one clade of section Ilex and two of section Lobatae (Fig. 2). Clades where Agrilus hosts are underrepresented (host‐poor clades) are found within section Cyclobalanopsis, section Quercus and interestingly, given the presence of two host‐rich clades, also within section Lobatae (Fig. 2). These findings are consistent with the presence of a moderately strong overall phylogenetic signal in host status indicated by Fritz and Purvis' D (D = 0.56 Fritz & Purvis, ²⁰¹⁰; Fig. 3a), where the distribution differs significantly from both a Brownian expectation (P = 0.001) and a random one (P < 2e‐16; Fig. 3a).

Fig. 3.

Phylogenetic signal, and posterior mean estimates and 95% credible intervals (CI) for the chosen model, that is model 43. (a) Fritz and Purvis' D statistic for the distribution of Agrilus hosts in the Quercus phylogeny. We found significant levels of clumping of Agrilus hosts, with the phylogenetic signal being moderately strong (D = 0.56, dotted line). Our signal differs significantly from (i.e. does not overlap with) both the Brownian (pink distribution) and the random expectation (blue). (b) Posterior mean estimates and 95% CI of predictors and random effects on host status for model 43. In this binomial model, related Quercus (oak) species have a similar probability of hosting Agrilus, that is, the variation among the ${\mu }_i$ (phylogenetic random effect) is phylogenetically correlated (intercept σ = 0.48, CI = 0.31–0.69). This means that species have different baseline probabilities of being hosts, with more closely related species having more similar values. See Supporting Information Table S13 for the group‐level (i.e. random) effects of each level (Quercus species). The estimates of α ₂ (coefficient of the species‐specific phylogenetic effect) show that closely related oak species are more likely to be hosts of a particular Agrilus species (estimate = −1.51, 95% CI = −2.30 to −0.75). The estimates of α ₁ (coefficient of the geographic distance effect) show that oak species spatially close to hosts of a given Agrilus species, j, are more likely to also be hosts (estimate = −3.61, CI = −4.39 to −2.90). Given that all CIs are narrow, the effect of these variables appears to be strong, albeit with the effect of the (fixed) phylogenetic distance being the least pronounced (Fig. S2). The negative overall intercept (−3.16, CI = −5.82 to −0.93) indicates that the naive probability of any given random Quercus species being a host is low. This model did not perform significantly worse than those that included a random effect for Agrilus species or a proxy for sampling effort, indicating that this simpler model can be applied to predict hosts for Agrilus species with limited available information.

Variable predictors of Agrilus host status

Among the 132 candidate models (see Table 1 for a summary and Tables S4–S12 for full statistics and comparisons), our selected model was the one best supported by the LOOIC comparisons (Table S8; see the ‘Materials and Methods’ section). Only one other best‐performing model (model 44, Table 1) is equally simple; the remaining 11 models are more complex (Tables 1, S8).

Table 1.

Comparison of selected computed statistical models explaining Agrilus host status of Quercus species.

Model number	Variables	LOOIC (SE)	Δ ELPD (SE)
43	host.status ~ geomean + phylomin + (1 | group)	741 (61.96)	0.00 (0.00)
44	host.status ~ geomean + phylomean + (1 | group)	750 (61.29)	−4.38 (3.22)
61	host.status ~ geomean + GBIF + (1 | group)	760 (61.64)	−9.41 (5.59)
65	host.status ~ geomean + (1 | group) + (1 | agrilus)	758 (61.47)	−8.59 (4.59)
83	host.status ~ geomean + phylomin + GBIF + (1 | group)	743 (62.54)	−0.97 (2.97)
84	host.status ~ geomean + phylomean + GBIF + (1 | group)	751 (61.99)	−5.17 (4.32)
101	host.status ~ geomean + phylomin + (1 | group) + (1 | agrilus)	742 (62.09)	−0.52 (0.62)
103	host.status ~ geomean + phylomean + (1 | group) + (1 | agrilus)	746 (61.80)	−2.61 (2.7)
108	host.status ~ geomedian + phylomin + (1 | group) + (phylomin | agrilus)	754 (62.92)	6.37 (7.36)
113	host.status ~ geomean + GBIF + (1 | group) + (1 | agrilus)	757 (61.84)	−8.12 (5.28)
121	host.status ~ geomean + phylomin + GBIF + (1 | group) + (1 | agrilus)	745 (62.67)	−1.95 (3.00)
123	host.status ~ geomean + phylomean + GBIF + (1 | group) + (1 | agrilus)	748 (62.37)	−3.58 (3.88)
128	host.status ~ geomedian + phylomin + GBIF + (1 | group) + (phylomin | agrilus)	753 (63.37)	−6.04 (8.00)

