Impact of area, isolation, and distance from source on bacterial interaction networks
SPIEC-EASI networks of bacterial interactions changed in betweenness centrality and modularity associated with environment (Figures 6 and 7). In REGUA sampling sites, low-abundance bacteria and bacteria in the genera Arsenophonus , Wolbachia , and Bartonella had higher betweenness centrality than other bacteria in these networks, acting to connect graph neighborhoods (Figure 6). As networks became more modular (i.e. decreasing habitat fragment area), there were fewer bridge nodes with greater extremes of high and low betweenness centrality. In smaller habitat patches, low-abundance bacteria andMycoplasma had much larger betweenness centrality compared to other bacteria within each network. Arsenophonus ,Wolbachia , and Bartonella did not have high betweenness centrality in the networks of smaller habitat fragments. The exception to this pattern is fragment F4, where betweenness centrality of bridge nodes was not extremely large compared to the betweenness centrality of the other bacteria in the graph and Arsenophonus ,Wolbachia , and Bartonella had high betweenness centrality. While fragment F4 had small habitat area, it was the only fragment outside of REGUA that supported all streblid bat fly species.
The raw modularity of each network was higher in fragments outside REGUA than in sites within the protected area, although this pattern was not statistically significant (Figure 7A; Wilcoxon Rank Sum test, p-value= 0.133). This is largely due to fragment F9, which showed much lower modularity than any other network. When we remove fragment F9, the difference in modularity between REGUA sites and fragments is significant (Wilcoxon Rank Sum test, p-value=0.033). Null distribution-centered modularity measures, which are standardized to account for variation in network size and shape, are highest in REGUA, the largest fragment (F10), and the fragment with the second highest bat fly diversity after the largest fragment (F4; Figure 7B). Smaller fragments and those distant from a source had low centered modularity, and fragment F9 was no longer an outlier. Standardizing modularity by comparison with the mean and standard deviation of the set of measured networks did not control for network size and shape variation, and mimicked the pattern exhibited by raw modularity. Calculation of the Z-score modularity using the null distribution for each site indicated that all measured networks have significantly higher modularity than a randomly connected network of the same size and shape.
Sample size (i.e., number of parasite individuals) was lowest in small fragments and highest in large fragments, with the exception of F4 which has intermediate area, isolation, and distance to source measurements. While there was no impact of sample size on ASV richness in each network, a greater number of samples may allow detection of more edges between nodes (more interactions between ASVs). Efforts to reconstruct SPIEC-EASI networks using subsampling of the sites with the largest sample sizes failed because SPIEC-EASI networks on subsamples never reached stability. To examine the topology of the network independent of size (number of connections), we decomposed the graph into 4-node graphlets and examined correlations between the incidences of graphlet orbits and ordinated the networks using PCoA (Mahana et al., 2016; Ruiz et al., 2017; Yaveroğlu et al., 2014). The position of networks in the ordination differed from what we would expect if landscape variables were the primary driver of network topology (Figure 7D). The REGUA2 sampling site was found to be quite different from the other REGUA sampling sites which were instead closest to fragment F9, the second largest fragment. Outside of the REGUA sites, there was no pattern of differentiation with decreasing area, increasing isolation, and increasing distance to source.
Discussion