Fig 1.
Conceptual depiction of the proposed automated protocol.
1-2. Inputs consist of non-curated species lists. In our case study, we compiled lists from ≈160 RAMSAR wetland communities worldwide. 3. These raw species lists were enriched with functional traits (e.g., size, diet) automatically mined from public databases. 4-5. Missing data were estimated using two complementary strategies: (i) previously known allometric body mass/size relationships, and (ii) taxonomic inference, where missing values were inferred from the closest phylogenetic relatives. 6. The resulting enriched lists containing trait values were introduced as input parameters in the allometric niche model, a theoretical method capable of generating realistic food webs. 7. In the resulting (automatically constructed) networks, nodes correspond to the species found in the original ecosystem, and links represent highly plausible trophic interactions. 8. This procedure can create large ensemble of homogeneous and realistic trophic networks amenable for quantitative and comparative analysis. Credits: Panel 1 map: Panel 1 world map showing Ramsar sites sourced from the Ramsar Sites Information Service (RSIS, rsis.ramsar.org), licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). Animal silhouettes in other panels were obtained from OpenClipart (openclipart.org), which releases all content under CC0 (public domain).
Fig 2.
(A). Topological consistency (TC, X-axis) denotes the fraction of replicates of a given ecosystem in which a specific trophic interaction exists.
A large TC indicates highly stable interactions. The Y-axis represents the relative frequency of links (within a given replicate) that exhibit certain TC. As the Figure shows, although most links vary across replicates (facultative interactions), a core, stable sub-network of highly consistent links (specialized necessary interactions) remains stable. Each line represents one replicate of a given ecosystem. (B). The automatically constructed trophic networks show a scale-free degree distribution (DEG). X-axis: number of trophic interactions per node. Y-axis: relative frequency of species having that degree (within a replicate). Line colors as in the (Log scale) insets. Red: all interactions. Purple: only-input interactions (GENnerality). Orange: only-output interactions (VULnerability). Grey and dashed lines: individual (1000) replicates and average distributions under the pure ANM. Insets use logarithmic axes (log-scaled, not log-transformed data) to highlight the scale-free-like structure of the distribution.
Fig 3.
Representative two-dimensional projections of the multivariate network space defined by the quantitative descriptors summarized in Box 1.
Each point corresponds to one automatically generated network (one ecosystem × one replicate). When available, data (represented as empty black circles) corresponding to empirically measured properties of real food webs (mainly from refs. [29,32]) are also shown for comparison. (A) In-degree/out-degree asymmetry vs. absolute deviation in vulnerability relative to the pure allometric niche model (denoted by the M superscript). (B) Network connectivity vs. absolute deviation in total degree relative to the pure niche model. (C) Species richness vs. maximum number of trophic interactions involving a single species. (D) Species richness vs. absolute deviation in generality relative to the pure niche model. (E) Species richness vs. network connectivity. (F) Species richness vs. average number of trophic links per species. (G) Species richness vs. network nestedness. (H) Species richness vs. number of distinct trophic chains in the network. In panels A and B, colors encode species richness (N) for the corresponding network. In panels C-H, colors are used to identify the ecosystem of origin (from 1 to 160) to visually group replicates belonging to the same site.
Fig 4.
Example of a biologically plausible trophic network automatically generated by our algorithm.
(A). Case example: the RAMSAR wetland Orihuela del Tremedal (Spain) whose original species list (N = 110) included un-resolved names. (B). An example (1st replicate) network reveals that simulated trophic interactions are generally consistent with the known lifestyles of the species present in the ecosystem. For illustrative purposes, the silhouettes of representative species of the main trophic levels have been juxtaposed to the network. Node size reflects relative body mass; while colors indicate functional guilds (blue = plants, green = herbivores, orange = carnivores). Notice that some highly specific relationships such as insect-eater plants (circle in the central, lower part of the Figure), are adequately captured by the automated algorithm. This is achieved by introducing a few algorithmic rules devoted to handle these highly specific interactions (see Materials and Methods and SI). Credits: Panel A (left to right): world map from Natural Earth (naturalearthdata.com, public domain); map of Spain from the Instituto Geográfico Nacional (IGN), licensed under Creative Commons Attribution 4.0 International (CC BY 4.0); original photograph taken by M. Brun-Usan. Panel B: animal silhouettes obtained from OpenClipart (openclipart.org) and released under the CC0 (public domain) dedication.
Fig 5.
Quantitative comparison between empirically measured trophic interactions and those predicted by the automated protocol.
The Figure compares, on a link-by-link basis, 100 automatically generated network replicates with the corresponding empirical food webs reported for five wetland-like ecosystems with varying species richness and connectivity [6]: Benguela Coastal Ecosystem (N = 29), Ythan Estuary (N = 92), Tuesday Lake (N = 56), Skipwith Pond (N = 37), and Broadstone Stream (N = 37). Underestimated links (red) correspond to interactions present in the empirical network but missing in the reconstructed networks, whereas overestimated links (yellow) are interactions present in the reconstructed networks but absent from the empirical data. (A) Comparison using a default estimated average connectivity of ≈0.12. (B) Comparison using the empirically measured connectivity for each ecosystem. (C) Comparison with equally connected random networks, in which each possible link is realized with a probability equal to the empirical connectivity.
