Congruent trophic pathways underpin global coral reef food webs

Significance

Species loss can weaken the trophic interactions that underpin ecosystem functioning. Coral reefs are the world’s most diverse marine ecosystem, harboring interaction networks of extraordinary complexity. We show that, despite this complexity, global coral reef food webs are governed by a suite of highly consistent energetic pathways, regardless of regional differences in biodiversity. All networks are characterized by species with narrow dietary preferences, arranged into distinct groups of predator–prey interactions. These characteristics suggest that coral reef food webs are robust to the loss of prey resources but vulnerable to local extinctions of consumer species.

Abstract

Ecological interactions uphold ecosystem structure and functioning. However, as species richness increases, the number of possible interactions rises exponentially. More than 6,000 species of coral reef fishes exist across the world’s tropical oceans, resulting in an almost innumerable array of possible trophic interactions. Distilling general patterns in these interactions across different bioregions stands to improve our understanding of the processes that govern coral reef functioning. Here, we show that across bioregions, tropical coral reef food webs exhibit a remarkable congruence in their trophic interactions. Specifically, by compiling and investigating the structure of six coral reef food webs across distinct bioregions, we show that when accounting for consumer size and resource availability, these food webs share more trophic interactions than expected by chance. In addition, coral reef food webs are dominated by dietary specialists, which makes trophic pathways vulnerable to biodiversity loss. Prey partitioning among these specialists is geographically consistent, and this pattern intensifies when weak interactions are disregarded. Our results suggest that energy flows through coral reef communities along broadly comparable trophic pathways. Yet, these critical pathways are maintained by species with narrow, specialized diets, which threatens the existence of coral reef functioning in the face of biodiversity loss.

Our planet’s ecosystems are underpinned by a remarkable diversity of species and their complex ecological interactions, which govern energy and nutrient fluxes across multiple trophic levels (1). As anthropogenic pressures reduce local biodiversity worldwide (2–4) either through severe decreases in abundance or complete local extinctions, the effects of these losses can propagate through ecological networks and induce unexpected ecosystem-scale changes (5–8). Understanding, predicting, and possibly averting such cascading processes requires detailed knowledge of how pairwise species interactions (e.g., predator–prey, host–parasite, and plant–animal interactions) scale up to the complex structure typical of ecological networks. Although the field of network ecology has burgeoned in the last two decades (e.g., refs. 9 and 10), the effort to identify species interactions has been unevenly distributed across ecosystems. Since the number of ecological links increases exponentially with species richness, hyperdiverse systems such as tropical rainforests or coral reefs pose particularly strong challenges to researchers seeking to quantify ecological interactions across entire ecosystems (11).

Tropical ecosystems are highly diverse, disproportionally threatened, and critically important from an ecological and societal perspective (12). Among them, coral reefs produce and recycle a remarkable amount of biomass, supporting the livelihoods of over 500 million people worldwide (13). The invaluable productivity of reefs is mediated by a tremendous number of interactions across multiple trophic levels (14–16), and most of this complexity, especially pertaining to cryptic organisms (17), remains largely unresolved (18). In contrast, large, mobile fishes are the most intensively studied reef taxa due to their conspicuousness, functional importance, and commercial value (19). Several large datasets exist that quantify fish-centric, trophic interactions based on the visual inspection of fish gut contents (20–24). Yet large-scale evaluations of reef trophic structure is typically performed using categorical, qualitative trophic guilds as proxies (25–28), while quantitative analyses of detailed trophic networks have been restricted to modeling simplified food webs (29–31) or to analyses within a single region (32, 33). Since global analyses on reef fish communities highlight marked variation in the representation of different body size classes (28) and trophic guilds (34) across bioregions, historical and biogeographic factors may have also favored the emergence of multiple trophic organizations across global coral reefs. However, the lack of holistic, large-scale analyses of coral reef food webs based on empirical data prevents us from identifying general characteristics of coral reef energy fluxes at a global scale.