132 binary logistic models with different variable combinations were computed, divided into three groups, and compared (see Supporting Information Tables S4–S8 for full comparisons, and see the ‘Materials and Methods’ section for more information). The best‐fitting models from each group were then compared with one another; results are shown here (excluding models with non‐significant predictors). Note that $\text{LOOIC}$ (Leave‐One‐Out cross‐validation Information Criterion) $=-2\times \text{ELPD}$ (Expected Log Predictive Density). Key: geo = normal‐weighted geographic distance metric (either ‘mean’ or ‘median’, see the ‘Materials and Methods’ section); phylo = phylogenetic effect specific for each Quercus–Agrilus interaction (either ‘minimum’ or ‘mean’, see the ‘Materials and Methods’ section); GBIF = number of GBIF records per oak species (proxy for sampling effort); group=gr(quercus.sp, cov=oak_phylo_cov), that is, the grouping variable of the general oak phylogenetic random effect; agrilus = Agrilus species. SE, standard error.

Our final model, Eqn 1, reveals that Agrilus beetle host use in Quercus is governed by both phylogenetic and spatial structure (Fig. 3b). Host sharing is strongly influenced by geographic proximity (Figs 3, S2): oak species are more likely to host a given Agrilus species (j) if they grow near another host. This spatial effect, quantified by α ₁, declines rapidly with distance, with the best‐performing metric for $x_{1\mathit{ij}}$ being the normal‐transformed distance (σ = 50 km). Model comparison using LOOIC confirms the importance of this effect, as models including α₁ consistently outperform those without it, that is, with α ₁ = 0 (e.g. model 2 vs 1 and 121 vs 115; Tables S9, S11).

The model also captures two distinct phylogenetic effects on host status. First, there is a strong general tendency for Agrilus hosts, regardless of which Agrilus species they host, to be phylogenetically clustered (Figs 3, S2). This is reflected in the phylogenetic correlation of the variation among the ${\mu }_i$ values (Fig. 3), and supported by LOOIC comparisons showing that models with this random phylogenetic effect outperform those without (e.g. model 8 vs 1 or 43 vs 10; Table S9; see Table S13 for the group‐level – that is, random – effects by Quercus species). Second, closely related oaks also tend to host the same particular Agrilus species (j), an effect quantified by α ₂. Although smaller than the spatial effect α ₁ (Figs 3, S2), this signal is still important: models including α ₂ outperform those excluding it (e.g. model 1 vs 4 or 121 vs 113; Tables S9, S11). The best‐performing metric for this effect is the minimum phylogenetic distance to another known host ($x_{2\mathit{ij}}$).

The phytophagy level of each Agrilus species – that is, the number of oak species it exploits – is already encapsulated by the spatial and phylogenetic terms (α ₁ and α ₂), since models that include an Agrilus species random effect for species‐level prevalence perform no better (e.g. model 9 vs 1 or 20 vs 2; Table S9), or have very small effect sizes (Tables 1, S3; Fig. S3). Similarly, including a proxy for sampling effort has little impact on explanatory power (e.g. model 7 vs 1 or 88 vs 48; Tables S9, S10), or has very small effect sizes (Tables 1, S3).