Fig 6.
(A). The number of semi-independent modules remains relatively low (around 4) and constant for most ecosystems. (B). The average clustering (CLUST) increases in large ecosystems. (C). The average modularity (MOD) decreases in large ecosystems. (D). Some of the modularity-related measures are linearly correlated as indicated by the corresponding Pearson’s correlation coefficient. Anticorrelation between CLUST and MOD suggests that modules become less independent as ecosystems grow in size. (E-F). Two case examples of modular structure in our dataset: (E). Naikaikemi (Japan, N = 28), and (F). Whooping Crane (USA, N = 36). Colors represent the optimal division maximizing modularity, dashed line separates carnivores-scavengers from the remaining organisms and dashed nodes represent “bridge” species between modules. (G). An idealized network showing how the modules are, in general, organized across networks.
Fig 7.
Robustness to loss of taxonomical resolution.
(A). Relationship between species richness and taxonomic richness across ecosystems, shown for multiple hierarchical levels (number of genera, families, orders, classes, and phyla). Each point corresponds to an ecosystem. This panel shows how the number of different genera in the ecosystems (each representing functionally similar organisms) escalates in proportion to the number of species N, which helps explain why our method is relatively insensitive to moderate taxonomic rarefactions (see next panels). (B) Taxonomic rarefactions were introduced by progressively collapsing species into higher taxonomic ranks (species → genus → family → order), producing genus-, family- and order-levels nodes. The panel illustrates how species belonging to the same genus (same color) in a schematic species-level network collapse into a single node in the corresponding genus-level network. For each collapsed node, body size was set to the mean of member species, and diet/habitat were aggregated as the union of their categories. After each taxonomic rarefaction, the trophic networks were recalculated, and their topological features compared with the species-level network. (C). Most topological descriptors remain, at least up to the family rank, within certain confidence interval (defined as the mean () ± the variance (σ) of that descriptor in the species-level network). (D). Many of the links in the networks constructed using medium taxonomic ranks (orange) remain similar to the species-level networks (yellow lines and Fig 2A). (E). Many above-species networks retain the scale-free architecture of the original, species-level networks (yellow lines and Fig 2B).
Fig 8.
Robustness to functional rarefactions.
Y-axes in panels A–C represent the absolute difference between each measured network attribute and the mean value of the corresponding attribute in the original (unperturbed) networks, expressed in units of the standard deviation of that attribute across replicates of the original networks. Accordingly, values close to zero indicate negligible differences, values around one indicate deviations comparable to the intrinsic variability of the original networks, and values greater than one indicate differences exceeding that variability. (A). The ecosystems’ structure is more deeply affected when herbivores are removed, while removing plants or carnivores has a milder effect. However, the removal of any trophic group has catastrophic effects on the network structure if >N/2 of species are removed (see Figs 7, S2 and panel (C)). (B). The ecosystems’ structure is deeply affected when medium-sized organisms are removed. As the four size intervals defined do not cover all (>12) orders of magnitude present in the whole dataset, the % of species removed (X-axis) never reaches the N/2 limit as in A. (D). A summary of the general trends shown in previous panels taking the average slope (A.S.) as a proxy of network change. Most topological attributes vary more intensely, and above a random rarefaction (C), when herbivores and/or medium-sized animals are removed. In (B-C), grey points correspond to individual (perturbed) replicates. Red dashed lines: linear regressions of the represented data (with slope A.S.).
Fig 9.
(A). Classical networks cannot capture the complex interactions of amphibious/biphasic life cycles. (B). A multilayer network with larval-adult and aquatic-terrestrial compartments allows representing, for instance, adult frogs feeding on adult dragonflies and dragonflies’ nymphs feeding on frogs’ tadpoles. (C). Organisms movements across layers depend on key functional traits that can be mined from public databases. These traits can inform rules to automatically build multilayer networks via a matrix operator (D). An automatically constructed multilayer network (Ntsikeni Reserve, South Africa). Most trophic edges (purple) occur within sub-habitat compartments, yet some adult organisms can move and feed across sub-habitats (green bi-directional arrows). Many ontogenic edges (orange) connect the aquatic larvae with terrestrial adults (e.g., dragonflies). Big arrows and dashed circles have been manually added to show the main functional groups and their connections. N = 170, half of the plants removed for the sake of clarity. Credits: Panels A-B: animal silhouettes obtained from OpenClipart (openclipart.org), released under the CC0 (public domain) dedication. Panel D (left to right): world map from Natural Earth (naturalearthdata.com, public domain); map of South Africa from Natural Earth (naturalearthdata.com, public domain); photograph from RAMSAR (rsis.ramsar.org), licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).