Get 20M+ Full-Text Papers For Less Than $1.50/day. Subscribe now for You or Your Team.

Learn More →

Multilayer networks reveal the spatial structure of seed-dispersal interactions across the Great Rift landscapes

Multilayer networks reveal the spatial structure of seed-dispersal interactions across the Great... ARTICLE DOI: 10.1038/s41467-017-02658-y OPEN Multilayer networks reveal the spatial structure of seed-dispersal interactions across the Great Rift landscapes 1 1 1 1 1 Sérgio Timóteo , Marta Correia , Susana Rodríguez-Echeverría , Helena Freitas & Ruben Heleno Species interaction networks are traditionally explored as discrete entities with well-defined spatial borders, an oversimplification likely impairing their applicability. Using a multilayer network approach, explicitly accounting for inter-habitat connectivity, we investigate the spatial structure of seed–dispersal networks across the Gorongosa National Park, Mozam- bique. We show that the overall seed–dispersal network is composed by spatially explicit communities of dispersers spanning across habitats, functionally linking the landscape mosaic. Inter-habitat connectivity determines spatial structure, which cannot be accurately described with standard monolayer approaches either splitting or merging habitats. Multi- layer modularity cannot be predicted by null models randomizing either interactions within each habitat or those linking habitats; however, as habitat connectivity increases, random processes become more important for overall structure. The importance of dispersers for the overall network structure is captured by multilayer versatility but not by standard metrics. Highly versatile species disperse many plant species across multiple habitats, being critical to landscape functional cohesion. CFE – Centre for Functional Ecology, Department of Life Sciences, University of Coimbra, Calçada Martim de Freitas, 3000-456 Coimbra, Portugal. Correspondence and requests for materials should be addressed to S.T. (email: stimoteo@gmail.com) NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications 1 | | | 1234567890():,; ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y ver the recent decades, ecological networks have proved a Contrarily to the high spatial turnover in species and interac- valuable framework to simultaneously evaluate the role of tions, the functional role of species is considered relatively Ospecies, their interactions, and the importance of the stable . Centrality measures have been largely used to assess the 27,28 emerging community structure for the persistence and stability of topological position of a species in the structure of networks . biological communities . Such studies revealed that ignoring the In a multilayer context, such overall centrality can be estimated complex web of interactions between plants and animals in which with Google’s PageRank algorithm, which has been successfully 29 30 many vital ecosystem functions are rooted might jeopardize the used to guide conservation strategies . 2,3 long-term functioning and persistence of ecosystems . To date, Here, we investigate how a mutualistic multilayer network is most studies have considered networks as entities with discrete structured across habitats and the importance of species to the borders defined by the experimental design, ignoring the potential cohesion of seed dispersal across a complex landscape. To this across-border connections , or alternatively as aggregations of end, we collected seed–dispersal interactions across the Gor- several spatially and temporal sampling occasions into an overall ongosa National Park, Mozambique, to build the most complete, 5 31 network . In nature, however, these sub-networks are linked by seed–dispersal network of the African continent to date , common species and by processes that span over several spatial including all potential guilds of seed dispersers. Gorongosa and temporal scales, contributing to the functional connectivity of underwent a severe defaunation that affected many of the large 6 32,33 ecosystems . The importance of the spatial dimension of net- herbivores, and its recovery is now en route . In this context, 7–9 works of interactions is becoming increasingly recognized , seed–dispersal is particularly vital for plants to recolonize newly highlighting the key function of species that cross habitat available patches or disturbed ground , and is likely a key driver 10 35 boundaries acting as mobile links that connect the different of long-term habitat dynamics in Gorongosa patchy landscapes . habitats. Recent work has provided further evidence of the Our objectives are twofold. First, we aim to explore the spatial importance of the often-neglected inter-habitat links and their distribution of seed–dispersal modules (i.e., communities of unequivocal ecological relevance . Perhaps ironically, the appli- tightly interacting plants and their dispersers) spanning across the cation of such tools that proved particularly suited to tackle the different habitats of the Gorongosa National Park. We will do so intrinsic complexity of ecosystems is limited by the amount of by evaluating the modularity of multilayer networks formed by complexity that can be sampled and analyzed, leading to a frag- discrete, yet interconnected layers representing different habitats. mentation of real networks, likely to result in oversimplifications, We used different null models to explore how the strength of the and eventually to incomplete or erroneous conclusions about interlayer connectivity affects the overall structure of the spatial 12,13 network structure, dynamics, and stability . Similarly, ignor- multilayer network, and to what extent this multilayer approach ing the role of different species as spatial couplers of ecosystems improves the currently used monolayer analyses of disconnected may hinder our understanding of natural processes, e.g., the flux and aggregated networks. Second, we aim to assess the relative of energy, or nutrients, between aquatic and terrestrial systems, contribution of each disperser species to the cohesion of seed pollen transfer by insects across the landscape, or the dispersal of dispersal across habitats. We will do so by exploring dispersers seeds of invasive species by birds . Recently, some authors multilayer versatility, which expresses their contribution to the have started to tackle this issue by treating different habitats, or mobile link function both within and between habitats. We dis- patches of habitat as a set of layers within a larger multilayer cuss the potential of this new metric by comparing it to the 15–17 network . An expansion of the concept of beta-diversity has information provided by traditional species-level descriptors. been proposed to measure dissimilarities between networks, by exploring species and interactions turnover between groups of Results 16,18 independent (i.e., formally disconnected) networks .Ina Overview of seed dispersal in Gorongosa. During this one year, further step, Frost et al. quantified the connectivity between we collected 1399 fecal samples (1174 mammal dung piles and spatial layers (habitats) of a host-parasitoid network, though they 236 bird droppings) produced by 98 animal species, of which 508 did not explore the effect of habitat connectivity to the structure (29%) had at least one undamaged seed. Overall, 12,159 unda- of the spatial network. However, only now ecologists have started maged seeds from 94 plant species were retrieved from the feces to explicitly include interlayer edges in the analysis of the actual of 29 dispersers, comprising 508 links. Focal observations pro- structure of “ecological multilayer networks” , taking advantage duced 85 further links (14% from the total), whereas camera traps of recent theoretical developments and analytic tools from other contributed with 15 new links (2.5%). In total, we compiled 608 13,19 research areas . links between 32 animal species and 101 plant species, in four 1,20,21 A key structural pattern in networks is modularity , habitats (Fig. 1). measuring the extent to which species form cohesive groups Overall, primates were responsible for most interactions, (modules) where species interact more often within the same namely Papio ursinus (chacma baboon, 35%) and Cercopithecus module than with species in other modules . These modules pygerythrus (vervet monkey, 10%), followed by Loxodonta provide insights into the phylogenetic history and trait con- africana (elephant, 22%) and Civettictis civetta (African civet, vergence of unrelated species, resulting from local co-adaption, 7%) (Fig. 1). The three most commonly dispersed plant species and ecological convergence in the use of resources . By mea- represented 41% of all recorded interactions, namely Ziziphus suring multilayer modularity, the connectivity between layers, i.e., mucronata (Rhamnaceae, 15%), Sclerocarya birrea (Anacardia- interlayer edge strength, is explicitly accounted for, with the ceae, 13%), and Hyphaene natalensis (Arecaceae, 13%). advantages of detecting modules that span across layers. It also We estimated that our sampling effort captured 77% of the allows the identification of nodes that can belong to different disperser species and 44% of the plants with similar levels of modules in different layers, thus particularly relevant for main- sampling completeness across the four habitats (Supplementary taining the continuity of ecosystem functions in space or Fig. 1 and Supplementary Table 1). 12,24 time . However, the ideal way to quantify interlayer edge strength is still a matter of research in multilayer network research, and the investigation of the relative importance of intra- Modular structure of the spatial multilayer network. To eval- and interlayer processes is essential to understand the structure of uate the extent to which the seed–dispersal interactions are sorted 12,24,25 36 multilayer networks . into distinct communities of tightly interacting species ,we 24,37 calculated the multilayer modularity of the spatial network of 2 NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y ARTICLE Birds Carnivores Elephant Insects Primates Rodents Antelopes 15 2 3 4 6 7 Dispersers Plants a b c d e f Overall network ++ + Grassland Transition forest Mixed dry forest Miombo Fig. 1 Quantitative seed–dispersal network of the Gorongosa National Park, Mozambique. Both the aggregated (top) and the individual habitat (bottom) networks are based on the same sampling effort and are represented on the same scale. The boxes in the top level represent disperser species and those on the bottom level represent the plant species dispersed. The gray lines linking the two levels represent pairwise species interactions, and their width proportional to the interaction frequency. The aggregated network was obtained by pooling all interactions across the four habitats, and summing their frequencies. Main seed dispersers: 1. Pycnonotus tricolor,2. Civettictis civetta,3. Loxodonta africana,4. Cercopithecus pygerythrus,5. Papio ursinus,6. Hystrix africaeaustralis, and 7. Redunca arundinum. Most commonly dispersed plants: a Centaurea praecox,b Grewia inaequilatera,c Hyphaene natalensis,d Sclerocarya birrea,e Tamarindus indica, and f Ziziphus mucronata. The full list of species can be seen in Fig. 3 and Supplementary Fig. 2, for animals and plants, respectively. The silhouettes used in this figure are all sourced from Open Clipart and were made available under a CC0 1.0 licence Gorongosa (see “Methods” section and Supplementary Methods predicted by both null models tended to converge to that of the for details on the multilayer modularity algorithm). Using a observed network at very high values of interlayer strength multilayer formalism , this network is defined by the animal (Fig. 2a and Supplementary Data 1), indicating an increasing seed–dispersal interactions (intralayer links) in each habitat importance of random processes in structuring the networks. (layer), with habitat connectivity (interlayer links) provided by This suggests that when habitat connectivity is very high the the common species. Ultimately, interlayer links should be overall network structure becomes less determined by the identity interpreted as the movement of matter or energy between layers, of animals connecting them, and might be more contingent on in our case the effective movement of animals and seeds between the structure of seed dispersal within habitats. habitats, and quantified in a way that estimates the intensity of To understand the added value of the multilayer approach in these movements (interlayer strength). Multilayer modularity was relation to the traditional monolayer approach, we compared the calculated across a range of interlayer strength (0–10), assuming results from the multilayer analysis with those provided by the that any co-occurring species between habitats effectively con- currently standard approaches of either merging all data into a nected them with the same intensity, to test how the structure of single aggregated network (Q ), in which interactions aggregated the spatial network is affected by habitat connectivity. The mul- occurring at multiple habitats are summed across habitats, or tilayer modularity of the Gorongosa seed–dispersal network was considering each habitat as a discrete and disconnected network. very high across the whole range of values of habitat connectivity The structure of the aggregated network is influenced by the (Q = 0.903–0.993; Fig. 2a), with an overall increasing distribution of the interactions among the species, with multilayer trend toward an asymptote just below 1. We used two null models modularity being significantly lower than predicted by the (see “Methods” section for details) to test how the structure of the intralayer null model (mean Q = 0.43 vs. Q = aggregated null models spatial network is influenced by the seed–dispersal process within 0.59, p < 0.001; Supplementary Fig. 3). However, it ignored each habitat (intralayer null model) and by the identity of the habitat connectivity because it cannot incorporate such informa- animals connecting these habitats (interlayer null model). The tion. In the disconnected network, habitats are considered totally structure of the empirical network, across the range of habitat independent from each other, thus equivalent to calculate connectivity values, was statistically different than that predicted modularity for each of them . Modularity was similar or slightly by both null models, though in opposite directions: reshuffling higher than that of the aggregated network and much lower than interactions within each habitat (intralayer null model) over- that of the multilayer network, ranging from 0.43, in the Mixed estimated modularity, whereas reshuffling the identity of the forest, to 0.56, in the Grassland (Fig. 2 and Supplementary Fig. 3). habitat-connecting animals in each habitat (interlayer null model) The number of modules detected in the multilayer structure is underestimated modularity (Fig. 2a). The identity of the dis- mostly constant, oscillating between 11 and 12 across most values persers and the intensity of movements between habitats (inter- of interlayer strength, except for very small values, where some layer strength) play a more important role for the spatial structure additional modules were detected (Figs. 2b, 3, and 2). The of the seed–dispersal network than the pattern of seed dispersal intralayer null model consistently predicted more modules than within each individual habitat. Nonetheless, the modularity observed, while the interlayer null model consistently predicted NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications 3 | | | ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y away bird, is consistently assigned to the same module of the bush 1.5 *** babies, Fig. 3). 