Predicting Agrilus host status

Our chosen model predicts Quercus–Agrilus interactions effectively (Figs 4, S4, S5), with the log‐odds scores for observed interactions being much higher (median log‐odds = −1.35) than those for the rest (median log‐odds = −8.05; Fig. 4a). To characterise the model's performance, we divided this continuous distribution into predicted positive interactions (meaning that the Agrilus beetle j is predicted to be able to exploit the oak species i) and negative interactions (it is not), using our binary threshold (log‐odds = −4.15; see the ‘Materials and Methods’ section). Of the known interactions, 90.5% are correctly predicted to be positive; 90.6% of the interactions scored as 0 s for model fitting are predicted to be negative (Fig. 4b). We hypothesise that the lower left‐hand quadrant of Fig. 4b (predicted positive, but appeared negative in the training dataset, i.e. ‘false positives’) includes Quercus–Agrilus interactions that either occur but have not yet been reported, or could occur in the future (these interactions include some recently recorded in Agrilus invasive ranges that were coded as negative for model training; see the ‘Model performance’ in the Discussion section). Consequently, they identify potential hosts at risk (Table S14). This higher risk category includes a total of 696 interactions, encompassing all 32 Agrilus species and 105 Quercus species (including all 50 known hosts in our dataset).

Fig. 4.

Continuous and binary predictions for the chosen model. (a) Continuous predictions (log‐odds) of Agrilus host status for each Quercus species – Agrilus species pair. The x‐axis represents the observed host status for each Quercus species – Agrilus species pair, and the y‐axis shows the log‐odds of the model prediction, with higher values indicating a greater probability of a ‘positive’ interaction. Violin plots show the distribution density of predicted values; boxplots show the median and interquartile range (IQR), with whiskers extending to 1.5 times the IQR, and points indicating outliers. The dotted line indicates the threshold used in (b) to divide predictions into predicted non‐hosts and predicted hosts (see the ‘Materials and Methods’ section). (b) Binary predictions of Agrilus host status for each Quercus species – Agrilus species pair. Numbers indicate the number of Quercus species – Agrilus species pairs that belong to each quadrant. 90.52% of known Quercus species – Agrilus species interactions are predicted as such, and 90.64% of unobserved interactions are predicted as negatives. We hypothesise that the lower left‐hand quadrant (696 Quercus species – Agrilus species pairs in total) might indicate future potential Quercus–Agrilus interactions and thus, potential future hosts at risk (Supporting Information Table S14). (c–e) Continuous predictions for A. auroguttatus, A. bilineatus and A. sulcicollis, respectively. The dotted line represents the binary threshold. Pink labelled points indicate oak species that have been reported as new hosts in the introduced ranges of these beetles (scored as nonhosts for the purposes of our model).

To evaluate the model's performance when information about a particular interaction is wrong or missing, we conducted two types of LOO analyses. In the first (Fig. 5a), known positive interactions were coded as 0s, assessing the ability of our model to correctly identify known hosts when they are incorrectly scored as non‐hosts; this effectively simulates currently unrecognised host species. In the second analysis (Fig. 5b), missing interactions were incorporated using approximate LOO predictions, estimating model predictions as if each interaction was missing from the training data.

Fig. 5.

Leave‐One‐Out (LOO) predictions. Only shown for the 116 known interactions (i.e. Quercus–Agrilus pairs where the oak species is known to host the given beetle) in our dataset. (a) Predictions from the more stringent analysis where each 1 (i.e. known ‘positive’ interaction) was converted into a 0 at a time, against the original predictions from the model. (b) Predictions from the analysis where Leave‐One‐Out estimates were used to assess the effect of missing interactions, against the original predictions from the model. (c) Predictions from the second LOO analysis (shown in b) against predictions from the first one (shown in a)). The dashed grey line indicates the 1 : 1 line and the blue points indicate values that are over the binary threshold on one axis (i.e. in the ‘original predictions’ in the case of (a) and (b)) but not the other; there are none in the case of (c), illustrating that these approaches serve as good proxies for each other.