The structure of interaction networks is heavily influenced by species’ ecological specialization and the number and nature of trophic linkages that are shared among species (i.e., network nodes) (9, 35–37). Widespread trophic generalism implies high connectivity among species and the potential for a perturbation to cascade through an entire network (38). Yet such generalism is likely to increase functional redundancy, which buffers ecosystem functioning against biodiversity loss (39). In contrast, extreme trophic specialization suggests a lack of redundancy but may favor a modular food web structure that limits the propagation of perturbations (40). On coral reefs, various studies conducted at a local scale have highlighted extensive trophic overlap among reef fish species and the dominance of dietary generalists (41, 42). However, recent biochemical and molecular analyses have identified fine-scale dietary partitioning among ecologically similar or closely related species (43–48). Disentangling these nuances of coral reef food webs and their consequences for the response of reef ecosystems to ever-increasing perturbations requires detailed, quantitative evidence from a variety of coral reef systems (49).

Here, we collated community-wide gut content analyses of coral reef fishes to build and compare coral reef consumer–resource networks across six biogeographic regions. Specifically, using network theory, we examined 1) the consistency of food web structure (i.e., modularity and node overlap) across bioregions, 2) global patterns of trophic specialization in reef fish species, and 3) the vulnerability of trophic pathways on coral reefs worldwide.

Results and Discussion

We synthesized gut content data from six distinct regions across coral reefs worldwide (Madagascar, the Marshall Islands, New Caledonia, Okinawa, Hawaii, and the West Indies) to generate a dataset containing a total of 688 reef fish species across 84 families (SI Appendix, Tables S1–S3) and 649 prey items belonging to 31 different functional or taxonomic categories (SI Appendix, Table S4). Using these data, we built six consumer–resource networks that summarize trophic linkages between fish families and their prey. We found that, despite large biogeographic differences in species richness and composition, all six food webs exhibited strikingly similar structure (Fig. 1). The observed proportion of trophic interactions that were consistent across all regions (12%) was twice larger than expected from a null model controlling for consumer diet breadth and size, resource availability, and resource exploitation (i.e., the number of consumers using a given resource; Z = 8.3 and P < 0.001). The proportion of shared interactions further increased when the number of networks was reduced. Nearly 50% of trophic interactions were shared between at least three regions, compared to 30% of shared interaction under by the null model (Fig. 2A; Z = 13.4 and P < 0.001). When we examined trophic interaction strength (i.e., prey item frequency or volumetric percentage), we found that interactions shared among all the regions were also the strongest interactions. These results differ significantly from our null model even after controlling for consumer size, consumer diet breadth, resource availability, and resource exploitation (Fig. 2B). Overall, our results suggest that the major energetic pathways across resource–consumer networks exhibit strong congruence across reefs worldwide, but they are supplemented by weak, opportunistic, and regionally idiosyncratic trophic links. The comparability of network structure across marine regions suggests that the convergent evolutionary processes that led to similar morphologies in reef fish communities across biogeographic regions (50, 51) may have also promoted the emergence of congruent trophic pathways on coral reefs.

Fig. 1.

Quantitative bipartite networks of coral reef food webs across geographic regions. (A) The location of the six coral reef food webs. (B) Food webs in i) Madagascar, ii) Okinawa, iii) New Caledonia, iv) the Marshall Islands, v) Hawaii, and vi) the West Indies. Food webs are plotted as bipartite networks with link widths indicating the interaction strength (i.e., prey item frequency or volumetric percentage). For each food web, the top level corresponds to the 84 fish families, and the lower level corresponds to the 31 prey item categories. Networks are plotted with the entire set of nodes found across the six regions. Fish families and prey items are sorted alphabetically. See SI Appendix, Tables S3 and S4, respectively, for fish predator families and prey item categories in the food webs.

Fig. 2.

Proportion and strength of shared trophic interactions among the six coral reef food webs. (A) The proportion of shared interactions across the examined food webs. Observed values are drawn with a solid line, while means and errors of the null distributions from 1,000 random food webs are drawn with a dashed line. The error bars indicate the 5 and 95% quantiles. (B) The strength of shared interactions (mean ± SE). Interaction strengths were obtained from the volumetric percentage for the West Indies and prey item frequency for Hawaii, Madagascar, the Marshall Islands, New Caledonia, and Okinawa.

