Species roles
Here we present a case study comparing the roles of two pollinator species over time. Data was from four mountaintop plant-pollinator communities in the Seychelles, sampled over the flowering season in eight consecutive months between September 2012 and April 2013 (Kaiser-Bunbury et al. 2017; Supplementary Table 2). Restoration by removal of exotic plants from these communities resulted in pollinator species becoming more generalised. This pattern was driven largely by two abundant, highly generalist pollinator species, one native (Lasioglossum mahense) and one non-native (Apis mellifera) (Kaiser-Bunbury et al. 2017). These two abundant, super-generalist species could have similar strategies for partner selection and therefore play similar roles in the community. This is the result found in the original study: both species had similar levels of specialisation (quantified using the specialisation index d¢, which measures the extent to which species deviate from a random sampling of available partners (Blüthgen et al. 2006)): 0.17 ± 0.10 and 0.22 ± 0.18 for Lasioglossum mahense and Apis mellifera respectively. Alternatively, two abundant, super-generalists could minimise competition by exploiting different areas of ‘interaction niche space’ and therefore have different roles. To test these alternatives, we calculated the motif role signatures of both species at each site in each monthly network, giving a detailed view of how each species is embedded in the community over time. We used permutational multivariate analysis of variance (PERMANOVA), stratified by site, to assess if there are significant differences between the roles the two species play in the four communities. PERMANOVA is similar to ANOVA but compares multivariate differences within and between groups without assuming normality or Euclidean distances (Anderson 2001). We used Bray-Curtis distance as the dissimilarity measure, as it is suitable for a variety of ecological data, including motifs (Faith et al. 1987, Anderson and Robinson 2003, Baker et al. 2015). PERMANOVAs were run with 10000 permutations.
The PERMANOVA analysis showed that Lasioglossum mahense and Apis mellifera had significantly different roles over time (F1,62, p = 0.0496), exploiting different areas of
interaction niche space. This means that, while Kaiser-Bunbury et al. (2017) used the species-level metric d¢ to show that both species were super-generalists, a motif approach reveals that they are generalist in different ways. This result is visualised in Fig. 9. More positive values of the first NMDS axis are associated with motif positions where more specialist pollinators compete with generalist pollinators for a shared plant resource, while negative values are associated with positions where generalist pollinators visit specialist plants with little competition. More positive values of the second NMDS axis are associated with positions where pollinators visit plants which are also visited by generalist species; negative values are associated with positions where pollinators visit plants which are also visited by specialist species. Lassioglossum mahense generally occupies higher values of both NMDS axes than Apis mellifera. Therefore, while both species are generalists, Lassioglossum mahense is in greater competition with generalist pollinators than Apis mellifera which visits more specialist plants and competes with more specialist pollinators. These differences in indirect interactions are essential for understanding the ecology of these two species and are missed using the d¢ index alone. All PERMANOVA tests and NMDS analyses were conducted in the R package vegan (Oksanen et al. 2016).
Figure 9: The movement of Lasioglossum mahense and Apis mellifera through interaction niche space
over eight months in four sites (Bernica, Salazie, Tea Plantation and Trois Freres). Each vertex represents the role of a species in a monthly network. Numbers ‘1’ and ‘8’ indicate the first and last sampling month, respectively. Shaded polygons are convex hulls containing the vertices of each species.
Motif over- and under-representation
To determine whether particular motifs occur more or less than would be expected by chance, it is possible to compare empirical motif frequencies to those produced by a null model (Milo et al. 2002). We compared the motif distributions of 122 empirical pollination networks from the Web of Life repository (web-of-life.es) to 100 randomisations of a null model where the probability of a link occurring between plant i and animal j is equal to the mean of the normalised degree of species i and j (Bascompte et al 2003). In other words, the probability of a cell in the interaction matrix being occupied is equal to the mean of the occupancy of that cell’s row and column (Bascompte et al 2003). Ecologically, this means that the likelihood of two species
interacting is proportional to their level of generalisation (degree). This null model is used to determine the significance of structural properties in pollination networks as it allows structure to be determined beyond that which results from the degree distribution alone (Bascompte et al 2003). The level of over- or under-representation of motif i was expressed as a z-score (Milo et al 2002; Stouffer et al 2007):
F' =GH=E− GMMMMMMMMIJKL NOPQRS
where GMMMMMMMM and NIJKL OPQRS are the mean and standard deviation of the randomised motif counts, respectively.
Results are shown in Figure 10. Motifs 2, 7, 13, 14, 15, 16, 17 occur significantly more than random in the majority of networks, while motifs 3 and 10 occur significantly less than random in the majority of networks (though only in just above half of networks for motif 10). The over-represented motifs all involve one or two plants interacting with between two and four pollinators. This asymmetry in the number of species in each level suggests that there are a large number of indirect interactions between animal pollinators mediated through a smaller number of plants. Conversely, under-represented motifs 3 and 10 involve two or three plants interacting with one or two pollinators, respectively. This asymmetry suggests that indirect interactions between plants, mediated by one or two pollinators, are less common than would be expected from degree distribution alone. The structure of under-represented motif 10 is particularly interesting. Of all the motifs containing three plants and two pollinators, motif 10 has the lowest possible connectance. Other motifs with the same number of species in each level, but with more dense patterns of connections, are not under-represented. This implies that, when plants and pollinators form asymmetric local structures with more plants than pollinators, these are tightly connected with a high degree of cohesion. These more tightly connected motifs will also have indirect effects which are harder to predict because there are many possible pathways through which the mutualistic benefit could flow. This is in contrast to the under-represented motif 10, which has a relatively simple and predictable structure. This finding matches results by Carvalheiro et al (2014), who found that indirect interactions between plants via shared pollinators are more frequent among abundant plants that tend to be more generalised and are thus more likely to be
involved in more densely connected motifs. Overall, therefore, we conclude that pollination networks have high levels of indirect interactions between pollinators mediated by a small number of plants, and fewer indirect interactions between plants mediated by a small number of pollinators. However, when indirect interactions between plants mediated by pollinators do occur, these are in tightly connected clusters that may have complex, and possibly hard to predict, dynamics.
Figure 10: Patterns of motif over- and under-representation in pollination networks relative to a null
model (Bascompte et al 2003). (a) Each line shows the pattern of motif over- and under-representation
for a single network. z-scores were normalised following &'= F'/U∑ F0 0; where Pi is the normalised
profile of network i, z is a z-score and j is an index over motifs (Stouffer et al 2007). (b) For each motif, the figure shows the proportion of networks which had more (black) or less (grey) of that motif than
−0.5 0.0 0.5 1.0 1 3 5 7 9 11 13 15 17 Motif Nor malised z − score
a
−0.5 0.0 0.5 1 3 5 7 9 11 13 15 17 Motif Propor tion of netw or ksb
0.00 0.25 0.50 0.75 1.00 1 3 5 7 9 11 13 15 17 Motif Propor tion of netw or ksc
the null model. (c) For each motif, the figure shows the proportion of networks which had significantly (± 1.96 standard deviations) more (red) or less (blue) of the motif than the null model. Grey indicates motif frequencies were not significantly different to the null model.