Both analyses yield similar results, with individual log‐odds prediction values showing a high degree of agreement (Fig. 5c). This indicates the second approach, which is less conservative and computationally intensive, would be sufficient for this assessment of model performance. In both cases, the model correctly classifies the oak as a potential host in 83.6% of instances, suggesting the model could be used to identify host species at risk from Agrilus, even where knowledge of existing host ranges is incomplete. Only eight interactions are predicted as positive in the original analysis but as 0 in the LOO analyses (Fig. 5a,b). These are: Q. castaneifolia – A. graminis, Q. arizonica – A. parkeri, Q. prinoides – A. obsoletoguttatus, Q. ilicifolia – A. arcuatus, Q. durata – A. angelicus, Q. shumardii – A. quercicola, Q. marilandica – A. geminatus, and Q. acutissima – A. ussuricola. Finally, we compared the second LOO approach against a fivefold cross‐validation analysis (Fig. S6). Predictions were highly consistent, with strong Pearson correlations (r = 0.93 across all values and r = 0.97 for known positives), further supporting the robustness of our modelling approach to missing interactions.

Discussion

In this study, we have investigated whether current patterns in the association between species of Agrilus beetles and Quercus trees can be used to predict future novel interactions, which have the potential to be severely damaging to the newly acquired hosts. Such interactions have recently arisen, as in the case of the emerald ash borer (A. planipennis) (Herms & McCullough, 2014), which is decimating Fraxinus (ash) populations in North America, and A. mali, which is attacking apple trees in western China (Bozorov et al., ²⁰¹⁹). These examples emphasise the need for effective approaches to predict the next major threat, so that mitigation can be put in place.

The importance of phylogenetic and spatial variables in predicting future interactions

In the past, efforts to disentangle host–enemy dynamics have often integrated phylogenetic and trait data. For instance, Mech et al. (2019) devised a logistic regression model for predicting the future risk posed by non‐native insects to North American conifers, and Barrow et al. (2019) employed host environmental, trait, and phylogenetic information to construct phylogenetic mixed‐effects models (using brms, Bürkner, ²⁰¹⁷) to predict parasite infection of birds. While both approaches successfully modelled host–enemy interactions, our model provides accurate predictions while maintaining simplicity, as it incorporates only three variables and does not rely on trait data, which can be arduous to compile (albeit some trait databases are available, e.g. Kattge et al., ²⁰²⁰). Instead, our mixed‐effects model uses relatively simple predictors of host status, based on readily available phylogenetic and geographic information which are typically more centralised, accessible and generalisable.

In our model, geographic range overlap is a strong predictor of host use, supporting the idea that ecological opportunity plays a role in shaping host–pest associations. An oak species is more likely to host a given Agrilus species if at least part of its distribution overlaps with that of another host, as repeated exposure to a pest increases the likelihood of host jumps (Gilbert & Parker, 2016). This variable also likely captures other relevant patterns, such as beetle polyphagy; a generalist Agrilus species is more likely to be present on nearby trees, which may explain why the model performed well without a species‐level random effect (see the ‘Variable predictors of Agrilus host status’ in the Results section). Range overlap may also reflect ecological similarities between plants with similar distributions, which can lead to convergent traits and shared climatic tolerances (Losos, 2011; Givnish, ²⁰¹⁶; Wang et al., ²⁰²⁵), partly substituting for more explicit environmental niche modelling. This single predictor, therefore, encapsulates multiple ecological drivers of host use, allowing for a simple model that is both efficient and interpretable.