We measured the breadth of fish species’ trophic niches, computed as the proportion of resources to which they were connected, to explore whether coral reef food webs were dominated by species with narrow or wide dietary preferences. On average, fish species were linked to 19% ± 0.05 of prey resources. However, this value significantly decreased when only stronger interactions were considered. Indeed, when we removed interactions with strengths <0.25 and <0.5 (Materials and Methods), the average proportion of resources used by each species decreased to 5% ± 0.01 and 2% ± 0.008, respectively (SI Appendix, Table S5). Fishes may feed on a large variety of resources, but most consumed prey types represent accessory prey obtained through opportunistic predation, while only few resources serve as staples for a given species.

Furthermore, we explored whether fish species generally use complementary or shared resources. We used the Node Overlap and Segregation (NOS) metric (52), which compares the tendency of consumer nodes to share food items against a probabilistic null expectation that accounts for the body size of consumers. Our analysis revealed dietary segregation (NOS < 0) for a substantial number of species, ranging from 24% of species in Okinawa to 39% of species in the West Indies (Fig. 3). However, when we removed the weakest interactions associated with opportunistic predation, niche segregation became dominant for the majority of species in all locations. NOS also revealed a modular network structure (Q_NOS = 0.59 to 0.80, where 0 represents no modularity and 1 represents strong modularity). We explored this pattern by computing the widely employed Newman’s modularity metric (Q) (53). Remarkably, all six consumer–resource networks were composed of five modules and exhibited higher modularity than predicted under our null model (Fig. 4A). The West Indies network was the most modular (Q = 0.36), followed by Hawaii, New Caledonia, and the Marshall Islands (Q = 0.30, Q = 0.29, and Q = 0.31, respectively). The Okinawan and Malagasy networks were the least modular (Q = 0.25 and Q = 0.26, respectively).

Fig. 3.

Modularity and reef fish species’ network role. (A) The modularity metric calculated for observed networks and one thousand random networks (mean ± 5% and 95% quantiles). (B) The distribution of reef fish species according to their network role. Points represent species over the entirety of species described within each food.

Fig. 4.

Distribution of NOS values in prey items between predators. Each panel represents a food web. A threshold was applied on interaction strength, resulting in the retention of all interactions (cyan), interactions with a strength superior to 0.25 (light gray), or interactions with a strength superior to 0.5 (dark gray).

To examine the role of individual species in the consumer–resource networks, we measured their contribution to within- versus between-module links. The majority of fish species (565 or 67.2% of species) were identified as specialists that interact with only a few resources [i.e., peripheral nodes, which contribute little to intermodule links (54)] (Fig. 4B). This included most species within the families Serranidae, Acanthuridae, Scorpaenidae, Synodontidae, Carangidae, Chaetodontidae, Apogonidae, and Muraenidae, which are some of the most dominant reef fish families (55). Connectors (i.e., generalist species with many links to other modules) were the second most common network role (32.2%; 271 species). These generalists predominantly belonged to the Lethrinidae, Mullidae, Tetraodontidae, and Balistidae. Two species, Crenimugil crenilabis and Lutjanus gibbus, were detected as module hubs in the Marshall Islands. As for network hubs, only three species were detected, including Lethrinus nebulosus in New Caledonia, Ocyurus chrysurus in the Virgin Islands, and Epinephelus merra in Okinawa. The distribution of species’ roles was almost identical across the six food webs (SI Appendix, Fig. S1).

Trophic specialization may promote the persistence of high biodiversity by facilitating species coexistence. However, extensive specialization may also result in highly vulnerable energetic pathways, since the loss of single consumer species may result in the net loss of unique trophic interactions (56). To explore this hypothesis, we performed simulations of biodiversity loss while quantifying net energy fluxes through the network. Specifically, we simulated random species removal according to three approaches: consumer removal, resource removal, and both consumer and resource removal. We designed a simple model, assuming that the sum of the standardized interaction strengths could be used as a proxy for trophic flow (i.e., amount of energy transferred to the consumer). After every species removal, we updated the trophic flow measure, allowing for a reallocation of trophic fluxes among extant species based on the initial interaction weight, as well as accounting for the new competition patterns. These simulations revealed that energy fluxes through coral reef consumer–resource network are more vulnerable to the loss of consumers (i.e., top-down effect) than to the loss of resources (i.e., bottom-up effects). This pattern is consistent across the networks from different bioregions. Furthermore, the vulnerability of net energy flux to consumer loss exceeded expectations based on a highly conservative null model where the redistribution of resources was possible only between consumers with similar body size (Fig. 5).