1.4 The strength of each interaction can vary across habitats, reflecting different animal resource preferences in different 1.3 contexts, and therefore, species can change their module 24,38 affiliation between habitats . We calculated species adjust- 1.2 ability as the proportion of animal or plant species that switch Intra-layer null model module affiliation at least once between any pair of habitats . 1.1 Observed Most species do not change module affiliation across habitats, 1.0 exhibiting a relatively low or non-existing adjustability (Supple- n.s. mentary Fig. 4). When the intensity of species movement between 0.9 Inter-layer null model habitats (interlayer strength) is low, animals and plants tend to ** interact with distinct set of species in each habitat and a higher 0.8 proportion of species will change their module affiliation between habitats. As the intensity of these movements intensifies, and *** habitat connectivity increases, species adjustability becomes negligible and interactions tend to occur among the same species across all habitats. However, this stabilization on interaction 20 partners happens at different levels of habitat connectivity for animals and plants (Fig. 3; Supplementary Fig. 4; and Supple- Intra-layer null model mentary Data 1). For both animal and plant species, adjustability was generally more affected by the identity of the animal (interlayer null model) than by the pattern of interaction within Inter-layer null model Observed habitats (intralayer null model). For low interlayer strength, animal adjustability was significantly lower than predicted by the 10 intralayer null model, but higher than predicted by the intralayer null model (Supplementary Fig. 4). However, both null models *** performed better at greater values of interlayer strength. The 02468 10 interaction pattern within habitats (intralayer null model) had a Inter−layer strength variable effect on plant adjustability: the observed plant adjust- ability was significantly higher for very low, but also for high Fig. 2 Modularity and number of modules observed in empirical networks habitat connectivity, but lower observed adjustability between and predicted by two different null models. Mean maximized modularity (a) these values. The interlayer null model consistently predicted and mean number of modules (b) of the observed networks (black) across significantly higher plant adjustability for lower interlayer the range of interlayer strength (0–10), and comparison against the two null strengths (Supplementary Fig. 4). Thus, animals are more likely models: intralayer null model (blue), and interlayer null model (red to disperse the same plant species across habitats than plant symbols). Values presented as the mean (±SEM) of 100 runs of the species are to rely on the same dispersers, and animal movement modularity function, for each interlayer strength. The significance of the across the landscape exerts a stronger influence in the spatial observed modularity and number of modules was compared against those structure of the seed–dispersal network. of the null networks with a one-sample t test. *p < 0.050, **p < 0.010, ***p < 0.001, n.s. non-significant. Full results are presented in Supplementary Data 1 Contribution of disperser species to seed dispersal cohesion. We did not detect differences on animal species richness across fewer modules than observed, across the whole range of interlayer the four main habitats of Gorongosa (G test: G = 1.84, p = 0.61; strength (Fig. 2b and Supplementary Data 1). In the aggregated Fig. 4a and Supplementary Table 2). Mixed forest holds a greater network, the average number of modules detected was 11.4, richness of plants than the other three habitats, but this was only which was in line with those detected in the multilayer network significant in comparison to Grassland and Miombo (Fig. 4b and (Supplementary Fig. 3 and Fig. 2), and significantly higher than Supplementary Table 2). As for richness of interactions, Mixed those predicted by the intralayer null model (mean modules forest had more interactions than Transition forest, and both observed = 11.4 vs. null model = 9.2; t(99) = −28.60, p < 0.001). habitats had more interactions than Grassland and Miombo (all The mean number of modules in each habitat of the disconnected pairwise G tests: p < 0.002; Fig. 4c and Supplementary Table 2). network was variable and ranged from 6 to 13 (Supplementary Dispersers’ specialization did not differ significantly among Fig. 3 and Fig. 3). In the spatial multilayer network, modules are habitats (Χ = 2.49, df = 3, p = 0.49; Fig. 4d and Supplementary subsets of species that strongly interact across the different layers Table 3). 25,39 of the network . For animals, this corresponds to species that We calculated each disperser multilayer versatility, which is occur and disperse seeds from the same plant species in more equivalent to an overall measure of centrality to identify those than one habitat (Fig. 3 and Supplementary Fig. 2). For example, that are topologically important to the structure of the spatial most primates (baboon, vervet monkey, and Otolemur crassicau- network . For this effect, we used a unimodal projection of the datus (bush baby)) all disperse Z. mucronata and are consistently network, in which two animal species are connected if they placed in the same module in the multilayer and in the aggregated disperse the same plant species , thus providing an insight over networks, but not when habitats are weakly connected or their likely “functional redundancy” . Links between species were considered independent. It is worth to note that module quantified by weighting the number of shared interactions by the 42,43 affiliations do not necessarily group phylogenetically related assemblage size , minimizing the loss of information asso- species, but species that feed on similar resources, which in seed ciated with unimodal projections . Multilayer versatility revealed dispersal might be mostly determined by behavioral and that few dispersers are disproportionately important, namely the morphological constraints (e.g., Corythaixoides concolor, the go- baboon and the elephant, followed by a long tail of species with 4 NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications | | | Mean maximized modularity Mean number of modules NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y ARTICLE Multilayer Non-multilayer = 0.1  = 0.5  = 1.0  = 2.0  = 10.0 Tragelaphus sylvaticus Tragelaphus strepsiceros Tragelaphus angasii Tockus nasutus Tchagra senegalus Redunca arundinum Pycnonotus tricolor Psammodes sp. Potamochoerus larvatus Phacochoerus africanus Papio ursinus Panthera leo Ourebia ourebi Otolemur crassicaudatus Oriolus larvatus Numida meleagris Mongoose Loxodonta africana Lagonosticta rhodopareia Kobus ellipsiprymnus Hystrix africaeaustralis Hippotragus niger Genetta tigrina Corythaixoides concolor Connochaetes taurinus Civettictis civetta Chlorocichla flaviventris Cercopithecus pygerythrus Cephalophus natalensis Ant Andropadus importunus Aepyceros melampus Fig. 3 Module affiliation of animal species in the spatial multilayer network of Gorongosa. Module affiliation is shown for five different interlayer edge strengths (left block), and for the monolayer networks, considering either the aggregated network or each habitat individually (right block). In each case, the run with the highest maximized modularity was used (module affiliation for plant species is shown in Supplementary Fig. 2). Within each network, different colors represent different modules. Colors in different blocks are independent lower versatility (Fig. 5a). The importance of these species comes multilayer versatility and both dispersers mean specialization (d′) from being central in the structure of the seed–dispersal network and the number of habitats where they occur (r = −0.255, because they share plant partners with many other animals, but p = 0.208, Fig. 5a; r = 0.383, p = 0.053, respectively, Supplemen- also because they share plant species across different habitats. The tary Table 4). However, dispersers multistrength was only versatility of dispersers in the multilayer network was correlated moderately correlated with their importance (i.e., versatility) on with their versatility in the aggregated network (r = 0.671, the multilayer network (r = 0.514, p = 0.007; Supplementary s s p < 0.001; Fig. 5b and Supplementary Table 4), but the Table 4). Species multistrength extends the concept of its importance of species with low versatility is underestimated in monolayer counterpart, expressing the total number of links of the aggregated network (Fig. 5b). There were relatively few shared a species across all layers of the network , i.e., the total shared links among habitats (total edge overlap = 8.2%). However, all interactions with all its neighboring species across the habitats. habitat pairs, except Miombo and Grassland, shared more than However, contrary to versatility, multistrength does not account 20% of the interactions (Fig. 6). for the distribution of these links in relation to the other species, We evaluated if the information condensed by multilayer or the number of layers in which these links occurs. Thus, versatility could be captured by other species-level metrics, although both metrics are related, multistrength will not reflect namely specialization d′, number of habitats, and species the importance of a species for the overall structure of the multistrength. We did not find a significant correlation between multilayer network as much as versatility. NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications 5 | | | Grassland Miombo Mixed forest Transition Grassland Miombo Mixed forest Transition Grassland Miombo Mixed forest Transition Grassland Miombo Mixed forest Transition Grassland Miombo Mixed forest Transition Aggregated Grassland Miombo Mixed forest Transition ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y a b 25 75 aaaa ab ab a b 20 60 10 30 5 15 0 0 c d 250 1.0 aa b c aa a a 200 0.8 0.6 0.4 50 0.2 0 0.0 Fig. 4 Network- and species-level descriptors of the interactions from each habitat of Gorongosa. Differences among the main habitats of Gorongosa in terms of a animal species richness, b plant species richness, c number of interactions, d species specialization (mean ± SEM). Different letters indicate statistical pairwise differences: a–c, pairwise G tests (see Supplementary Table 2 for full results); d, generalized linear mixed models (63 observations/ occurrences of 32 animal species in four habitats; Supplementary Table 3 for full results) 15,16,18 Discussion habitats , has long been recognized as critical for the Species and communities are not randomly distributed across the dynamic of patchy habitats across complex landscapes . Never- planet, but they are strongly structured by spatial attributes tra- theless, and despite the current interest on species interactions ditionally recognized by ecologists (e.g., niches, habitats, land- networks, these are yet to explicitly accommodate this interlayer scapes, and biomes). Traditionally, species interaction networks dynamic when analyzing the structure of spatial networks. have been studied as discrete entities with borders defined by the Here, we implement for the first time a multilayer approach to researchers based on different landscape attributes. However, evaluate the spatial structure of an ecological network explicitly species interactions do not abruptly finish at habitat borders, and incorporating the interlayer strength connecting networks from therefore the decision of merging or segregating data from these adjacent habitats. Our spatial multilayer seed–dispersal network spatial units is far from trivial. Nevertheless, ecologists are still exhibited a highly modular structure, i.e., species tend to interact faced with a paucity of tools to evaluate when such combination with subsets of species (i.e., modules) within subsets of spatially of data is useful, or when it might increase the noise around the coupled habitats. By explicitly including non-zero interlayer links, patterns of interest, thus obscuring important conclusions. i.e., the habitat connectivity promoted by the common species, it Although still based on the recognition of different habitats, the is possible to account for the interdependence of the network implementation of a multilayer approach provides a valuable tool structure across multiple habitats , and identify modules that that allows for better decisions regarding the merits of segregating spread across habitat borders . or merging spatially (or temporal) explicit data. For a more realistic module detection, interlayer strength However, interlayer connectivity has never been explicitly should ideally be measured empirically to reflect the effective incorporated in the analysis of the modular structure of spatial movement of the individual species across habitat borders . ecological networks. In this study, we investigate the spatial Unfortunately, obtaining such data at the community and land- structure of a seed–dispersal network spanning across multiple scape levels, i.e., all species, across all habitats can be incredibly habitats explicitly considering interlayer connectivity. We made challenging. The alternative of assigning the same interlayer use of a highly comprehensive data set collected in a highly strength to all species, i.e., assuming that all species connect diverse African landscape including all potential disperser guilds. habitats with the same intensity, is a clearly undesirable simpli- This adds to the sparse knowledge of seed dispersal in Africa, but fication as species connect habitats with different intensities has direct implications for our understanding of seed–dispersal because of their differential ability to move across and establish in 46,47 networks across the globe. a given habitat . Incorporating such empirical data could have Our results show a highly modular structure of the spatial important implications on modules found by the modularity multilayer network that is influenced by the strength of the function; the relative importance between intra- and interlayer connectivity between habitats, with about half the communities of process would be different for each individual species, thus seed dispersers detected bridging most of the habitats (Figs. 3 and affecting its probability of changing module affiliation . 