The two distinct phylogenetic effects captured in our model are consistent with the presence of a phylogenetic signal in the host range of Agrilus species that use oaks (see the ‘Phylogenetic signal in host status’ in the Results section). This evidence for phylogenetic signal aligns with that observed in other plant–enemy interactions (Parker & Gilbert, 2004; Gilbert & Webb, ²⁰⁰⁷; Pearse & Hipp, ²⁰⁰⁹; Yguel et al., ²⁰¹¹; Gilbert et al., ²⁰¹²; Pearse & Altermatt, ²⁰¹³; Barrow et al., ²⁰¹⁹; Mech et al., ²⁰¹⁹; Cirtwill et al., ²⁰²⁰; Pearse & Rosenheim, ²⁰²⁰; Lynch et al., ²⁰²¹), including in Quercus herbivores (Pearse & Hipp, 2009), consistent with the idea that many host traits are phylogenetically conserved (Gilbert & Parker, 2016). Many of these studies also make use of phylogenetic information for predictive purposes. For example, Pearse & Altermatt (2013) used a simple generalised linear model to predict associations between German Lepidoptera and introduced plant species, using host breath information and minimum phylogenetic distance of each plant species to any known host of each lepidopteran. They built their model using information on native interactions and used it to predict values for non‐native species, validating the results with known host status (Pearse & Altermatt, 2013). Both the presence and strength of this phylogenetic signal are, however, scale‐dependent, with larger scales expected to result in stronger signals (Kamilar & Cooper, 2013; Pearse et al., ²⁰¹³). For example, Cavender‐Bares et al. (2006) proposed that the likelihood of observing phylogenetic clustering was higher in broadly‐defined communities; they found that functional traits of plant communities in Florida (US) tended to be conserved at broad phylogenetic scales (woody plants), but failed to find this same pattern when focusing on a single lineage (Quercus) (Cavender‐Bares et al., ²⁰⁰⁶).

In our model, the general phylogenetic tendency included as a random effect reflects a baseline preference of Agrilus to use closely related oak species, with a bias towards certain Quercus clades (Fig. 2). This variable might also capture ascertainment bias – as oak species which have attracted more observations seem to be non‐randomly distributed through the phylogenetic tree (see host‐rich clades in Fig. 2) – and may partly explain why our proxy for sampling effort made no discernible contribution to the model (see the ‘Variable predictors of Agrilus host status’ in the Results section). By contrast, the Agrilus species‐specific phylogenetic signal makes the weakest contribution to model predictions, likely because the dataset is already constrained to Agrilus that use oaks. As such, the general random effect is probably capturing most of the phylogenetic information; if a Quercus species is a common host, other Agrilus species that exploit oaks might also be able to utilise it. Part of this effect might also be captured by the geographic metric (consistent with Tobler's law; Tobler, ¹⁹⁷⁰), as closely related species tend to occur nearer each other due to shared evolutionary history and habitat filtering (Gilbert & Parker, 2016). In fact, the oak genus itself is divided into geographically structured clades (Kremer & Hipp, 2020), meaning that more closely related hosts are likely to also be closer in space.

Previous studies have also sought to combine phylogenetic and geographic data when forecasting biotic interactions. For example, Robles‐Fernández & Lira‐Noriega (2017) used phylogenetic data to predict hosts of ambrosia beetles and subsequently looked at host distribution to highlight geographical hotspots of interactions, whereas Albery et al. (2020) built models based on phylogenetic proximity and geographic overlap, amongst other variables, to predict viral sharing between mammal hosts. However, to the best of our knowledge, our efforts represent the first attempt to integrate phylogenetic and geographic information to build simple and robust models for predicting future interactions within a plant–pest network. Given that phylogenetic data and occurrence records are already available for many taxa, together with a growing number of examples for phylogenetic signal in host ranges, we believe our approach could prove useful for predicting future interactions in many other systems.

Model performance

Our model (Fig. 4) successfully classifies Quercus–Agrilus interactions with high accuracy, predicting 90.5% of known interactions and achieving 83.6% accuracy in LOO analyses. Notably, the lower left quadrant of Fig. 4b comprises 696 interactions not currently reported within the beetles' native ranges. While these predictions represent false positives with respect to the input data, our validation analyses indicate the model can detect plausible interactions even when they are recorded as negative. As such, we believe this quadrant captures both overlooked and potential future interactions (Table S14).

