In ecological communities, interactions between consumers and resources lead to the emergence of ecological networks and a fundamental problem to solve is to understand which factors shape network structure. Empirical and theoretical studies on ecological networks suggest predator body size is a key factor structuring patterns of interaction. Because larger predators consume a wider resource range, including the prey consumed by smaller predators, we hypothesized that variation in body size favors the rise of nestedness. In contrast, if resource consumption requires specific adaptations, predators are expected to consume distinct sets of resources, thus favoring modularity. We investigate these predictions by characterising the trophic network of a species-rich Amazonian snake community (62 species). Our results revealed an intricate network pattern resulting from larger species feeding on higher diversity of prey, promoting nestedness, and specific lifestyles feeding on distinct resources, promoting modularity. Species removal simulations indicated that the nested structure is favored mainly by the presence of five species of the family Boidae, which because of their body size and generalist lifestyles connect modules in the network. Our study highlights the particular ways traits affect the structure of interactions among consumers and resources at the community level.

The network structure of interactions between snakes and their food resources

We analysed the snake diet derived from a long-term study carried out in a Central Amazonia site on the natural history of forest snakes (Martins and Oliveira 1998). We described the resource use by snakes as an interaction matrix A in which if a snake i feeds on a given resource j and zero otherwise. The matrix A defines a bipartite network in which one set of nodes is represented by snake species, the other set of nodes by resource types and the links describe interactions between snake species and food resource types. Our food resources are not described at species level but at coarse categories such as small mammals, medium mammals and big mammals (see details in Supplementary material - dataset). Similar approach led to insights in the study of food webs (Cohen 1977) and individual-based networks (Araújo et al. 2008). In fact, there is no intrinsically correct level of description when characterizing an ecological network (Guimarães 2020). We opted for these coarse categories because they are in agreement with (i) the evidence that snakes are specialized in broad categories of resources, e.g., serpentiform organisms that include snakes, amphisbaenians, and caecilians (see Martins and Oliveira 1998); (ii) the level of detail available from the diet analyses of snakes. Having said that, to verify if our level of network description affects our analyses we performed a set of sensitivity analyses (details below).

We used four metrics to characterise the structure of the interactions network analysed: (i) degree distribution, which is the description on how the number of food resources a given snake can feed on (the degree) varies across snake species; (ii) connectance (C), which is the proportion of all possible interactions actually recorded in the network. Connectance values range from 0 (non-connected network) to 1 (maximum connectance); (iii) modularity (M), a measure of the extent to which the network is formed by groups (modules) of snake species in which snake within a module overlap in much of their resources, whereas snakes in different modules show no or weak resource use overlap; and (iv) nestedness (N), which consists of an interaction pattern in which the specialists interact with sets of resources with which the generalists also interact. Detailed descriptions of the metrics are available in the Supplementary material Appendix 1.

We used Q_B metric, defined by Barber (2007), to characterise modularity, with values ranging from 0 (non-modular network) to 1 (completely modular). A simulated annealing algorithm (Guimerà & Amaral 2005) was used to optimise the Q_B value. Modularity analyses were performed using the Modular program (Marquitti et al. 2014). All the above and the following analyses were performed using R version 3.5.1 (R Core Team 2018), with the exception of modularity. We performed a set of sensitivity analyses to verify if our results are dependent on our approach to compute modularity (Supplementary material Appendix 1).

The NODF metric was used to characterise the nestedness degree (Almeida-Neto et al. 2008) and its values ranges from 0 (non-nested network) to 100 (perfect nestedness). The degree of nestedness and modularity were then compared with a theoretical benchmark provided by the null model 2 of Bascompte et al. (2003) (see detailed description in Supplementary material Appendix 1). We generated 1000 null model matrices to estimate nestedness and modularity. If a network shows a degree of nestedness or modularity larger than expected by the null model 2, then there is evidence of ecological or evolutionary processes acting on the network organisation that goes beyond those shaping the degree of specialisation of the snake species (e.g, Bascompte et al. 2003).