4). The network is dominated by a few highly versatile species Exploring the modular structure across a range of interlayer edge that secure both local (habitat level) and global (landscape level) strengths is an alternative to obtaining empirical data, and has dispersal of seeds, ensuring the spatial continuity of the seed– been often done in other fields to understand its importance for 24,25,38,48 dispersal process. processes spanning across different layers . Our analysis Landscapes are intrinsically dynamic, being constantly shaped revealed that the modularity and the number of modules are by local disturbance and ecological succession . Understanding mostly affected at extremely low levels of interlayer strength, how animals move between habitats providing key mobile links , suggesting that the spatial community structure can be main- and how ecological interactions are distributed across tained even if the strength of the habitat connectivity is low 6 NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications | | | Grassland Transition Mixed forest Miombo Grassland Transition Mixed forest Miombo No. of interactions Animal richness Mean specialization d ′ Plant richness NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y ARTICLE a b 1.0 1.0 rho = −0.255, p = 0.208 rho = 0.671, p < 0.001 0.8 0.8 0.6 0.4 0.6 0.2 0.4 0.0 0.2 0.2 0.4 0.6 0.8 1.0 Multilayer versatility Fig. 5 Correlates of multilayer versatility. Correlation between animal species versatility in the multilayer network (bars) and the mean specializationd’ (dots) (a), and between multilayer versatility of the monolayer versatility of the aggregated network (b). Full data available in Supplementary Table 4 Shared interactions within and between layers, which is objectively defined by the relative strength of the inter- to intralayer edges. Regarding the modules' composition, these grouped-together Miombo Transition forest species that are not always phylogenetically close (e.g., primates 0.27 were grouped with the go-away bird), suggesting that functional and morphological matching, such as gape-size and seed/fruit size, are more important drivers of seed–dispersal interactions . Interestingly, we detected low adjustability for most species and 0.12 module affiliations remained mostly constant across habitats. Module switching occurred only for some species (e.g., primates, elephants, or civets), and at very low values of habitat con- 0.37 nectivity (Fig. 3). Most animals, however, tend to disperse the same plant species in different habitats, thus maintaining a Mixed dry forest Grassland similar functional role across the landscape , even if habitat connectivity is very low. This can only be detected if the habitats 0.22 are explicitly linked in the analysis of network structure. Resource availability largely determines animal movements at the land- scape level . In turn, the inter-habitat movement of dispersers is likely to affect plant regeneration dynamics, and thus resource Fig. 6 Similarity in terms of shared interactions, between the habitat pairs availability. The capacity of species to adjust their interactions to of Gorongosa. Similarity is assessed using the edge overlap between the specific contexts (thus increasing overall adjustability) is likely habitats in the spatial multilayer network, i.e., proportion of shared links important for species persistence in changing environments, across habitats. The width of the links is proportional to the correlation while at the same time tends to promote a greater connectivity between each habitat pair. The silhouettes used in this figure are all (e.g., seed dispersal) across habitats. In Gorongosa, some of the sourced from Open Clipart and were made available under a CC0 1.0 species that changed module affiliation have generally wide-range licence movements and can distribute seeds between habitats, thus giving a key contribution to plant genetic diversity and spatial dis- 34,49 relative to the strength of the interactions within the habitats. The tribution of plant populations through seed dispersal . This is inter- to intralayer edge strength ratio is essential for the outcome particularly true for intrinsically large-scale processes, such as of modularity estimation, affecting the probability of nodes being seed dispersal, while other processes, such as belowground 38,48 assigned to distinct modules . As such, when interlayer cou- mycorrhizal associations, seem to be structured at much finer pling increases, the number of detected modules is expected to scales . 24,25 decrease . Importantly, the structure of the seed–dispersal Although the number of interactions differed among habitats, network was not fully captured by the aggregated network or by and the overall edge overlap was relatively low, neither the considering each habitat as an independent network. Conse- richness of dispersers, the mean number of dispersed plants, or quently, the result obtained by using a multilayer approach is not dispersers’ specialization varied significantly. biased by any decision regarding aggregating or disconnecting the Multilayer versatility allowed us to identify the animal species different layers of the network. Instead, the resulting structure is a that are most important for dispersing seeds simultaneously at the consequence of the relative importance of the processes occurring NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications 7 | | | x = y Versatility/Specialization d ′ Papio ursinus Loxodonta africana Cercopithecus pygerythrus 0.33 Civettictis civetta Phacochoerus africanus Aepyceros melampus Hystrix africaeaustralis Redunca arundinum Chlorocichla flaviventris Andropadus importunus Ourebia ourebi Mongoose Hippotragus niger Oriolus larvatus Kobus ellipsiprymnus Numida meleagris Connochaetes taurinus 0.28 Cephalophus natalensis Tragelaphus sylvaticus Corythaixoides concolor Otolemur crassicaudatus Genetta tigrina Potamochoerus larvatus Pycnonotus tricolor Tragelaphus strepsiceros Tragelaphus angasii Aggregated versatility ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y local/habitat scale and at the global/landscape scale by linking The effective conservation and restoration of natural areas multiple habitats. In the Gorongosa landscape, primates, ele- requires an integrated view of how species and their interactions phants, and African civets are central nodes in the network and maintain functional ecosystems on complex landscapes, and a likely important for its stability and cohesion . Although the multilayer approach is a most valuable tool to explore these versatility of the aggregated network can correctly identify the factors. Here, we took a step further in the analysis of spatial most important dispersers, it underestimates the importance of mutualistic networks, and using interlayer edges strength, we species that are restricted to one or a few habitats (e.g., Genetta explicitly considered the interactions between plants and their tigrine (genet) or Tragelaphus strepsiceros (kudu); Supplementary dispersers across multiple habitats in the analysis of the network Fig. 5). The relatively low correlation between multilayer versa- structure. Furthermore, we identified key spatial coupler species, tility and multistrength, and the non-significant relationship with which play a pivotal role in long-term vegetation dynamics in the number of habitats where each species occurs and its spe- Gorongosa by ensuring the dispersal of genes across the land- cialization (d′)reflect the information gain of using multilayer scape. These key spatial couplers, namely primates, elephants and versatility, which could not be captured by conventional metrics. African civets, should be highly regarded in the protection of this The most versatile seed dispersers of Gorongosa were those essential ecosystem service. switching module affiliation between habitats. They are known As many other types of networks, mutualistic networks are not 52–54 for incorporating a high proportion of fruits into their diets , temporally static, and they do not abruptly stop at habitat bor- having relatively long gut retention times, and traveling for long ders . Therefore, forcing the analysis of biotic interactions into 53–55 distances . This allows seeds to escape the high intraspecific spatially delimited “network snapshots” and ignoring habitat competition near their parent plants by diversifying the deposi- connectivity will inevitably limit the insights that can be gained tion site of ingested seeds, and thus increasing their chances of by this approach . Here, we show that a multilayer approach can 56 32 successful recruitment . Baboons are ubiquitous in Gorongosa be used to link ecological processes that occur in different spatial and assume a central role as seed dispersers across the whole layers of a network providing insights that may not be captured park, and elephants, whose populations are still recovering in using traditional representations of monolayer networks. Over- Gorongosa , are also essential to the dispersal of plant species, looking the multiple relationships between nodes on different particularly those with very large fruits and seeds (e.g., H. nata- layers may lead to an inaccurate network structure, but also to lensis or Borassus aethiopum) . Surprisingly, despite being locally misidentification of the real role of species in the whole-network 19,24,29 abundant and often considered important seed dispersers in structure . Alternatively, the explicit incorporation of the many ecosystems , birds had a low versatility. While bird ver- temporal and spatial dynamics into a multilayer network satility might have been underestimated as a result of the different approach represents an important next step in the study of sampling method used (mist-netting), the low proportion of bird animal-plant interactions. droppings with seeds (only 7 out of the 96 bird species captured where found to disperse seeds, and these were only found in 19 Methods out of the 236 bird droppings analyzed), and the consistently low Field site and sampling. This work was carried out in the Gorongosa National sampling completeness estimated to all sampling methods Park, Mozambique (hereafter Gorongosa), a hyperdiverse park covering 4067 (transects: 25%, mist-netting: 13%, and focal observations: 13%) 2 km at the southern end of the Great Rift Valley (18°47′43.2d″S 34°28′09.1″E). We suggests that the lower bird versatility is actually structural rather defined four major habitats based on the vegetation structure and flooding regime: (1) grassland, periodically inundated grassland, with few shrubs and virtually no than a sampling artifact. Potential biases may emerge if any trees; (2) Transition forest, characterized by short flooding periods and dominated particular animal or plant groups are under or oversampled due by trees of Faidherbia albida, H. natalensis,or Acacia xanthophloea with a mostly to the use of different sampling methods. This problem might be open understory; (3) Mixed forest, occasionally flooded and formed by a diverse countervailed by performing analysis on rarefied or unweighted mixture of tree species with a dense and closed understory; and 4) Miombo, one of the most extensive habitats in Africa, unaffected by floods, and dominated by networks, though these are subjected to their own caveats. We are Brachystegia spp. trees with a dense understory (see ref. for details). Throughout only aware of a single study that explored the potential con- a year, we reconstructed seed–dispersal interactions from the four habitats by sequences of merging data originated from different sampling retrieving intact seeds from animal dung collected in the field. Sampling took place methods in the assembly of seed–dispersal matrices ; the authors in 12 occasions, evenly spaced between June 2014 and May 2015, except from December to February when floods make the park mostly inaccessible. In order to concluded that this was in fact beneficial due to the com- sample all potential disperser guilds, we employed complementary sampling pro- plementarity of the methodologies. Moreover, it must be noted tocols. In each sampling occasion, one transect ca. 2000 m and 5 m wide was that the consistently low estimates of sampling completeness are performed in each habitat and separated from any other transects by at least 350 m. likely to be largely underestimated for at least two reasons: first, Overall, 48 transects were run (ca. 96 km), corresponding to a surveyed area of because species accumulation curves are based on the assumption 480,000 m . A team of two observers simultaneously collected animal dung sam- ples, corresponding to the deposition of a dung pile of a single animal, and that communities are closed, an assumption that is not met in identified the disperser species by direct observation of defecating animals, or using year-round studies, where new species “enter” the community of 63,64 the expertise of local park rangers and field guides , and recorded direct potential interactions as a consequence of advancing phenology observations of animals ingesting fruits. Bird dispersal was evaluated by collecting (i.e., fruiting season); and second, because a large proportion of droppings from birds captured during mist-netting sessions on each sampling occasion, run for 5 h after dawn. Birds were kept inside individual holding bags, the potential interactions will never be detected because they are and released after producing a dropping . All samples were carefully screened not really possible due to spatial, temporal, and phenological under a ×40 magnifying microscope, and all undamaged seeds identified against a 60,61 mismatch between species . These “forbidden links” (or “true reference collection of seeds/fruits collected in the field. Seeds that could not be zeros”) can amount to a large proportion of the unobserved identified visually were barcoded, and their DNA sequences compared against online databases, with species identified based on the best match and on a checklist potential links (up from 44 to 77%) in seed–dispersal networks . of Gorongosa plants (see Supplementary Information for details). Most seeds Interestingly, the significant edge overlap between habitat pairs (92%) were identified to the species level, 7% to genus level, and less than 1% to the (i.e., the proportion of shared links) confirms the a priori family level or grouped into morphotypes, hereafter referred to as “species” for assumption that seed dispersal does not stop at habitat borders. simplicity. Sampling was further complemented with the analysis of motion- triggered camera traps, operating for five nights per habitat per sampling occasion Taken together with the results from multilayer modularity, to record interactions from non-conspicuous animals or those feeding at night. We showing that species tend to maintain their module affiliation estimated sampling completeness for animal and plant species in each habitat. The throughout the landscape, these results suggest that species traits use of different methods to record interactions could result in different interaction (such as mobility and size) may largely determine their multilayer sampling completeness, thus we also estimated completeness of interaction sam- versatility. pling for each method. This was done by estimating the proportion of observed 8 NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y ARTICLE presented as the proportion of networks generated by the null models with richness in relation to the total asymptotic richness estimated by the non- 66 67 modularity lower than the observed networks. For each network, we calculated the parametric estimator Chao2 using function specpool from package vegan ,in R 68 mean number of modules and the mean adjustability of animal and plant species software . as the proportion of species in each level of the network that changed