Fig. 5.

Energetic robustness of coral reef food webs to consumer loss. Panels show the average trophic flow (i.e., the sum of all interaction weights) after random extinction simulations of fishes (Top; blue), resources (Bottom; pink), or both fishes and resources (both; yellow) across the six networks. Removal scenarios were compared to 1,000 null networks to assess the effect of trophic specialization on the resilience of energy fluxes on coral reefs. The solid lines represent null models, and the dashed lines represent observed values.

Overall, our results reveal that coral reef consumer–resource networks are modular and predominantly composed of species with unique and sparse trophic linkages. High levels of modularity have been described in ecological interaction networks [e.g., pollination networks (57)] and can have diverse causes (58). On coral reefs, modularity may be linked to the existence of a wide variety of subhabitats for food exploitation [e.g., pelagic versus benthic (43, 55)], phylogenetic effects, or trait matching (49). However, modularity on coral reefs may also stem from the emergence of novel niches across evolutionary time that offer competitive advantages to specialists (56, 59), which ultimately favors the existence of the exceptional fish biodiversity on coral reefs (36, 60).

Despite high segregation and modularity, the presence of generalist species across networks (32.2%) suggests that reef fish network structure is more complex than a series of distinct modules. Although less common than previously hypothesized (41, 61), the trophic niches of these generalists are governed by omnivory and opportunistic feeding, which may have emerged as a competitive strategy to reduce interspecific competition under resource scarcity or as an adaptation to colonize new environments (62). However, node overlap was mostly due to weak interactions in our networks. While several predator–prey, plant–pollinator, and plant–frugivore interaction networks are based on high dietary resolution, our global study required the relatively coarse classification of prey, which inevitably impacts our results. While the more-widespread use of molecular techniques may help overcome these issues (47, 48), even visually identified prey items grouped into relatively coarse categories resulted in strongly compartmentalized trophic structuring on reefs and consistently narrow species-specific prey preferences.

The degree of specialization and modularity in consumer–resource networks intuitively affects their responses to the removal of species from the network, either through severe decreases in abundance (i.e., functional extinction) or complete local extinction. Losses of generalist species (even if they were not dominant) may propagate through the entire network, while the impact of removing specialists may be contained within individual modules (57). However, the prevalence of niche segregation makes ecosystem functioning vulnerable since the loss of each species represents the removal of a unique trophic pathway (63, 64). Indeed, our results reveal that the total interaction strength of networks, here interpreted as a proxy for energy fluxes, may be severely reduced by the loss of consumer species (37); in fact, across local networks, the majority of fish species are distinct in their core feeding habits. This suggests that previous estimations of functional redundancy, even in hyperdiverse groups, may underestimate the degree of ecological diversification present on reefs (65). Consequently, biodiversity loss on coral reefs may substantially alter energy fluxes, compromise ecosystem functioning, and significantly interrupt the ecosystem services that coral reefs provide to humanity (66).

Overall, despite strong biogeographical differences in species richness and composition, coral reef food webs share a backbone of major trophic pathways on a global scale, which are likely driven by convergent ecosystem-scale evolutionary processes. Furthermore, trophic specialization is prevalent throughout coral reef food webs, a pattern that may explain the origin and maintenance of biodiversity on coral reefs. While high modularity strengthens the trophic backbone of coral reefs, which clearly enhances overarching food web stability, species-specific specialization suggests that energetic pathways on coral reefs are vulnerable to species loss. Consequently, perturbations that substantially reshuffle reef fish communities may severely impair energy fluxes on coral reefs, ultimately resulting in significant changes to ecosystem functioning in hyperdiverse tropical coral reef ecosystems.

Materials and Methods

Food Webs Datasets.