Three Agrilus species in our dataset, that is A. bilineatus, A. sulcicollis and A. auroguttatus, were recently introduced to areas outside their native ranges, where they have been observed exploiting novel Quercus hosts (Jendek & Grebennikov, 2009; Coleman & Seybold, ²⁰¹¹; Coleman et al., ²⁰¹²; Petrice & Haack, ²⁰¹⁴; Haack & Petrice, ²⁰¹⁹). These were treated as non‐hosts in our analyses to test predictive capacity (Fig. 4c–e). The model successfully predicted 50% of these associations. Agrilus bilineatus and A. sulcicollis both have their recently recorded hosts (Quercus robur and Quercus rubra, respectively) ranked as the top novel prediction (Fig. 4d–e). Agrilus auroguttatus shows a more complex pattern; only one of four reported novel hosts (Quercus kelloggii) exceeds the binary threshold, while another (Quercus agrifolia) falls just below it (Fig. 4c). This reflects the fact that, while all four novel hosts are geographically distant from the known hosts of this beetle, Q. kelloggii and Q. agrifolia (both within host‐rich clades, Fig. 2) are phylogenetically more closely related to other Agrilus hosts (Fig. S7). Interestingly, A. auroguttatus is the only one of these three beetles with a known host (Q. peduncularis) under the prediction threshold (Fig. 4c). In total, only ten Agrilus species have a known interaction under this threshold (Fig. 4b), typically involving single‐host species or phylogenetically and geographically isolated oaks (Fig. S7). This suggests the model may be less reliable in cases of unusual host use patterns, cryptic diversity, or taxonomic uncertainty (e.g. A. auroguttatus was historically synonymised with A. coxalis; Hespenheide, ¹⁹⁷⁹; Coleman et al., ²⁰¹²).

The geographic distribution of the predicted hosts largely mirrors the overall distribution of Quercus, and that of known Quercus hosts (Fig. S8). This is logical, as these regions provide a greater number of species that might act as hosts and more opportunity for novel interactions to arise (see Fig. S8c–f for maps showing the number of known and predicted hosts per region). Phylogenetically, predictions broadly (but not entirely) reflect known host‐rich clades (Fig. S9). This suggests that, while detecting host‐rich clades is a helpful first step, this approach will miss some vulnerable species.

Implications of model predictions

Individual Quercus and Agrilus species involved in many predicted novel interactions may be of particular interest when considering where best to focus efforts to mitigate future risks. Two oak species, Q. acutissima (with two known interactions, Table S1) and Q. rubra (with six), stand out as potentially suitable hosts for all 32 Agrilus species. While our model might be overestimating their hosting ability, these species seem strong candidates for general hosts, with wide distributions (when considering both native and introduced ranges; GBIF, ²⁰²²) – and thus geographically close to other hosts – and falling within host‐rich phylogenetic clades. Nevertheless, Agrilus species may vary in their ability to exploit them. For example, introduced species lacking natural enemies and escaping host defences (Brockerhoff & Liebhold, 2017) might be more aggressive and able to take over the niche (Work & McCullough, 2000; Fortuna et al., ²⁰²²). Alternatively, Agrilus species already established might take precedence over non‐aggressive newcomers. Nevertheless, co‐occurrence of multiple Agrilus species on a single host may be enabled by preferences for attacking different parts of trees (e.g. while most Agrilus larvae tend to girdle through the trunk, A. angelicus prefers young twigs; Swiecki & Bernhardt, ²⁰⁰⁶), or trees at different stages (e.g. Agrilus hastulifer predominantly utilises dead trees, DEFRA, ²⁰²⁶, and Agrilus horni prefers young suckers, Nord et al., ¹⁹⁶⁵). In addition to Q. acutissima and Q. rubra, a further 26 Quercus species are predicted hosts for ten or more different species of Agrilus (Table S14). Likewise, there are some Agrilus species that are involved in a very large number of predicted novel interactions. For example, A. obsoletoguttatus is predicted to have 52 novel potential oak hosts. Interestingly, with ten reported host species and nine different host genera (Jendek & Poláková, 2014), it is, in fact, a generalist beetle capable of using a diverse range of hosts. While this species has not been reported causing harm to hosts in the past, it has also not yet escaped its native range (GBIF, ²⁰²⁵). Two further Agrilus species are predicted to have more than 40 novel oak hosts – A. graminis and A. angustulus (Table S14). In common with A. obsoletoguttatus, both species have numerous known hosts, including multiple oak species (nine in the case of A. graminis and 10 for A. angustulus; Table S1), and thus their polyphagous nature is already documented.