module affiliation at least once between habitats. The mean number of modules and Multilayer network construction. We assembled a multilayer quantitative adjustability of the observed networks was compared against that of the networks 12,25 bipartite seed–dispersal network for each habitat, which were visualized using the R generated by the null models with a one-sample t test . package bipartite . Interaction frequency was quantified by calculating a pooled Multilayer modularity calculations and the interlayer null model were frequency of occurrence, collating the information from the different sources (fecal performed in MATLAB . analysis, mist-netting, camera traps, and direct observations). This pooled fre- quency of occurrence, resulted from the direct sum of all fecal samples containing Contribution of disperser species to seed dispersal cohesion. First, we assessed 5,33,70 at least one seed of a given plant species , the number of transects where a whether the specialization of seed dispersers differed consistently between habitats given focal interaction was detected, and the number of camera-trap recordings (1 by calculating animal specialization (d′), which quantifies their selectiveness for night = 1 sample), where a given interaction was detected. seeds within the range of resources used . However, the number of interactions of Using a multilayer formalism , we assembled a multilayer network formed by: a species is considered to reflect both resources availability and consumer activity. (a) a set of nodes (called “physical nodes”) representing the animal dispersers and This metric takes into account the pattern of interaction of a species in relation to the plant species whose seeds are dispersed, (b) a set of layers representing the the available resources, while being robust to sampling effort, network size and different habitats; (c) a set of “state nodes” that correspond to the manifestation of 74 75 asymmetry , also see ref. . We used a GLMM, with Gamma errors, and included each node on a given layer; and (d) a set of two types of edges connecting the nodes disperser species as a random factor. The model was analyzed with the R package pairwise, namely, intralayer edges (i.e., animal seed–dispersal interactions); and lme4 . If any level of the independent variable was significant, we assessed its interlayer edges, connecting state nodes between layers (i.e., animal or plant species 2 77 overall effect with a Wald Χ test available in R package car . The overall fit of the between habitats). Interlayer edges encode the movement of animal and plant model was assessed with the Akaike’s information criterion against a reduced species between habitats. While the temporal scale of the movements is not exactly model, which only included the intercept. Pairwise comparisons between habitats the same, both animal and plant genes frequently cross and establish in were tested with Tukey HSD test with the R package multicomp . neighboring habitats , and for that reason can be considered effective habitat Secondly, we explored seed–dispersers’ importance in the network by connectors with the strength of the interlayer links quantifying the intensity of the calculating their overall versatility as seed dispersers. Versatility identifies species habitat coupling provided by this movement. Spatial multilayer networks have a that are topologically important for the structure of the multilayer network . categorical (non-ordinal) coupling, i.e., interlayer edges are not constrained in any Versatility was calculated using software muxViz 1.0 , which adapts the Google’s specific order and any pair of layers can be connected . The quantification of the PageRank algorithm to describe the position of a node within the structure of a interconnectivity of the multilayer network is used in the calculation of some network based on a random walk between adjacent nodes (see Supplementary (modularity and multistrength), but not all the network diagnostic (versatility and Information for further details), being equivalent to a global measure of centrality. edge overlap). The implementation of this method requires bipartite networks to be projected To assess differences in the richness of animals, plants, and plant–animal onto unimodal networks. While some projection methods entail some loss of interactions among habitats, we used a G test. If an overall effect was present, we information , we applied a weighted projection which estimates interaction weight performed pairwise G tests to identify differences between habitat pairs. This based on the proportion of shared interactions (i.e., seed species shared by analysis was performed with the R package RVAideMemoire . 42,43 disperser species) relatively to the total network size , thus minimizing the loss of information. The projection was performed with function projecting_tm from the R package tnet . This algorithm is particularly suitable to multilayer Modular structure of the spatial multilayer network. To evaluate the extent to 29 40 networks as it condensates information on dispersers niche overlap , based on which the seed–dispersal interactions of Gorongosa are sorted into distinct com- the importance of their shared dispersed seeds . munities, we calculated multilayer modularity (Q ). Multilayer modularity, multilayer To understand how animals shared their links across the different habitats, we as its application to monolayer counterparts, quantifies to what extent nodes are calculated edge overlap with software muxViz 1.0 , which quantifies the organized into modules of strongly interacting nodes interacting more frequently proportion of common links between animals across habitats. than expected by chance . The Q modularity function was maximized applying a We evaluated the information condensed by multilayer versatility in respect to “generalized Louvain” method . This method proceeds until the network config- the species versatility of an aggregated network, and to other species-level metrics, uration that maximizes the weight of the edges inside the modules in relation to a calculating its correlation with aggregated species versatility, specialization d′, null model is found (see Supplementary Information for details). Following Pilosof et al. , the modularity function was adapted to reflect the multistrength, and the number of habitats in which a species is present. The specialization index d′ measures the distribution of a species interactions with each bipartite nature of our networks. The “generalized Louvain” method requires the specification of two parameters: the resolution limit γ, and the interlayer coupling partner over the total number of partners available . Multistrength is an extension of its monolayer version, and sums the total weight of the links incident on a node ω. The resolution limit defines the detail to which the network will be resolved into communities, and can be viewed as the importance given to the null model across all layers, taking into account the links connecting nodes in the different 38 24,38 layers (see Supplementary Information for details). It expresses the total number compared to the empirical data , and we used the default resolution γ = 1 . The of shared interactions of a dispersers species with all its neighboring species across interlayer coupling quantifies the strength of the connection between layers of a all habitats. Versatility and multistrength were calculated using muxViz 1.0 . network, i.e., the effect that species have in connecting the different layers. However, measuring such interlayer strength is intrinsically challenging and there is no absolute method to do so. To explore the importance of interlayer Code availability. MATLAB scripts for the estimation of modularity and gen- strength for the detection of modules, we calculated modularity along a range of erating the interlayer null model are available from https://figshare.com/s/ values of interlayer strength (from 0 to 10), assuming uniform interlayer strength cb5723b5adf640ec0451, and R code for generating the intralayer null model, across all species, i.e., all species connecting any two pair of habitats have the same analysis of the modular structure, and generating the files to be used with muxViz effect in the interlayer process . The stochastic nature of this algorithm means that are available from https://figshare.com/s/197888d27ec3d4a0b9d0. a different maximum is reached on each run, thus we run the modularity function 12,25 100 times, and averaged the results obtained . We compared the results Data availability. The seed–dispersal interaction network matrices are available on obtained with a multilayer network to that of two different representations of the reasonable request. same network: (a) an aggregated network (Q ), where all interactions across aggregated the different layers were pooled to create one overall aggregated network, with the frequency of interactions that occur in multiple habitats being summed, and (b) a Received: 18 October 2016 Accepted: 18 December 2017 disconnected network where habitats are considered fully independent from each other, i.e., interlayer strength is set to zero, and thus modularity is calculated for each of them . We test the modular structure of the observed network comparing it to the 12,25 structure of networks built under the assumptions of two null models : (1) an “intralayer null model” that keeps the number of links (i.e., connectance), while References redistributing the individual interactions across the whole matrix as implemented by r00_both in vegan , was used to assess the influence of structure within each 1. Thebault, E. & Fontaine, C. Stability of ecological communities and the layer/habitat in the modular structure of the multilayer network ; and (2) an architecture of mutualistic and trophic networks. Science 329, 853–856 (2010). “interlayer model” following the same rationale of the nodal model , i.e., keeping 2. Tylianakis, J. M., Laliberté, E., Nielsen, A. & Bascompte, J. Conservation of the same matrix, but redistributing species identities independently in each layer, to species interaction networks. Biol. Conserv. 143, 2270–2279 (2010). assess if the modular structure is dependent on the identity of the species 3. Jordano, P. Chasing ecological interactions. PLoS Biol. 14, e1002559 (2016). connecting the habitats. We run each null model 100 times along the same range of interlayer strength (0–10). The significance of the observed modularity was compared against the distribution of the modularity of the null networks, and NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications 9 | | | ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y 4. Timóteo, S., Ramos, J. A., Vaughan, I. P. & Memmott, J. High resilience of seed 38. Bazzi, M. et al. Community detection in temporal multilayer networks, with an dispersal webs highlighted by the experimental removal of the dominant application to correlation networks. Multiscale Model. Simul. 14,1–41 (2016). disperser. Curr. Biol. 26, 910–915 (2016). 39. Pilosof, S., Morand, S., Krasnov, B. R. & Nunn, C. L. Potential parasite 5. Heleno, R. H., Ramos, J. A. & Memmott, J. Integration of exotic seeds into an transmission in multi-host networks based on parasite sharing. PLoS ONE 10, Azorean seed dispersal network. Biol. Invasions 15, 1143–1154 (2013). 1–19 (2015). 6. Allen, C. R. et al. Quantifying spatial resilience. J. Appl. Ecol. 53, 625–635 40. Mello, M. A. R. et al. Keystone species in seed dispersal networks are mainly (2016). determined by dietary specialization. Oikos 124, 1031–1039 (2015). 7. Bascompte, J. & Jordano, P. Mutualistic Networks 206 (Princeton University 41. Yachi, S. & Loreau, M. Biodiversity and ecosystem productivity in a fluctuating Press, Princeton, NJ, 2014). environment: the insurance hypothesis. Proc. Natl Acad. Sci. USA 96, 8. Hagen, M. et al. Biodiversity, species interactions and ecological networks in a 1463–1468 (1999). fragmented world. Adv. Ecol. Res. 46,89–120 (2012). 42. Opsahl, T. Projecting Two-mode Networks Onto Weighted One-mode Networks 9. García-Callejas, D., Molowny-Horas, R. & Araújo, M. B. Multiple interactions https://toreopsahl.com/2009/05/01/projecting-two-mode-networks-onto- networks: towards more realistic descriptions of the web of life. Oikos 125, weighted-one-mode-networks/ (2009). 336–342 (2017). 43. Newman, M. E. J. Scientific collaboration networks. II. Shortest paths, weighted 10. Lundberg, J. & Moberg, F. Mobile link organisms and ecosystem functioning: networks, and centrality. Phys. Rev. E 64, 16132 (2001). implications for ecosystem resilience and management. Ecosystems 6, 44. Opsahl, T. Structure and Evolution of Weighted Networks 104–122 (University 0087–0098 (2003). of London, Queen Mary College, London, 2009). 11. Knight, T. M., McCoy, M. W., Chase, J. M., McCoy, K. A. & Holt, R. D. Trophic 45. Turner, M. G. Landscape ecology: the effect of pattern on process. Annu. Rev. cascades across ecosystems. Nature 437, 880–883 (2005). Ecol. Syst. 20, 171–197 (1989). 12. Pilosof, S., Porter, M. A., Pascual, M. & Kéfi, S. The multilayer nature of 46. Hampe, A. Plants on the move: the role of seed dispersal and initial population ecological networks. Nat. Ecol. Evol. 1, 101 (2017). establishment for climate-driven range expansions. Acta Oecol. 37, 666–673 13. Kivela, M. et al. Multilayer networks. J. Complex Netw. 2, 203–271 (2014). (2011). 14. Olesen, J. M. et al. From broadstone to Zackenberg: space, time and hierarchies 47. Bélisle, M. Measuring landscape connectivity: the challange of behavioral in ecological networks. Adv. Ecol. Res. 42,1–69 (2010). landscape ecology. Ecology 86, 1988–1995 (2005). 15. Frost, C. M. et al. Apparent competition drives community-wide parasitism 48. Bassett, D. S. et al. Robust detection of dynamic community structure in rates and changes in host abundance across ecosystem boundaries. Nat. networks. Chaos 23, 13142 (2013). Commun. 7, 12644 (2016). 49. Levine, J. & Murrell, D. The community-level consequences of seed dispersal 16. Poisot, T., Canard, E., Mouillot, D., Mouquet, N. & Gravel, D. The dissimilarity patterns. Annu. Rev. Ecol. Evol. Syst. 34, 549–574 (2003). of species interaction networks. Ecol. Lett. 15, 1353–1361 (2012). 50. Rodríguez-Echeverría, S. et al. Arbuscular mycorrhizal fungi communities from 17. Bodin, Ö. & Norberg, J. A network approach for analyzing spatially structured tropical Africa reveal strong ecological structure. New Phytol. 213, 380–390 populations in fragmented landscape. Landsc. Ecol. 22,31–44 (2007). (2017). 18. Trojelsgaard, K., Jordano, P., Carstensen, D. W. & Olesen, J. M. Geographical 51. Jordán, F., Liu, W. & Davis, A. J. Topological keystone species: measures of variation in mutualistic networks: similarity, turnover and partner fidelity. Proc. positional importance in food webs. Oikos 112, 535–546 (2006). R. Soc. B Biol. Sci. 282, 20142925–20142925 (2015). 52. Ray, J. C. & Sunquist, M. E. Trophic relations in a community of African 19. Boccaletti, S. et al. The structure and dynamics of multilayer networks. Phys. rainforest carnivores. Oecologia 127, 395–408 (2001). Rep. 544,1–122 (2014). 53. Slater, K. & du Toit, J. T. Seed dispersal by chacma baboons and syntopic 20. Guimarães, P. R., Pires, M. M., Marquitti, F. M. & Raimundo, R. L. eLS1–9 ungulates in southern African savannas. South Afr. J. Wildl. Res. 32,75–79 (Wiley, Hoboken, NJ, 2016). (2002). 