Network metrics and meta-community structure
The infection data assembled here originate from different ecosystems. Therefore, we considered all communities inferred from these data as meta-communities of cichlids and species of Cichlidogyrus . To investigate the effects of adaptive radiation, we compared the meta-communities of species infecting hosts from Lake Tanganyika (LT) and the region inhabited by the Lake Victoria superflock (LV) (see Verheyen et al. 2003) with the whole cichlid-Cichlidogyrusnetwork. We also inferred meta-communities through the Louvain community detection algorithm, an approach based on optimisation of network modularity (see Blondel et al. 2008) implemented in the Rpackage igraph v1.2.6 (Csardi & Nepusz 2006). The algorithm was applied to the entire (natural and invasive) documented host ranges with hosts and parasites treated equally as nodes. To characterise meta-community structure, we inferred several network metrics that are widely applied to weighted links, including the weighted nestedness based on overlap and decreasing fill (NODFw) (Almeida-Neto & Ulrich 2011), weighted connectance (Cw) (Bersier et al. 2002), specialisation asymmetry (SA) (Blüthgenet al. 2007), interaction evenness (Ei) (Bersieret al. 2002), and the standardised interaction diversity (H2’) (Blüthgen et al. 2006) using Rpackage bipartite v2.15 (Dormann et al. 2008, 2009; Dormann 2011).
We calculated network metrics for meta-communities including more than 10 species (Fig. 2) both for the natural ranges and the full realised host repertoire and geographical distribution (including the result of anthropogenic translocations). To correct for varying sampling intensity, we produced two null distributions (NM): Patefield’s algorithm (Patefield 1981), which randomly redistributes rows and columns of the interaction matrix (NM1) and the redistribution algorithm proposed by Vázquez et al. (2007) (NM2), which maintains the network connectance, i.e. only realised interactions are redistributed. We generated 1000 null matrices through the function nullmodel in bipartite and assessed significance as proportion of null estimates greater than the observed estimates.