In order to highlight the unique inferences provided by the network approach, we compared the results of the network analysis with the results of a multivariate analysis. Multivariate analysis methods are widely used in ecology due to their ability to analyse complex systems registered in an interaction matrix (Prado et al. 2002). Among the several types of multivariate analyses, we chose correspondence analysis (CA) because of its ability to reveal reciprocal relationships between two sets of equal interest (Greenacre 1984; Lewinsohn and Prado 2008), in our case, snakes and their food resources.

The role of snake species in network structure

If the network of interactions analysed follows the organisation pattern structured by body mass (i.e. presenting higher nestedness ​​than expected by the null model 2), we hypothesized snake average body mass to be positively correlated with the number of resources consumed by the snake species. To explore this prediction, we investigate the association between average body mass and the role of species in the network structure. We recorded the estimates of the average body mass of each snake species in our network (data available in Feldman et al. 2016). Average body mass was log-transformed prior to analysis (Supplementary material Appendix 2 Table A1).

In order to understand the individual contribution of each species of snake to nestedness, we used a jackknife resampling approach in which we removed a snake species and recomputing the degree of nestedness. We repeated the procedure for all snake species in the network and then we computed a change in nestedness: ΔN_i = N - N_i, in which N is the degree of nestedness of the complete network and N_i is the degree of nestedness after the removal of a snake species i. If body size is shaping the contribution to nestedness, we should expect that the ΔN_i will assume increasingly positive values as larger snakes are removed from the network, indicating that nestedness is higher in the presence of these larger snake species.

The relationship between lifestyles and network structure

Because dietary specialisation in snakes can be related to habitat occupation (see Martins et al. 2002, Alencar et al. 2017), we expect snake lifestyles to affect the degree of dietary specialisation (e.g., an aquatic snake would rely upon aquatic prey). If this is true, the distribution of lifestyles in the different modules will not be random. We evaluated this prediction using two analyses. First, we analysed the frequency of snake lifestyles in different modules. We estimated the probability of the observed number of species of a given lifestyle in a given module be reproduced by randomly assigning species across modules, but preserving the number of snake species in each lifestyle and the number of snake species in each module (n = 1000 randomisations). Then, we analysed the dissimilarity on lifestyles between modules. To do so, we used the Bray-Curtis index, available in the vegan package in R (Oksanen et al. 2018) (see detailed description in Supplementary material Appendix 1). Dissimilarity between a pair of modules range from 0 (modules are identical in the composition of lifestyles) to 1 (no lifestyle occurs in both modules).

Sensitivity analyses focuses on the level of resource resolution

Sampling effects may affect the description of network patterns. Therefore, we performed a sensitivity analysis to explore how robust is the description of network patterns to changes in our dataset. We add information to the use of resources by snakes by using data from other Amazonian regions, based on evidence that there is no significant intraspecific variation on the snake's diet across different localities in Amazonia (Martins and Oliveira, 1998; Supplementary material Appendix 1 and Appendix 2 Table A1).

Snake diet often include food resources that are mainly consumed and resources that are only eventually consumed. We performed a sensitivity analysis to check if the patterns reported in our study are robust enough when considering the presence or absence of secondary resources in the snake diet. We described two matrices of interactions: (1) a matrix in which only main resources were considered; (2) and a matrix in which both main and secondary resources were considered. We defined if a resource is main or secondary according to information about snake diet preferences available in Martins and Oliveira (1998). Then, we calculate the nestedness and modularity values in the presence and absence of secondary resources. The nestedness values of the two networks were compared with a null model generated with 5,000 random removals of food resources from each of the analysed networks. Finally, we calculated whether there was a significant difference between the nestedness of the network in the presence and absence of secondary resources.

Because taxonomic resolution might influence the detection of patterns in the network (Rezende et al. 2009), we performed another sensitivity analysis to check if the type of resource categorization could affect the network patterns. Thus, we described two other matrices of interaction with different degrees in the resources taxonomic resolution: less specific network (Supplementary material Appendix 1 Figure A2 and Appendix 3 Table A1) and more specific network (Supplementary material Appendix 1 Figure A3 and Appendix 4 Table A1).

 Additional information that may be needed is available in the supplementary material captions.