Using niche modeling to detect unobserved interactions in host-parasite networks
Host-parasite interactions are quantified based on field observations, and are assumed to be a exhaustive record of host-parasite associations. However, it is likely that some interactions are missed as a result of limited sampling effort or low parasite prevalence. Prediction of missing links is a central problem in the study of social, economic, and biological networks. Prediction of host-parasite associations could address how non-native hosts integrate into host-parasite networks, or identify the potential for pathogen spillover to humans or domestic animals. Here, we test the predictive accuracy of a conditional density estimator in predicting host-parasite associations. We test the algorithm using artificial data, and then apply it to data on parasites of small mammals sampled as part of the Sevilleta Long-Term Ecological Research effort.
The conditional density niche modeling algorithm performed well on simulated networks that varied in terms of network features and trait resolution. Specifically, the inclusion of binary or uninformative variables did not strongly reduce accuracy of model predictions. Predictive accuracy increased with matrix size, such that communities with fewer than 11 host or parasite species tended to have slightly reduced accuracy (AUC ~ 0.75). Our model performed well on empirical data as well, predicting host-parasite interactions with fairly high accuracy (AUC = 0.82). This application of niche modeling does not rely on knowledge of network structure (e.g., node degree), such that predictive accuracy is achieved using only host and parasite trait data.