21. Fontaine, C. et al. The ecological and evolutionary implications of merging 54. Campos-Arceiz, A. & Blake, S. Megagardeners of the forest - the role of different types of networks. Ecol. Lett. 14, 1170–1181 (2011). elephants in seed dispersal. Acta Oecol. 37, 542–553 (2011). 22. Guimerà, R. & Amaral, L. A. N. Functional cartography of complex metabolic 55. Bunney, K., Bond, W. J. & Henley, M. Seed dispersal kernel of the largest networks. Nature 433, 895–900 (2005). surviving megaherbivore-the African savanna elephant. Biotropica 49, 395–401 23. Donatti, C. I. et al. Analysis of a hyper-diverse seed dispersal network: (2017). modularity and underlying mechanisms. Ecol. Lett. 14, 773–781 (2011). 56. Schupp, E. W., Jordano, P. & Gómez, J. M. Seed dispersal effectiveness revisited: 24. Mucha, P. J., Richardson, T., Macon, K., Porter, M. A. & Onnela, J.-P. a conceptual review. New Phytol. 188, 333–353 (2010). Community structure in time-dependent, multiscale, and multiplex networks. 57. Guimarães, P. R., Galetti, M. & Jordano, P. Seed dispersal anachronisms: Science 328, 876–878 (2010). rethinking the fruits extinct megafauna ate. PLoS ONE 3, e1745 (2008). 25. Bassett, D. S. et al. Dynamic reconfiguration of human brain networks during 58. Howe, H. F. & Smallwood, J. Ecology of seed dispersal. Annu. Rev. Ecol. Syst. learning. Proc. Natl Acad. Sci. USA 108, 7641–7646 (2011). 13, 201–228 (1982). 26. Emer, C., Memmott, J., Vaughan, I. P., Montoya, D. & Tylianakis, J. M. Species 59. Escribando-Avila, G., Lara-Romero, C., Heleno, R. & Traveset, A. in Seed roles in plant–pollinator communities are conserved across native and alien Dispersal Networks in the Tropic (eds Datillo, W. & Rico-Gray, V.) (Springer, ranges. Divers. Distrib. 22, 841–852 (2016). New York). 27. Jordan, F. Keystone species and food webs. Philos. Trans. R. Soc. B Biol. Sci. 60. Bascompte, J. & Jordano, P. Mutualistic Networks: Monographs in Population 364, 1733–1741 (2009). Biology (Princeton University Press, Princeton, NJ, 2014). 28. Martín González, A. M., Dalsgaard, B. & Olesen, J. M. Centrality measures and 61. Jordano, P. Sampling networks of ecological interactions. Funct. Ecol. 30, the importance of generalist species in pollination networks. Ecol. Complex 7, 1883–1893 (2016). 36–43 (2010). 62. Wilson, E. O. & Naskrecki, P. A Window on Eternity: A Biologist’s Walk 29. De Domenico, M., Sole-Ribalta, A., Omodei, E., Gomez, S. & Arenas, A. Through Gorongosa National Park (Simon and Schuster, New York, 2014). Ranking in interconnected multilayer networks reveals versatile nodes. Nat. 63. Cillie, B. Pocket Guide to Mammals of Southern Africa (Sunbird Publishers, Commun. 6, 6868 (2015). Verlag, 2010). 30. McDonald-Madden, E. et al. Using food-web theory to conserve ecosystems. 64. Murray, K. D. Scatalog: Quick ID Guide to Southern African Animal Droppings Nat. Commun. 7, 10245 (2016). (Struik Nature, Cape Town, 2011). 31. Schleuning, M., Blüthgen, N., Flörchinger, M., Braun, J. & HM Specialization 65. Heleno, R. H., Ross, G., Everard, A., Memmott, J. & Ramos, J. A. The role of and interaction strength in a tropical plant-frugivore network differ among avian ‘seed predators’ as seed dispersers. Ibis 153, 199–203 (2011). forest strata. Ecology 92,26–36 (2011). 66. Chao, A. Estimating the population size for capture-recapture data with 32. Stalmans, M., Peel, M. & Massad, T. Aerial Wildlife Count of the Parque unequal catchability. Biometrics 43, 783 (1987). Nacional da Gorongosa. (2014). 67. Oksanen, J. et al. Vegan: community ecology package R package version 2. 4–1. 33. Correia, M., Timóteo, S., Rodríguez-Echeverría, S., Mazars-Simon, A. & https://CRAN.R-project.org/package=vegan (2013). Heleno, R. Refaunation and the reinstatement of the seed-dispersal function in 68. R Core Team. R: a language and environment for statistical computing (R Gorongosa National Park. Conserv. Biol. 31,76–85 (2017). Foundation for Statistical Computing, Vienna, 2015). 34. Matlack, G. R. Plant species migration in a mixed-history forest landscape in 69. Dormann, C. F., Frund, J., Bluthgen, N. & Gruber, B. Indices, graphs and null Eastern North America. Ecology 75, 1491–1502 (1994). models: analyzing bipartite ecological networks. Open Ecol. J. 2,7–24 (2009). 35. Hanski, I. Metapopulation dynamics. Nature 396,41–49 (1998). 70. Vázquez, D. P., Morris, W. F. & Jordano, P. Interaction frequency as a 36. Guimerà, R. & Amaral, L. A. N. Cartography of complex networks: modules surrogate for the total effect of animal mutualists on plants. Ecol. Lett. 8, and universal roles. J. Stat. Mech. Theory Exp. 2005, P02001 (2005). 1088–1094 (2005). 37. Jutlla, I. S., Jeub, L. G. & Mucha, P. J. A Generalized Louvain Method for 71. Shea, K. How the wood moves.pdf. Science 1231, 1231–1233 (2007). Community Detection Implemented in MATLAB http://netwiki.amath.unc.edu/ GenLouvain (2014). 10 NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y ARTICLE 72. Hervé, M. RVAideMemoire: diverse basic statistical and graphical functions R Author contributions Package version 0.9-68. https://CRAN.R-project.org/package=RVAideMemoire S.T., R.H., H.F., and S.R.-E. designed the study; S.T. and M.C. collected the data; S.T. led (2016). the analysis and wrote the first draft; and all authors contributed to manuscript revisions. 73. Bioinformatics toolbox: user’s guide R2016b. Version 9.1.0.441655 (The MathWorks Inc., Natick, Massachusetts, 2015). Additional information 74. Blüthgen, N., Menzel, F. & Blüthgen, N. Measuring specialization in species Supplementary Information accompanies this paper at https://doi.org/10.1038/s41467- interaction networks. BMC Ecol. 6, 9 (2006). 017-02658-y. 75. Poisot, T., Canard, E., Mouquet, N. & Hochberg, M. E. A comparative study of ecological specialization estimators. Methods Ecol. Evol. 3, 537–544 (2012). Competing interests: The authors declare no competing financial interests. 76. Bates, D., Mächler, M., Bolker, B. & Walker, S. Fitting linear mixed-effects models using lme4. J. Stat. Softw. 67,1–48 (2015). Reprints and permission information is available online at http://npg.nature.com/ 77. Fox, J. & Weisberg, S. An R Companion to Applied Regression (Sage reprintsandpermissions/ Publications Inc., Thousand Oaks, CA, 2011). 78. Hothorn, T., Bretz, F. & Westfall, P. Simultaneous inference in general Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in parametric models. Biom. J. 50, 346–363 (2008). published maps and institutional affiliations. 79. De Domenico, M., Porter, M. A. & Arenas, A. MuxViz: a tool for multilayer analysis and visualization of networks. J. Complex Netw. 3, 159–176 (2015). 80. Brin, S. & Page, L. The anatomy of a large-scale hypertextual web search engine. Comput. Netw. 56, 3825–3833 (2012). Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Acknowledgements Commons license, and indicate if changes were made. The images or other third party This work was supported by FEDER funds from the “Programa Operacional Factores de material in this article are included in the article’s Creative Commons license, unless Competitividade – COMPETE” and by national funds from the Portuguese Foundation for Science and Technology – FCT through the research project PTDC/BIA-BIC/4019/ indicated otherwise in a credit line to the material. If material is not included in the 2012. S.R.-E., R.H., and M.C. were also supported by FCT (Grants IF/00462/2013, IF/ article’s Creative Commons license and your intended use is not permitted by statutory 00441/2013, and SFRH/BD/96050/2013, respectively). We thank Greg Carr, and the Greg regulation or exceeds the permitted use, you will need to obtain permission directly from Carr Foundation – Gorongosa Restoration Project, Dr Marc Stalmans and all the staff the copyright holder. To view a copy of this license, visit http://creativecommons.org/ from Gorongosa National Park for their logistic assistance during fieldwork, and for licenses/by/4.0/. sharing their knowledge and passion about Gorongosa. © The Author(s) 2017 NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications 11 | | | http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Nature Communications Springer Journals

Multilayer networks reveal the spatial structure of seed-dispersal interactions across the Great Rift landscapes

Loading next page...
 
/lp/springer-journals/multilayer-networks-reveal-the-spatial-structure-of-seed-dispersal-5rejw0Vhwt

References (96)

Publisher
Springer Journals
Copyright
Copyright © 2017 by The Author(s)
Subject
Science, Humanities and Social Sciences, multidisciplinary; Science, Humanities and Social Sciences, multidisciplinary; Science, multidisciplinary
eISSN
2041-1723
DOI
10.1038/s41467-017-02658-y
pmid
29321529
Publisher site
See Article on Publisher Site

Abstract

ARTICLE DOI: 10.1038/s41467-017-02658-y OPEN Multilayer networks reveal the spatial structure of seed-dispersal interactions across the Great Rift landscapes 1 1 1 1 1 Sérgio Timóteo , Marta Correia , Susana Rodríguez-Echeverría , Helena Freitas & Ruben Heleno Species interaction networks are traditionally explored as discrete entities with well-defined spatial borders, an oversimplification likely impairing their applicability. Using a multilayer network approach, explicitly accounting for inter-habitat connectivity, we investigate the spatial structure of seed–dispersal networks across the Gorongosa National Park, Mozam- bique. We show that the overall seed–dispersal network is composed by spatially explicit communities of dispersers spanning across habitats, functionally linking the landscape mosaic. Inter-habitat connectivity determines spatial structure, which cannot be accurately described with standard monolayer approaches either splitting or merging habitats. Multi- layer modularity cannot be predicted by null models randomizing either interactions within each habitat or those linking habitats; however, as habitat connectivity increases, random processes become more important for overall structure. The importance of dispersers for the overall network structure is captured by multilayer versatility but not by standard metrics. Highly versatile species disperse many plant species across multiple habitats, being critical to landscape functional cohesion. CFE – Centre for Functional Ecology, Department of Life Sciences, University of Coimbra, Calçada Martim de Freitas, 3000-456 Coimbra, Portugal. Correspondence and requests for materials should be addressed to S.T. (email: stimoteo@gmail.com) NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications 1 | | | 1234567890():,; ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y ver the recent decades, ecological networks have proved a Contrarily to the high spatial turnover in species and interac- valuable framework to simultaneously evaluate the role of tions, the functional role of species is considered relatively Ospecies, their interactions, and the importance of the stable . Centrality measures have been largely used to assess the 27,28 emerging community structure for the persistence and stability of topological position of a species in the structure of networks . biological communities . Such studies revealed that ignoring the In a multilayer context, such overall centrality can be estimated complex web of interactions between plants and animals in which with Google’s PageRank algorithm, which has been successfully 29 30 many vital ecosystem functions are rooted might jeopardize the used to guide conservation strategies . 2,3 long-term functioning and persistence of ecosystems . To date, Here, we investigate how a mutualistic multilayer network is most studies have considered networks as entities with discrete structured across habitats and the importance of species to the borders defined by the experimental design, ignoring the potential cohesion of seed dispersal across a complex landscape. To this across-border connections , or alternatively as aggregations of end, we collected seed–dispersal interactions across the Gor- several spatially and temporal sampling occasions into an overall ongosa National Park, Mozambique, to build the most complete, 5 31 network . In nature, however, these sub-networks are linked by seed–dispersal network of the African continent to date , common species and by processes that span over several spatial including all potential guilds of seed dispersers. Gorongosa and temporal scales, contributing to the functional connectivity of underwent a severe defaunation that affected many of the large 6 32,33 ecosystems . The importance of the spatial dimension of net- herbivores, and its recovery is now en route . In this context, 7–9 works of interactions is becoming increasingly recognized , seed–dispersal is particularly vital for plants to recolonize newly highlighting the key function of species that cross habitat available patches or disturbed ground , and is likely a key driver 10 35 boundaries acting as mobile links that connect the different of long-term habitat dynamics in Gorongosa patchy landscapes . habitats. Recent work has provided further evidence of the Our objectives are twofold. First, we aim to explore the spatial importance of the often-neglected inter-habitat links and their distribution of seed–dispersal modules (i.e., communities of unequivocal ecological relevance . Perhaps ironically, the appli- tightly interacting plants and their dispersers) spanning across the cation of such tools that proved particularly suited to tackle the different habitats of the Gorongosa National Park. We will do so intrinsic complexity of ecosystems is limited by the amount of by evaluating the modularity of multilayer networks formed by complexity that can be sampled and analyzed, leading to a frag- discrete, yet interconnected layers representing different habitats. mentation of real networks, likely to result in oversimplifications, We used different null models to explore how the strength of the and eventually to incomplete or erroneous conclusions about interlayer connectivity affects the overall structure of the spatial 12,13 network structure, dynamics, and stability . Similarly, ignor- multilayer network, and to what extent this multilayer approach ing the role of different species as spatial couplers of ecosystems improves the currently used monolayer analyses of disconnected may hinder our understanding of natural processes, e.g., the flux and aggregated networks. Second, we aim to assess the relative of energy, or nutrients, between aquatic and terrestrial systems, contribution of each disperser species to the cohesion of seed pollen transfer by insects across the landscape, or the dispersal of dispersal across habitats. We will do so by exploring dispersers seeds of invasive species by birds . Recently, some authors multilayer versatility, which expresses their contribution to the have started to tackle this issue by treating different habitats, or mobile link function both within and between habitats. We dis- patches of habitat as a set of layers within a larger multilayer cuss the potential of this new metric by comparing it to the 15–17 network . An expansion of the concept of beta-diversity has information provided by traditional species-level descriptors. been proposed to measure dissimilarities between networks, by exploring species and interactions turnover between groups of Results 16,18 independent (i.e., formally