We extracted the diets of Chondrichthyes (i.e., cartilaginous fishes) and Osteichthyes (i.e., bony fishes) from five published studies [Madagascar (20), Marshall Islands (21), Hawaii (22), West Indies (23), Okinawa (24)], and a partially published dataset from New Caledonia (67, 68). All dietary information was based on visual gut content analysis, where prey preference was quantified as either item frequency (20–22, 24, 67, 68) or volumetric percentage (23). The frequency of each prey i is equal to the number of fish guts that contain that prey i divided by the total number of nonempty guts of a species j. Only species with at least three nonempty guts were kept for subsequent analysis. We recorded 6,760 trophic interactions among 688 reef fish species and 84 families. One limitation of our data are that trophic interactions are quantified as bipartite networks, while in reality, fish species may be both consumers or resources. This was unavoidable due to the nature of the source datasets. Almost 95% of the 6,760 recorded interactions involve fishes feeding on invertebrates or algae, while almost all of the remaining links (involving fish as both resources and consumers) were not taxonomically resolved. Thus, our use of a bipartite resource–consumer network [similar to host–parasite systems (69) or terrestrial plant–herbivore networks; (70)] was justified and permitted us to use statistical tools developed for network analysis. Further limitations are discussed in the SI Appendix, SI Materials and Methods.

Prey Item Groupings.

Prey identification was heterogeneous across the six datasets, differing in taxonomic level and the use of common or scientific names (e.g., crabs versus Brachyura). All poorly informative (e.g., unidentified fragments, unknown species) and redundant items (e.g., crustacea fragments when co-occurring with an item already identified to lower taxonomic level such as shrimp) were discarded. In order to compare networks among the six datasets, prey items were assigned to a priori prey categories based on taxonomy, ecologically informative guilds (e.g., zooplankton), and the original item description (e.g., crab zoeae larvae). This categorization resulted in 31 distinct prey categories (SI Appendix, Table S4). Most groups followed official taxonomic classifications except for detritus, inorganic, worms, and zooplankton. In the West Indies dataset (23), items labeled as Algae and Detritus were assigned to both the categories detritus and benthic autotroph, and the volumetric percentage was divided in two. The category of worms includes several taxonomic groups and was created because of the numerous items described simply as worms in the New Caledonian dataset. The category zooplankton includes all eggs and larvae regardless of taxonomy, small crustaceans (e.g., copepods, amphipods), gastropods (e.g., pteropods), and other marine taxa (e.g., chaetognaths, ctenophores). At our taxonomic resolution, all resources were present in all locations, and their availability in each location can be assumed to be nonlimiting for any consumer.

Network Construction.

The consumer–resource networks were constructed as bipartite networks, where each set of nodes corresponds to a fish and their resource. Interaction strengths (i.e., volumetric percentage for the West Indies and prey item frequency for the other five food webs) were averaged by resource and consumer (i.e., fish families or species). Although food webs are usually unimodal, rather than bipartite, the examined food webs are the most exhaustive consumer–resource interactions currently available across the coral reef literature.

A network was defined by an incidence matrix N with r rows corresponding to resources and c columns corresponding to fish predators (i.e., consumers):

$$N\hspace{0.17em}=\left[\begin{array}{cc}{n}_{ij}& n_{ic}\\ n_{rj}& n_{rc}\end{array}\right],$$ [1]

where n_ij is the strength of trophic interaction (a proxy for trophic flow) between the i^th prey category and the j^th fish predator. Prey categories or fish predators absent from a region resulted in a_ij = 0. Each matrix was standardized, so the sum of each column equals 1. To ensure that each matrix had the same size, all predators (i.e., fish families) and prey occurring across all six food webs were included in all the matrices to estimate the proportions of shared interactions. For the subsequent analyses, food webs were composed of only predators and prey within a single food web.

Proportions of Shared Interactions and Trophic Specialization.

To obtain the number of shared interactions between regions, the quantitative matrices were then transformed into qualitative presence/absence matrices, and each was summed so that the final qualitative matrix had cell values ranging between zero and six. We then calculated the proportions of shared interactions on quantitative matrices among fish families and regions using the summed qualitative matrix to select interactions shared between two and six food webs. Fish families were chosen for comparison according to the low number of shared genera (n = 5) and species (n = 0) among the regions. We calculated the proportion of interactions, P, shared between two, three, four, five, and six food webs as:

$$P=\frac{{{\displaystyle \sum }}_{k=1}^{n=6}{{\displaystyle \sum }}_{i=1}^{p}ni}{{{\displaystyle \sum }}_{k=1}^{n=6}{{\displaystyle \sum }}_{j=1}^{q=84}sj},$$ [2]

where n_i is the interaction strength of all p interactions, and s_j is the sum of a column q, the number of predators within each matrix, for each matrix N. The significance of the results was assessed using a null model approach; we compared the empirical values to values obtained from randomized matrices. Specifically, we replicated the analysis of shared interactions with a set of 1,000 randomized versions of each of the six trophic interaction matrices. Implementing a conservative procedure, these were obtained by randomizing consumers’ diets (at the species level), then aggregating consumers’ diets at the family level. Randomization of consumers’ diets (at the species level) was performed using the following procedure:

• 1)two consumers, c₁ and c₂ were randomly selected;
• 2)if the ratio of the body sizes of c₁ and c₂ fell outside the range of 0.8 to 1.2, then the two consumers were discarded, and another pair was randomly selected until the ratio fell within the range;
• 3)the set of resources consumed by c₁ but not c₂ (s₁) and the set of elements consumed by c₂ but not c₁ (s₂) were identified;
• 4)if sets s₁ and s₂ were empty, one random item (i.e., resource) was selected from each set (r₁ and r₂);
• 5)resource r₁ was removed from c₁’s diet and added to c₂’s diet, while resource r₂ was removed from c₂’s diet and added to c₁’s diet;
• 6)and the original interaction strengths between c₁←r₁ and c₂←r₂ were reassigned, respectively, to the new interactions c₁←r₂ and c₂←r₁.

This procedure ensured that the number of consumers that use a given resource, as well as the number of resources used by a given consumer, did not change between the original and the randomized datasets. By controlling for these network features, we ensured that the randomized networks included the main structural components as their real-world counterparts. In addition, the total interaction weight from a given resource to all of its consumers was identical in the null and real dataset (based on the assumption that interaction weight is ecologically related to resource availability). Body size is an essential parameter governing size-dependent relationships such as predatory interactions, particularly among fish communities that are size structured (71, 72). Thus, by limiting prey reshuffling between fishes of similar sizes (i.e., body size ratio between 0.8 and 1.2), the generated random networks are less stochastic and more ecologically realistic. Anyway, reducing the strength of the size constraint to the null model had little effect on the final results (SI Appendix, Figs. S2 and S3).

Trophic niche breadth was measured for each species within each food web as the number of prey items consumed over the total number of prey items present in the food web. To detect a pattern between the specialization of consumers and the strength of their trophic interactions, this metric was calculated on the empirical food webs after applying a threshold on interaction strengths (i.e., >0.25 and >0.5).

NOS.

Nestedness is the tendency for nodes to share neighboring nodes, and it may play an important role in network stability (e.g., ref. 73). Here, we investigated nestedness using the NOS metric (52). The NOS metric quantifies the degree of overlap for each pair of nodes, then provides a network-wide measure of NOS by averaging all pairwise measures. NOS values ranges from −1 (perfect segregation) to 1 (perfect overlap). To quantify overlap and segregation, NOS compares the observed overlap with the average overlap that one would obtain by randomly reallocating resources among consumers. NOS uses a probabilistic approach based on combinatorics, providing an efficient alternative to performing an equivalent null model analysis based on network randomization. Furthermore, when assessing overlap/segregation in a trophic network, NOS accounts for “forbidden” links (i.e., links not observed in the actual network due to constraints such as functional trait incompatibility). This allows us to disentangle how specific ecological processes are impacted by network structure (in our case, energetic fluxes across food webs) (74).

For our analysis, we restricted the set of potential interactions by limiting the diet of a fish predator to the total pool of food items consumed by fishes of a similar size (i.e., fish species with a maximum body size ranging from 0.8 and 1.2 of the fish predator). The SD of NOS values also provides a derivative measure of modularity (52). A modular pattern emerges from a bimodal NOS value distribution with peaks at −1 and 1. To assess the emergence of network structure from interaction strength, we calculated the NOS values after only keeping interactions with strengths of >0.25 and >0.5.

Modularity and Species Roles in Networks.

