Network link prediction
Overall, the NLP algorithm performed solidly (plug-and-play : mean AUROC = 0.72, full model AUROC = 0.85; Poisson N-mixture : mean AUROC = 0.62) despite missing data (Fig. 5a). Several input parameters improved performance of the plug-and-play algorithm significantly (Fig. 5c) including the basin/basin type, host phylogenies, trophic level, and life style whereas the parameters for parasite morphology and phylogeny decreased the performance. A substantial amount of species interactions remains undetected, albeit less for LT and LV (Fig. 5b; Appendix S6). For LV , model performances were slightly better (0.78, 0.87; 0.76) but with the host phylogeny as the most important predictor. For LT , the models showed little discriminatory power (0.41, 0.87; 0.62).
Discussion
We investigated the patterns of host-parasite interaction of African cichlid fishes and their gill parasites belonging toCichlidogyrus , a proposed model system for macroevolutionary research (Pariselle et al. 2003; Vanhove et al. 2016). This study is the first to empirically investigate the effects of adaptive radiation events on species interactions [but see Maynardet al. (2018) for simulations]. The size of this species network (10529 infections, 477 interactions) is comparable to widely used host-parasite datasets in terms of species richness, e.g. the Global Mammal Parasite Database (GMPD) (Nunn & Altizer 2005), the Sevillata interaction network (Dallas & Presley 2014), and other fish-parasite (Lima Jr et al. 2012; Bellay et al. 2015) and plant-arthropod systems (López-Carretero et al. 2014; de Araújo et al. 2020; Oliveira et al. 2020; de Araújo & Maia 2021). The system is also the first to encompass closely related parasite species infecting a host system that is a model for speciation research (Seehausen 2006). Therefore, our dataset could be an asset for comparative studies in network ecology and ecological parasitology.