Virus isolation data improve host predictions for New World rodent orthohantaviruses

Identifying reservoir host species is crucial for understanding the ecology of multi-host pathogens and predicting risks of pathogen spillover from wildlife to people. Predictive models are increasingly used for identifying ecological traits and prioritizing surveillance of likely zoonotic reservoirs, but these often employ different types of evidence for establishing host associations. Comparisons between models with different infection evidence are necessary to guide inferences about the trait profiles of likely hosts and identify which hosts and geographical regions are likely sources of spillover. Here, we use New World rodent-orthohantavirus associations to explore differences in the performance and predictions of models trained on two types of evidence for infection and onward transmission: RT-PCR and live virus isolation data, representing active infections versus host competence, respectively. Orthohantaviruses are primarily carried by muroid rodents and cause the diseases haemorrhagic fever with renal syndrome (HFRS) and hantavirus cardiopulmonary syndrome (HCPS) in humans. We show that although boosted regression tree (BRT) models trained on RT-PCR and live virus isolation data both performed well and capture generally similar trait profiles, rodent phylogeny influenced previously collected RT-PCR data, and BRTs using virus isolation data displayed a narrower list of predicted reservoirs than those using RT-PCR data. BRT models trained on RT-PCR data identified 138 undiscovered hosts and virus isolation models identified 92 undiscovered hosts, with 27 undiscovered hosts identified by both models. Distributions of predicted hosts were concentrated in several different regions for each model, with large discrepancies between evidence types. As a form of validation, virus isolation models independently predicted several orthohantavirus-rodent host associations that had been previously identified through empirical research using RT-PCR. Our model predictions provide a priority list of species and locations for future orthohantavirus sampling. More broadly, these results demonstrate the value of multiple data types for predicting zoonotic pathogen hosts. These methods can be applied across a range of systems to improve our understanding of pathogen maintenance and increase efficiency of pathogen surveillance.