disconnected) networks .Ina Overview of seed dispersal in Gorongosa. During this one year, further step, Frost et al. quantified the connectivity between we collected 1399 fecal samples (1174 mammal dung piles and spatial layers (habitats) of a host-parasitoid network, though they 236 bird droppings) produced by 98 animal species, of which 508 did not explore the effect of habitat connectivity to the structure (29%) had at least one undamaged seed. Overall, 12,159 unda- of the spatial network. However, only now ecologists have started maged seeds from 94 plant species were retrieved from the feces to explicitly include interlayer edges in the analysis of the actual of 29 dispersers, comprising 508 links. Focal observations pro- structure of “ecological multilayer networks” , taking advantage duced 85 further links (14% from the total), whereas camera traps of recent theoretical developments and analytic tools from other contributed with 15 new links (2.5%). In total, we compiled 608 13,19 research areas . links between 32 animal species and 101 plant species, in four 1,20,21 A key structural pattern in networks is modularity , habitats (Fig. 1). measuring the extent to which species form cohesive groups Overall, primates were responsible for most interactions, (modules) where species interact more often within the same namely Papio ursinus (chacma baboon, 35%) and Cercopithecus module than with species in other modules . These modules pygerythrus (vervet monkey, 10%), followed by Loxodonta provide insights into the phylogenetic history and trait con- africana (elephant, 22%) and Civettictis civetta (African civet, vergence of unrelated species, resulting from local co-adaption, 7%) (Fig. 1). The three most commonly dispersed plant species and ecological convergence in the use of resources . By mea- represented 41% of all recorded interactions, namely Ziziphus suring multilayer modularity, the connectivity between layers, i.e., mucronata (Rhamnaceae, 15%), Sclerocarya birrea (Anacardia- interlayer edge strength, is explicitly accounted for, with the ceae, 13%), and Hyphaene natalensis (Arecaceae, 13%). advantages of detecting modules that span across layers. It also We estimated that our sampling effort captured 77% of the allows the identification of nodes that can belong to different disperser species and 44% of the plants with similar levels of modules in different layers, thus particularly relevant for main- sampling completeness across the four habitats (Supplementary taining the continuity of ecosystem functions in space or Fig. 1 and Supplementary Table 1). 12,24 time . However, the ideal way to quantify interlayer edge strength is still a matter of research in multilayer network research, and the investigation of the relative importance of intra- Modular structure of the spatial multilayer network. To eval- and interlayer processes is essential to understand the structure of uate the extent to which the seed–dispersal interactions are sorted 12,24,25 36 multilayer networks . into distinct communities of tightly interacting species ,we 24,37 calculated the multilayer modularity of the spatial network of 2 NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y ARTICLE Birds Carnivores Elephant Insects Primates Rodents Antelopes 15 2 3 4 6 7 Dispersers Plants a b c d e f Overall network ++ + Grassland Transition forest Mixed dry forest Miombo Fig. 1 Quantitative seed–dispersal network of the Gorongosa National Park, Mozambique. Both the aggregated (top) and the individual habitat (bottom) networks are based on the same sampling effort and are represented on the same scale. The boxes in the top level represent disperser species and those on the bottom level represent the plant species dispersed. The gray lines linking the two levels represent pairwise species interactions, and their width proportional to the interaction frequency. The aggregated network was obtained by pooling all interactions across the four habitats, and summing their frequencies. Main seed dispersers: 1. Pycnonotus tricolor,2. Civettictis civetta,3. Loxodonta africana,4. Cercopithecus pygerythrus,5. Papio ursinus,6. Hystrix africaeaustralis, and 7. Redunca arundinum. Most commonly dispersed plants: a Centaurea praecox,b Grewia inaequilatera,c Hyphaene natalensis,d Sclerocarya birrea,e Tamarindus indica, and f Ziziphus mucronata. The full list of species can be seen in Fig. 3 and Supplementary Fig. 2, for animals and plants, respectively. The silhouettes used in this figure are all sourced from Open Clipart and were made available under a CC0 1.0 licence Gorongosa (see “Methods” section and Supplementary Methods predicted by both null models tended to converge to that of the for details on the multilayer modularity algorithm). Using a observed network at very high values of interlayer strength multilayer formalism , this network is defined by the animal (Fig. 2a and Supplementary Data 1), indicating an increasing seed–dispersal interactions (intralayer links) in each habitat importance of random processes in structuring the networks. (layer), with habitat connectivity (interlayer links) provided by This suggests that when habitat connectivity is very high the the common species. Ultimately, interlayer links should be overall network structure becomes less determined by the identity interpreted as the movement of matter or energy between layers, of animals connecting them, and might be more contingent on in our case the effective movement of animals and seeds between the structure of seed dispersal within habitats. habitats, and quantified in a way that estimates the intensity of To understand the added value of the multilayer approach in these movements (interlayer strength). Multilayer modularity was relation to the traditional monolayer approach, we compared the calculated across a range of interlayer strength (0–10), assuming results from the multilayer analysis with those provided by the that any co-occurring species between habitats effectively con- currently standard approaches of either merging all data into a nected them with the same intensity, to test how the structure of single aggregated network (Q ), in which interactions aggregated the spatial network is affected by habitat connectivity. The mul- occurring at multiple habitats are summed across habitats, or tilayer modularity of the Gorongosa seed–dispersal network was considering each habitat as a discrete and disconnected network. very high across the whole range of values of habitat connectivity The structure of the aggregated network is influenced by the (Q = 0.903–0.993; Fig. 2a), with an overall increasing distribution of the interactions among the species, with multilayer trend toward an asymptote just below 1. We used two null models modularity being significantly lower than predicted by the (see “Methods” section for details) to test how the structure of the intralayer null model (mean Q = 0.43 vs. Q = aggregated null models spatial network is influenced by the seed–dispersal process within 0.59, p < 0.001; Supplementary Fig. 3). However, it ignored each habitat (intralayer null model) and by the identity of the habitat connectivity because it cannot incorporate such informa- animals connecting these habitats (interlayer null model). The tion. In the disconnected network, habitats are considered totally structure of the empirical network, across the range of habitat independent from each other, thus equivalent to calculate connectivity values, was statistically different than that predicted modularity for each of them . Modularity was similar or slightly by both null models, though in opposite directions: reshuffling higher than that of the aggregated network and much lower than interactions within each habitat (intralayer null model) over- that of the multilayer network, ranging from 0.43, in the Mixed estimated modularity, whereas reshuffling the identity of the forest, to 0.56, in the Grassland (Fig. 2 and Supplementary Fig. 3). habitat-connecting animals in each habitat (interlayer null model) The number of modules detected in the multilayer structure is underestimated modularity (Fig. 2a). The identity of the dis- mostly constant, oscillating between 11 and 12 across most values persers and the intensity of movements between habitats (inter- of interlayer strength, except for very small values, where some layer strength) play a more important role for the spatial structure additional modules were detected (Figs. 2b, 3, and 2). The of the seed–dispersal network than the pattern of seed dispersal intralayer null model consistently predicted more modules than within each individual habitat. Nonetheless, the modularity observed, while the interlayer null model consistently predicted NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications 3 | | | ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y away bird, is consistently assigned to the same module of the bush 1.5 *** babies, Fig. 3). 1.4 The strength of each interaction can vary across habitats, reflecting different animal resource preferences in different 1.3 contexts, and therefore, species can change their module 24,38 affiliation between habitats . We calculated species adjust- 1.2 ability as the proportion of animal or plant species that switch Intra-layer null model module affiliation at least once between any pair of habitats . 1.1 Observed Most species do not change module affiliation across habitats, 1.0 exhibiting a relatively low or non-existing adjustability (Supple- n.s. mentary Fig. 4). When the intensity of species movement between 0.9 Inter-layer null model habitats (interlayer strength) is low, animals and plants tend to ** interact with distinct set of species in each habitat and a higher 0.8 proportion of species will change their module affiliation between habitats. As the intensity of these movements intensifies, and *** habitat connectivity increases, species adjustability becomes negligible and interactions tend to occur among the same species across all habitats. However, this stabilization on interaction 20 partners happens at different levels of habitat connectivity for animals and plants (Fig. 3; Supplementary Fig. 4; and Supple- Intra-layer null model mentary Data 1). For both animal and plant species, adjustability was generally more affected by the identity of the animal (interlayer null model) than by the pattern of interaction within Inter-layer null model Observed habitats (intralayer null model). For low interlayer strength, animal adjustability was significantly lower than predicted by the 10 intralayer null model, but higher than predicted by the intralayer null model (Supplementary Fig. 4). However, both null models *** performed better at greater values of interlayer strength. The 02468 10 interaction pattern within habitats (intralayer null model) had a Inter−layer strength variable effect on plant adjustability: the observed plant adjust- ability was significantly higher for very low, but also for high Fig. 2 Modularity and number of modules observed in empirical networks habitat connectivity, but lower observed adjustability between and predicted by two different null models. Mean maximized modularity (a) these values. The interlayer null model consistently predicted and mean number of modules (b) of the observed networks (black) across significantly higher plant adjustability for lower interlayer the range of interlayer strength (0–10), and comparison against the two null strengths (Supplementary Fig. 4). Thus, animals are more likely models: intralayer null model (blue), and interlayer null model (red to disperse the same plant species across habitats than plant symbols). Values presented as the mean (±SEM) of 100 runs of the species are to rely on the same dispersers, and animal movement modularity function, for each interlayer strength. The significance of the across the landscape exerts a stronger influence in the spatial observed modularity and number of modules was compared against those structure of the seed–dispersal network. of the null networks with a one-sample t test. *p < 0.050, **p < 0.010, ***p < 0.001, n.s. non-significant. Full results are presented in Supplementary Data 1 Contribution of disperser species to seed dispersal cohesion. We did not detect differences on animal species richness across fewer modules than observed, across the whole range of interlayer the four main habitats of Gorongosa (G test: G = 1.84, p = 0.61; strength (Fig. 2b and Supplementary Data 1). In the aggregated Fig. 4a and Supplementary Table 2). Mixed forest holds a greater network, the average number of modules detected was 11.4, richness of plants than the other three habitats, but this was only which was in line with those detected in the multilayer network significant in comparison to Grassland and Miombo (Fig. 4b and (Supplementary Fig. 3 and Fig. 2), and significantly higher than Supplementary Table 2). As for richness of interactions, Mixed those predicted by the intralayer null model (mean modules forest had more interactions than Transition forest, and both observed = 11.4 vs. null model = 9.2; t(99) = −28.60, p < 0.001). habitats had more interactions than Grassland and Miombo (all The mean number of modules in each habitat of the disconnected pairwise G tests: p < 0.002; Fig. 4c and Supplementary Table 2). network was variable and ranged from 6 to 13 (Supplementary Dispersers’ specialization did not differ significantly among Fig. 3 and Fig. 3). In the spatial multilayer network, modules are habitats (Χ = 2.49, df = 3, p = 0.49; Fig. 4d and Supplementary subsets of species that strongly interact across the different layers Table 3). 