Because of the heterogeneous nature of assessing interaction strength across regions, we assessed modularity using binary matrices. This also permitted the use of other network metrics and null models to test for network structure, which has not been well developed for weighted matrices with ordinal numbers (75, 76). Modularity measures the tendency of a network to be arranged in clusters, in which groups of nodes are more connected to each other than with nodes from other modules (53). Although this modularity algorithm was developed for unipartite networks, it is among the most commonly used algorithm in ecological bipartite networks (57, 77). We compared our modularity values to those obtained using a null model approach (Materials and Methods, Proportions of Shared Interactions and Trophic Specialization), controlling for consumer diet breadth and size, resource availability, and resource exploitation. From the computation of modules, c and z values were calculated for each fish species (54, 57). These values correspond to the among-module connectivity and the within-module degree, respectively. Thresholds were set to infer a species’ role from the distribution of c and z values (54, 57). A species is peripheral when it has few links but most of its links are within its own module (z ≤ 2.5 and c ≤ 0.62). Connector species link modules together (z ≤ 2.5 and c > 0.62). Module hubs are highly connected to other species within their own modules (z > 2.5 and c ≤ 0.62). Finally, network hubs have homogeneously distributed links across all modules (z > 2.5 and c > 0.62). All analyses were performed with the statistical software R (R Core Team, 2018) using the packages bipartite (78) and nos (79).

Quantifying Network Robustness to Species Loss.

To examine how weighted bipartite networks respond to the loss of species, we performed three distinct extinction simulations to remove consumers, resources, or both consumers and resources. Our approach assumes that for each matrix corresponding to a consumer–resource network, the total number of interaction strengths can be used as a proxy for the number of resources that flow through the system. We sought to determine whether the degree of niche segregation among fish species reduced the overall capacity of the system to acquire resources. Indeed, in case of perfect niche segregation, if a consumer species is lost, no other species is able to access its resources; thus, the resource is theoretically rendered unused. On the other hand, in the case of resource loss, the greatest impact would emerge in a modular food web, where many consumers specialize on individual prey items. In this case, the loss of one resource determines the loss of an entire food web module and the associated energetic fluxes. In our simulations, the values for trophic interaction strengths were obtained from the raw data before standardization. A standardized counterpart of the N matrix, M, was generated as the following:

$$M={\left\{N\left[N^T/{\displaystyle \sum }_{i=1}^r{M}_{ij}\right]\right\}}^T.$$ [3]

Here, r is the number of rows in the matrix. Hence, the amount of the i^th resource available to each associated consumer is affected by competition, modeled by the ratio $N^T/{\displaystyle \sum }_{i=1}^r{M}_{ij}$ and modulated by the original interaction weights (N), a proxy for ecological and/or functional constraints to determine the accessibility of the i^th resource to the j^th consumer. We tracked all consumer–resource interaction weights (i.e., sum of all M entries) after each consumer was removed to calculate changes in energy fluxes. At each step, depending on the scenario, one species was removed from N. Then, we generated a new M matrix and calculated the corresponding energetic flow. By recalculating the M matrix at each step, we accounted for species loss in food webs by redistributing interaction weights. This permitted consumer diets to adjust to resource availability, mirroring a real, adaptive ecological system after perturbation. We simulated three different scenarios where we either progressively removed (in random order) only resources, only consumers, or either resources or consumers with equal probability. The procedure was reiterated until all target species were removed. We performed 1,000 simulations for each species loss scenario and biogeographical region.

We compared all simulated scenarios of trophic flow to a null model to show the effect of trophic specialization on the system’s resilience to consumer removal. For each network, we generated 10,000 random networks, each with the same size, shape, and incidence matrix marginal totals as the corresponding real-world networks (76) (SI Appendix, SI Materials and Methods). This approach allowed us observe marginal total constraints, which ensure that the number of resources used by any consumer, as well as the number of consumers using a given resource, are identical between the original matrix and the randomized one. Furthermore, this approach permitted the reshuffling of resources between consumers with comparable body sizes (body size ratio between 0.8 and 1.2).

Data Availability

All data and scripts have been deposited in a publicly accessible GitHub repository (https://github.com/ChloePZS/foodweb) and are also available on Zenodo (https://zenodo.org/record/5341340#.YS3ikt8682w). All other study data are included in the article and/or supporting information.