Limitations of the approach

We believe our model provides a robust framework for predicting future interactions between Agrilus beetles and their host plants, which can be extended to other enemy–host systems. However, some limitations and considerations must be acknowledged. Perhaps most obviously, it relies on the availability of phylogenetic information for the host taxa, which may not always be present. However, as more comprehensive datasets become accessible (e.g. Zuntini et al., ²⁰²⁴), use of phylogenetically informed models, such as ours, is increasingly feasible. In addition, while the LOO cross‐validation analysis demonstrates our model can infer unreported interactions, systematic geographic biases in host reporting (Jendek & Poláková, 2014) may still influence the results. As mentioned above, we expect the bias towards some phylogenetic clades (Fig. 2) to partly reflect ascertainment bias. Furthermore, while the overall distribution of oak species broadly agrees with that of hosts (Fig. S8), China – where many oak species are present but few hosts predicted – is a notable exception, suggesting our model may be underpredicting interactions for this region.

Whether a species actually becomes a host will depend on factors additional to those accounted for in our model. The probability of a host jump will depend on ecological factors such as plant relative abundance (Gilbert & Parker, 2016) – which influences the likelihood of repeated encounters – abiotic plant stress (Chamorro et al., ²⁰¹⁵) – which can alter the individual tree susceptibility, for example (Showalter et al., ²⁰¹⁸) – niche availability and resource competition (Gidoin et al., ²⁰¹⁵), or community structure (Gougherty & Davies, 2021). Pest spread can be facilitated in communities with closely related hosts (Gougherty & Davies, 2021), although polyphagous and monophagous pests may respond differently to habitat heterogeneity (with the former preferring species‐rich habitats) (Brockerhoff et al., ²⁰⁰⁶). Agrilus species with different diet breadths could therefore thrive in different environments, even if they share some hosts. Still, most Agrilus species (c. two‐thirds) are only known to exploit one plant genus (Jendek & Poláková, 2014), so they may perform better in homogeneous habitats. In fact, Dutch elm disease led to increased Fraxinus (ash) dominance in eastern North American forests (Gandhi & Herms, 2010; Barnes, ²⁰¹¹), which are now suffering from the A. planipennis (emerald ash borer) invasion (Gandhi & Herms, 2010). Moreover, some potential hosts may only occur in areas with environmental conditions otherwise unsuitable for pest survival, preventing establishment. Finally, regions with high trade activity may face elevated risks of species introductions. In future work, trade information could be used to complement our model by identifying routes or areas at high risk, similar to Montgomery et al. (2023), who integrated global trade connectivity, alongside other variables, into a stochastic network model to estimate the probability of pest and pathogen entry and establishment. In sum, we do not anticipate every predicted interaction (Table S14) will materialise, as we expect a gap between potential and realised host ranges (Braga & Janz, 2021). Rather, our approach is intended to prioritise likely interactions for further assessment (e.g. ground‐truthing) before translating these results into policy actions.

Even where predicted interactions do materialise, not all potential hosts will be equally affected. There is evidence of a conserved signal in host vulnerability for plant–pest interactions (Gilbert et al., ²⁰¹⁵), meaning that only a subset of host species will suffer severe damage. Unfortunately, for Agrilus and many other genera, data on the severity of impact are much less readily available than that of host status. Nonetheless, hosts that are not severely damaged themselves can still facilitate transmission to others. For example, alternative hosts have been reported to act as enemy reservoirs that increase the risk of transmission to host crops (Saeed et al., ²⁰¹⁵; Ocimati et al., ²⁰¹⁸).

Finally, hosts currently within the native ranges of Agrilus beetles that do not seem to be heavily impacted could suffer increased damage in the future. Agrilus beetles are known to cause greater damage during periods of abiotic stress to hosts (Wei et al., ²⁰⁰⁴; Haack & Petrice, ²⁰¹⁹) and when in association with other biotic agents (Brown et al., ²⁰¹⁵). Climate change may alter host distribution and ecosystem composition (Chen et al., ²⁰¹¹), potentially leading to associations with other biotic stressors or facilitating novel interactions (Lancaster, 2020). It may prompt the planting of exotic trees considered to be better adapted to future conditions (Holl & Brancalion, 2020), inadvertently bringing highly susceptible hosts into contact with native pests (Dang et al., ²⁰²²). In fact, it has been proposed that global warming might have already prompted the apparent expansion of A. biguttatus in the UK (Brown et al., ²⁰¹⁵). All these processes could result in native beetles causing significant damage in their current range, which in turn may facilitate their spread and interaction with naïve hosts.

