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

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.