25,39 of the network . For animals, this corresponds to species that We calculated each disperser multilayer versatility, which is occur and disperse seeds from the same plant species in more equivalent to an overall measure of centrality to identify those than one habitat (Fig. 3 and Supplementary Fig. 2). For example, that are topologically important to the structure of the spatial most primates (baboon, vervet monkey, and Otolemur crassicau- network . For this effect, we used a unimodal projection of the datus (bush baby)) all disperse Z. mucronata and are consistently network, in which two animal species are connected if they placed in the same module in the multilayer and in the aggregated disperse the same plant species , thus providing an insight over networks, but not when habitats are weakly connected or their likely “functional redundancy” . Links between species were considered independent. It is worth to note that module quantified by weighting the number of shared interactions by the 42,43 affiliations do not necessarily group phylogenetically related assemblage size , minimizing the loss of information asso- species, but species that feed on similar resources, which in seed ciated with unimodal projections . Multilayer versatility revealed dispersal might be mostly determined by behavioral and that few dispersers are disproportionately important, namely the morphological constraints (e.g., Corythaixoides concolor, the go- baboon and the elephant, followed by a long tail of species with 4 NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications | | | Mean maximized modularity Mean number of modules NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y ARTICLE Multilayer Non-multilayer = 0.1  = 0.5  = 1.0  = 2.0  = 10.0 Tragelaphus sylvaticus Tragelaphus strepsiceros Tragelaphus angasii Tockus nasutus Tchagra senegalus Redunca arundinum Pycnonotus tricolor Psammodes sp. Potamochoerus larvatus Phacochoerus africanus Papio ursinus Panthera leo Ourebia ourebi Otolemur crassicaudatus Oriolus larvatus Numida meleagris Mongoose Loxodonta africana Lagonosticta rhodopareia Kobus ellipsiprymnus Hystrix africaeaustralis Hippotragus niger Genetta tigrina Corythaixoides concolor Connochaetes taurinus Civettictis civetta Chlorocichla flaviventris Cercopithecus pygerythrus Cephalophus natalensis Ant Andropadus importunus Aepyceros melampus Fig. 3 Module affiliation of animal species in the spatial multilayer network of Gorongosa. Module affiliation is shown for five different interlayer edge strengths (left block), and for the monolayer networks, considering either the aggregated network or each habitat individually (right block). In each case, the run with the highest maximized modularity was used (module affiliation for plant species is shown in Supplementary Fig. 2). Within each network, different colors represent different modules. Colors in different blocks are independent lower versatility (Fig. 5a). The importance of these species comes multilayer versatility and both dispersers mean specialization (d′) from being central in the structure of the seed–dispersal network and the number of habitats where they occur (r = −0.255, because they share plant partners with many other animals, but p = 0.208, Fig. 5a; r = 0.383, p = 0.053, respectively, Supplemen- also because they share plant species across different habitats. The tary Table 4). However, dispersers multistrength was only versatility of dispersers in the multilayer network was correlated moderately correlated with their importance (i.e., versatility) on with their versatility in the aggregated network (r = 0.671, the multilayer network (r = 0.514, p = 0.007; Supplementary s s p < 0.001; Fig. 5b and Supplementary Table 4), but the Table 4). Species multistrength extends the concept of its importance of species with low versatility is underestimated in monolayer counterpart, expressing the total number of links of the aggregated network (Fig. 5b). There were relatively few shared a species across all layers of the network , i.e., the total shared links among habitats (total edge overlap = 8.2%). However, all interactions with all its neighboring species across the habitats. habitat pairs, except Miombo and Grassland, shared more than However, contrary to versatility, multistrength does not account 20% of the interactions (Fig. 6). for the distribution of these links in relation to the other species, We evaluated if the information condensed by multilayer or the number of layers in which these links occurs. Thus, versatility could be captured by other species-level metrics, although both metrics are related, multistrength will not reflect namely specialization d′, number of habitats, and species the importance of a species for the overall structure of the multistrength. We did not find a significant correlation between multilayer network as much as versatility. NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications 5 | | | Grassland Miombo Mixed forest Transition Grassland Miombo Mixed forest Transition Grassland Miombo Mixed forest Transition Grassland Miombo Mixed forest Transition Grassland Miombo Mixed forest Transition Aggregated Grassland Miombo Mixed forest Transition ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y a b 25 75 aaaa ab ab a b 20 60 10 30 5 15 0 0 c d 250 1.0 aa b c aa a a 200 0.8 0.6 0.4 50 0.2 0 0.0 Fig. 4 Network- and species-level descriptors of the interactions from each habitat of Gorongosa. Differences among the main habitats of Gorongosa in terms of a animal species richness, b plant species richness, c number of interactions, d species specialization (mean ± SEM). Different letters indicate statistical pairwise differences: a–c, pairwise G tests (see Supplementary Table 2 for full results); d, generalized linear mixed models (63 observations/ occurrences of 32 animal species in four habitats; Supplementary Table 3 for full results) 15,16,18 Discussion habitats , has long been recognized as critical for the Species and communities are not randomly distributed across the dynamic of patchy habitats across complex landscapes . Never- planet, but they are strongly structured by spatial attributes tra- theless, and despite the current interest on species interactions ditionally recognized by ecologists (e.g., niches, habitats, land- networks, these are yet to explicitly accommodate this interlayer scapes, and biomes). Traditionally, species interaction networks dynamic when analyzing the structure of spatial networks. have been studied as discrete entities with borders defined by the Here, we implement for the first time a multilayer approach to researchers based on different landscape attributes. However, evaluate the spatial structure of an ecological network explicitly species interactions do not abruptly finish at habitat borders, and incorporating the interlayer strength connecting networks from therefore the decision of merging or segregating data from these adjacent habitats. Our spatial multilayer seed–dispersal network spatial units is far from trivial. Nevertheless, ecologists are still exhibited a highly modular structure, i.e., species tend to interact faced with a paucity of tools to evaluate when such combination with subsets of species (i.e., modules) within subsets of spatially of data is useful, or when it might increase the noise around the coupled habitats. By explicitly including non-zero interlayer links, patterns of interest, thus obscuring important conclusions. i.e., the habitat connectivity promoted by the common species, it Although still based on the recognition of different habitats, the is possible to account for the interdependence of the network implementation of a multilayer approach provides a valuable tool structure across multiple habitats , and identify modules that that allows for better decisions regarding the merits of segregating spread across habitat borders . or merging spatially (or temporal) explicit data. For a more realistic module detection, interlayer strength However, interlayer connectivity has never been explicitly should ideally be measured empirically to reflect the effective incorporated in the analysis of the modular structure of spatial movement of the individual species across habitat borders . ecological networks. In this study, we investigate the spatial Unfortunately, obtaining such data at the community and land- structure of a seed–dispersal network spanning across multiple scape levels, i.e., all species, across all habitats can be incredibly habitats explicitly considering interlayer connectivity. We made challenging. The alternative of assigning the same interlayer use of a highly comprehensive data set collected in a highly strength to all species, i.e., assuming that all species connect diverse African landscape including all potential disperser guilds. habitats with the same intensity, is a clearly undesirable simpli- This adds to the sparse knowledge of seed dispersal in Africa, but fication as species connect habitats with different intensities has direct implications for our understanding of seed–dispersal because of their differential ability to move across and establish in 46,47 networks across the globe. a given habitat . Incorporating such empirical data could have Our results show a highly modular structure of the spatial important implications on modules found by the modularity multilayer network that is influenced by the strength of the function; the relative importance between intra- and interlayer connectivity between habitats, with about half the communities of process would be different for each individual species, thus seed dispersers detected bridging most of the habitats (Figs. 3 and affecting its probability of changing module affiliation . 4). The network is dominated by a few highly versatile species Exploring the modular structure across a range of interlayer edge that secure both local (habitat level) and global (landscape level) strengths is an alternative to obtaining empirical data, and has dispersal of seeds, ensuring the spatial continuity of the seed– been often done in other fields to understand its importance for 24,25,38,48 dispersal process. processes spanning across different layers . Our analysis Landscapes are intrinsically dynamic, being constantly shaped revealed that the modularity and the number of modules are by local disturbance and ecological succession . Understanding mostly affected at extremely low levels of interlayer strength, how animals move between habitats providing key mobile links , suggesting that the spatial community structure can be main- and how ecological interactions are distributed across tained even if the strength of the habitat connectivity is low 6 NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications | | | Grassland Transition Mixed forest Miombo Grassland Transition Mixed forest Miombo No. of interactions Animal richness Mean specialization d ′ Plant richness NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y ARTICLE a b 1.0 1.0 rho = −0.255, p = 0.208 rho = 0.671, p < 0.001 0.8 0.8 0.6 0.4 0.6 0.2 0.4 0.0 0.2 0.2 0.4 0.6 0.8 1.0 Multilayer versatility Fig. 5 Correlates of multilayer versatility. Correlation between animal species versatility in the multilayer network (bars) and the mean specializationd’ (dots) (a), and between multilayer versatility of the monolayer versatility of the aggregated network (b). Full data available in Supplementary Table 4 Shared interactions within and between layers, which is objectively defined by the relative strength of the inter- to intralayer edges. Regarding the modules' composition, these grouped-together Miombo Transition forest species that are not always phylogenetically close (e.g., primates 0.27 were grouped with the go-away bird), suggesting that functional and morphological matching, such as gape-size and seed/fruit size, are more important drivers of seed–dispersal interactions . Interestingly, we detected low adjustability for most species and 0.12 module affiliations remained mostly constant across habitats. Module switching occurred only for some species (e.g., primates, elephants, or civets), and at very low values of habitat con- 0.37 nectivity (Fig. 3). Most animals, however, tend to disperse the same plant species in different habitats, thus maintaining a Mixed dry forest Grassland similar functional role across the landscape , even if habitat connectivity is very low. This can only be detected if the habitats 0.22 are explicitly linked in the analysis of network structure. Resource availability largely determines animal movements at the land- scape level . In turn, the inter-habitat movement of dispersers is likely to affect plant regeneration dynamics, and thus resource Fig. 6 Similarity in terms of shared interactions, between the habitat pairs availability. The capacity of species to adjust their interactions to of Gorongosa. Similarity is assessed using the edge overlap between the specific contexts (thus increasing overall adjustability) is likely habitats in the spatial multilayer network, i.e., proportion of shared links important for species persistence in changing environments, across habitats. The width of the links is proportional to the correlation while at the same time tends to promote a greater connectivity between each habitat pair. The silhouettes used in this figure are all (e.g., seed dispersal) across habitats. In Gorongosa, some of the sourced from Open Clipart and were made available under a CC0 1.0 species that changed module affiliation have generally wide-range licence movements and can distribute seeds between habitats, thus giving a key contribution to plant genetic diversity and spatial dis- 34,49 relative to the strength of the interactions within the habitats. The tribution of plant populations through seed dispersal . This is inter- to intralayer edge strength ratio is essential for the outcome particularly true for intrinsically large-scale processes, such as of modularity estimation, affecting the probability of nodes being seed dispersal, while other processes, such as belowground 38,48 assigned to distinct modules . As such, when interlayer cou- mycorrhizal associations, seem to be structured at much finer pling increases, the number of detected modules is expected to scales . 24,25 decrease . Importantly, the structure of the seed–dispersal Although the number of interactions differed among habitats, network was not fully captured by the aggregated network or by and the overall edge overlap was relatively low, neither the considering each habitat as an independent network. Conse- richness of dispersers, the mean number of dispersed plants, or quently, the result obtained by using a multilayer approach is not dispersers’ specialization varied significantly. biased by any decision regarding aggregating or disconnecting the Multilayer versatility allowed us to identify the animal species different layers of the network. Instead, the resulting structure is a that are most important for dispersing seeds simultaneously at the consequence of the relative importance of the processes occurring NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications 7 | | | x = y Versatility/Specialization d ′ Papio ursinus Loxodonta africana Cercopithecus pygerythrus 0.33 Civettictis civetta Phacochoerus africanus Aepyceros melampus Hystrix africaeaustralis Redunca arundinum Chlorocichla flaviventris Andropadus importunus Ourebia ourebi Mongoose Hippotragus niger Oriolus larvatus Kobus ellipsiprymnus Numida meleagris Connochaetes taurinus 0.28 Cephalophus natalensis Tragelaphus sylvaticus Corythaixoides concolor Otolemur crassicaudatus Genetta tigrina Potamochoerus larvatus Pycnonotus tricolor Tragelaphus strepsiceros Tragelaphus angasii Aggregated versatility ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y local/habitat scale and at the global/landscape scale by linking The effective conservation and restoration of natural areas multiple habitats. In the Gorongosa landscape, primates, ele- requires an integrated view of how species and their interactions phants, and African civets are central nodes in the network and maintain functional ecosystems on complex landscapes, and a likely important for its stability and cohesion . Although the multilayer approach is a most valuable tool to explore these versatility of the aggregated network can correctly identify the factors. Here, we took a step further in the analysis of spatial most important dispersers, it underestimates the importance of mutualistic networks, and using interlayer edges strength, we species that are restricted to one or a few habitats (e.g., Genetta explicitly considered the interactions between plants and their tigrine (genet) or Tragelaphus strepsiceros (kudu); Supplementary dispersers across multiple habitats in the analysis of the network Fig. 5). The relatively low correlation between multilayer versa- structure. Furthermore, we identified key spatial coupler species, tility and multistrength, and the non-significant relationship with which play a pivotal role in long-term vegetation dynamics in the number of habitats where each species occurs and its spe- Gorongosa