In this study, we have set out to assess the risk that Agrilus species pose to Quercus trees. We have detected a phylogenetic signal in the distribution of oak hosts of Agrilus and built a relatively simple yet powerful model that successfully predicts host–beetle interactions by using readily available phylogenetic and geographic data. We hope these findings will be used to focus efforts and inform actions aimed at preventing future damaging interactions, and to aid the development of targeted approaches for early pest detection in this and other host–pest systems.

Competing interests

None declared.

Author contributions

LJK and RAN conceived and oversaw the project. EH‐G compiled the datasets and performed all analyses, with guidance from RAN and LJK EH‐G wrote the manuscript, with input from LJK and RAN.

Disclaimer

The New Phytologist Foundation remains neutral with regard to jurisdictional claims in maps and in any institutional affiliations.

Supporting information

Fig. S1 Binary threshold used to divide predictions into ‘negative’ and ‘positive’ interactions.

Fig. S2 Comparison of the simplest best‐fit model (model 43) against models with only geographic or phylogenetic information.

Fig. S3 Histogram of the estimate of the intercept value for the random effect of each Agrilus species in Model 65.

Fig. S4 Continuous predictions for model 43 divided by oak species.

Fig. S5 Continuous predictions for model 43 divided by Agrilus species.

Fig. S6 Comparison of fivefold cross‐validation log‐odds predictions against Leave‐One‐Out cross‐validation log‐odds predictions.

Fig. S7 Value of different parameters for multiple Agrilus species with at least one host not predicted as such.

Fig. S8 Choropleth maps showing the number of Quercus species, known oak hosts of Agrilus, and predicted hosts per country.

Fig. S9 Phylogenetic patterns of clustering of Agrilus hosts across the Quercus phylogeny, also showing the distribution of predicted hosts.

Notes S1 Formulation of the geographic distance metrics.

Notes S2 Formulation of geographic distance models.

Notes S3 Formulation of the phylogenetic distance metrics.

Table S1 Larval host information for Agrilus species that utilise Quercus species.

Table S2 LOOIC cross‐validation to assess the best‐performing geographic distance metric.

Table S3 Model statistics.

Table S4 LOOIC cross‐validation values for models 1–44 and comparisons against model with lowest LOOIC.

Table S5 LOOIC cross‐validation values for models 45–88 and comparisons against model with lowest LOOIC.

Table S6 LOOIC cross‐validation values for models 89–132 and comparisons against model with lowest LOOIC.

Table S7 LOOIC cross‐validation values for best‐fit models from all groups and comparisons against model with lowest LOOIC.

Table S8 LOOIC cross‐validation values for those best‐fit models from all groups with only contributing variables and comparisons against model with lowest LOOIC.

Table S9 LOOIC group 1 model comparisons.

Table S10 LOOIC group 2 model comparisons.

Table S11 LOOIC group 3 model comparisons.

Table S12 LOOIC comparison of best‐performing models.

Table S13 Group‐level (i.e. random) effect for each Quercus species (i.e. each level) in model 43.

Table S14 Interactions for which the Quercus species is not currently a confirmed larval host, but its predicted value in our statistical model is over the binary threshold.

Please note: Wiley is not responsible for the content or functionality of any Supporting Information supplied by the authors. Any queries (other than missing material) should be directed to the New Phytologist Central Office.

Data availability

All scripts used in this study, as well as all lightweight input and output files (i.e. text files and phylogenetic input data files), are hosted on the following GitHub repository: https://github.com/elvirahg/agrilus‐oaks‐predictions. The full model outputs (.RData) are available on Zenodo at: doi: 10.5281/zenodo.19537500, 10.5281/zenodo.19538352, 10.5281/zenodo.19538573, and 10.5281/zenodo.19538839.