by ensuring the dispersal of genes across the land- cialization (d′)reflect the information gain of using multilayer scape. These key spatial couplers, namely primates, elephants and versatility, which could not be captured by conventional metrics. African civets, should be highly regarded in the protection of this The most versatile seed dispersers of Gorongosa were those essential ecosystem service. switching module affiliation between habitats. They are known As many other types of networks, mutualistic networks are not 52–54 for incorporating a high proportion of fruits into their diets , temporally static, and they do not abruptly stop at habitat bor- having relatively long gut retention times, and traveling for long ders . Therefore, forcing the analysis of biotic interactions into 53–55 distances . This allows seeds to escape the high intraspecific spatially delimited “network snapshots” and ignoring habitat competition near their parent plants by diversifying the deposi- connectivity will inevitably limit the insights that can be gained tion site of ingested seeds, and thus increasing their chances of by this approach . Here, we show that a multilayer approach can 56 32 successful recruitment . Baboons are ubiquitous in Gorongosa be used to link ecological processes that occur in different spatial and assume a central role as seed dispersers across the whole layers of a network providing insights that may not be captured park, and elephants, whose populations are still recovering in using traditional representations of monolayer networks. Over- Gorongosa , are also essential to the dispersal of plant species, looking the multiple relationships between nodes on different particularly those with very large fruits and seeds (e.g., H. nata- layers may lead to an inaccurate network structure, but also to lensis or Borassus aethiopum) . Surprisingly, despite being locally misidentification of the real role of species in the whole-network 19,24,29 abundant and often considered important seed dispersers in structure . Alternatively, the explicit incorporation of the many ecosystems , birds had a low versatility. While bird ver- temporal and spatial dynamics into a multilayer network satility might have been underestimated as a result of the different approach represents an important next step in the study of sampling method used (mist-netting), the low proportion of bird animal-plant interactions. droppings with seeds (only 7 out of the 96 bird species captured where found to disperse seeds, and these were only found in 19 Methods out of the 236 bird droppings analyzed), and the consistently low Field site and sampling. This work was carried out in the Gorongosa National sampling completeness estimated to all sampling methods Park, Mozambique (hereafter Gorongosa), a hyperdiverse park covering 4067 (transects: 25%, mist-netting: 13%, and focal observations: 13%) 2 km at the southern end of the Great Rift Valley (18°47′43.2d″S 34°28′09.1″E). We suggests that the lower bird versatility is actually structural rather defined four major habitats based on the vegetation structure and flooding regime: (1) grassland, periodically inundated grassland, with few shrubs and virtually no than a sampling artifact. Potential biases may emerge if any trees; (2) Transition forest, characterized by short flooding periods and dominated particular animal or plant groups are under or oversampled due by trees of Faidherbia albida, H. natalensis,or Acacia xanthophloea with a mostly to the use of different sampling methods. This problem might be open understory; (3) Mixed forest, occasionally flooded and formed by a diverse countervailed by performing analysis on rarefied or unweighted mixture of tree species with a dense and closed understory; and 4) Miombo, one of the most extensive habitats in Africa, unaffected by floods, and dominated by networks, though these are subjected to their own caveats. We are Brachystegia spp. trees with a dense understory (see ref. for details). Throughout only aware of a single study that explored the potential con- a year, we reconstructed seed–dispersal interactions from the four habitats by sequences of merging data originated from different sampling retrieving intact seeds from animal dung collected in the field. Sampling took place methods in the assembly of seed–dispersal matrices ; the authors in 12 occasions, evenly spaced between June 2014 and May 2015, except from December to February when floods make the park mostly inaccessible. In order to concluded that this was in fact beneficial due to the com- sample all potential disperser guilds, we employed complementary sampling pro- plementarity of the methodologies. Moreover, it must be noted tocols. In each sampling occasion, one transect ca. 2000 m and 5 m wide was that the consistently low estimates of sampling completeness are performed in each habitat and separated from any other transects by at least 350 m. likely to be largely underestimated for at least two reasons: first, Overall, 48 transects were run (ca. 96 km), corresponding to a surveyed area of because species accumulation curves are based on the assumption 480,000 m . A team of two observers simultaneously collected animal dung sam- ples, corresponding to the deposition of a dung pile of a single animal, and that communities are closed, an assumption that is not met in identified the disperser species by direct observation of defecating animals, or using year-round studies, where new species “enter” the community of 63,64 the expertise of local park rangers and field guides , and recorded direct potential interactions as a consequence of advancing phenology observations of animals ingesting fruits. Bird dispersal was evaluated by collecting (i.e., fruiting season); and second, because a large proportion of droppings from birds captured during mist-netting sessions on each sampling occasion, run for 5 h after dawn. Birds were kept inside individual holding bags, the potential interactions will never be detected because they are and released after producing a dropping . All samples were carefully screened not really possible due to spatial, temporal, and phenological under a ×40 magnifying microscope, and all undamaged seeds identified against a 60,61 mismatch between species . These “forbidden links” (or “true reference collection of seeds/fruits collected in the field. Seeds that could not be zeros”) can amount to a large proportion of the unobserved identified visually were barcoded, and their DNA sequences compared against online databases, with species identified based on the best match and on a checklist potential links (up from 44 to 77%) in seed–dispersal networks . of Gorongosa plants (see Supplementary Information for details). Most seeds Interestingly, the significant edge overlap between habitat pairs (92%) were identified to the species level, 7% to genus level, and less than 1% to the (i.e., the proportion of shared links) confirms the a priori family level or grouped into morphotypes, hereafter referred to as “species” for assumption that seed dispersal does not stop at habitat borders. simplicity. Sampling was further complemented with the analysis of motion- triggered camera traps, operating for five nights per habitat per sampling occasion Taken together with the results from multilayer modularity, to record interactions from non-conspicuous animals or those feeding at night. We showing that species tend to maintain their module affiliation estimated sampling completeness for animal and plant species in each habitat. The throughout the landscape, these results suggest that species traits use of different methods to record interactions could result in different interaction (such as mobility and size) may largely determine their multilayer sampling completeness, thus we also estimated completeness of interaction sam- versatility. pling for each method. This was done by estimating the proportion of observed 8 NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y ARTICLE presented as the proportion of networks generated by the null models with richness in relation to the total asymptotic richness estimated by the non- 66 67 modularity lower than the observed networks. For each network, we calculated the parametric estimator Chao2 using function specpool from package vegan ,in R 68 mean number of modules and the mean adjustability of animal and plant species software . as the proportion of species in each level of the network that changed module affiliation at least once between habitats. The mean number of modules and Multilayer network construction. We assembled a multilayer quantitative adjustability of the observed networks was compared against that of the networks 12,25 bipartite seed–dispersal network for each habitat, which were visualized using the R generated by the null models with a one-sample t test . package bipartite . Interaction frequency was quantified by calculating a pooled Multilayer modularity calculations and the interlayer null model were frequency of occurrence, collating the information from the different sources (fecal performed in MATLAB . analysis, mist-netting, camera traps, and direct observations). This pooled fre- quency of occurrence, resulted from the direct sum of all fecal samples containing Contribution of disperser species to seed dispersal cohesion. First, we assessed 5,33,70 at least one seed of a given plant species , the number of transects where a whether the specialization of seed dispersers differed consistently between habitats given focal interaction was detected, and the number of camera-trap recordings (1 by calculating animal specialization (d′), which quantifies their selectiveness for night = 1 sample), where a given interaction was detected. seeds within the range of resources used . However, the number of interactions of Using a multilayer formalism , we assembled a multilayer network formed by: a species is considered to reflect both resources availability and consumer activity. (a) a set of nodes (called “physical nodes”) representing the animal dispersers and This metric takes into account the pattern of interaction of a species in relation to the plant species whose seeds are dispersed, (b) a set of layers representing the the available resources, while being robust to sampling effort, network size and different habitats; (c) a set of “state nodes” that correspond to the manifestation of 74 75 asymmetry , also see ref. . We used a GLMM, with Gamma errors, and included each node on a given layer; and (d) a set of two types of edges connecting the nodes disperser species as a random factor. The model was analyzed with the R package pairwise, namely, intralayer edges (i.e., animal seed–dispersal interactions); and lme4 . If any level of the independent variable was significant, we assessed its interlayer edges, connecting state nodes between layers (i.e., animal or plant species 2 77 overall effect with a Wald Χ test available in R package car . The overall fit of the between habitats). Interlayer edges encode the movement of animal and plant model was assessed with the Akaike’s information criterion against a reduced species between habitats. While the temporal scale of the movements is not exactly model, which only included the intercept. Pairwise comparisons between habitats the same, both animal and plant genes frequently cross and establish in were tested with Tukey HSD test with the R package multicomp . neighboring habitats , and for that reason can be considered effective habitat Secondly, we explored seed–dispersers’ importance in the network by connectors with the strength of the interlayer links quantifying the intensity of the calculating their overall versatility as seed dispersers. Versatility identifies species habitat coupling provided by this movement. Spatial multilayer networks have a that are topologically important for the structure of the multilayer network . categorical (non-ordinal) coupling, i.e., interlayer edges are not constrained in any Versatility was calculated using software muxViz 1.0 , which adapts the Google’s specific order and any pair of layers can be connected . The quantification of the PageRank algorithm to describe the position of a node within the structure of a interconnectivity of the multilayer network is used in the calculation of some network based on a random walk between adjacent nodes (see Supplementary (modularity and multistrength), but not all the network diagnostic (versatility and Information for further details), being equivalent to a global measure of centrality. edge overlap). The implementation of this method requires bipartite networks to be projected To assess differences in the richness of animals, plants, and plant–animal onto unimodal networks. While some projection methods entail some loss of interactions among habitats, we used a G test. If an overall effect was present, we information , we applied a weighted projection which estimates interaction weight performed pairwise G tests to identify differences between habitat pairs. This based on the proportion of shared interactions (i.e., seed species shared by analysis was performed with the R package RVAideMemoire . 42,43 disperser species) relatively to the total network size , thus minimizing the loss of information. The projection was performed with function projecting_tm from the R package tnet . This algorithm is particularly suitable to multilayer Modular structure of the spatial multilayer network. To evaluate the extent to 29 40 networks as it condensates information on dispersers niche overlap , based on which the seed–dispersal interactions of Gorongosa are sorted into distinct com- the importance of their shared dispersed seeds . munities, we calculated multilayer modularity (Q ). Multilayer modularity, multilayer To understand how animals shared their links across the different habitats, we as its application to monolayer counterparts, quantifies to what extent nodes are calculated edge overlap with software muxViz 1.0 , which quantifies the organized into modules of strongly interacting nodes interacting more frequently proportion of common links between animals across habitats. than expected by chance . The Q modularity function was maximized applying a We evaluated the information condensed by multilayer versatility in respect to “generalized Louvain” method . This method proceeds until the network config- the species versatility of an aggregated network, and to other species-level metrics, uration that maximizes the weight of the edges inside the modules in relation to a calculating its correlation with aggregated species versatility, specialization d′, null model is found (see Supplementary Information for details). Following Pilosof et al. , the modularity function was adapted to reflect the multistrength, and the number of habitats in which a species is present. The specialization index d′ measures the distribution of a species interactions with each bipartite nature of our networks. The “generalized Louvain” method requires the specification of two parameters: the resolution limit γ, and the interlayer coupling partner over the total number of partners available . Multistrength is an extension of its monolayer version, and sums the total weight of the links incident on a node ω. The resolution limit defines the detail to which the network will be resolved into communities, and can be viewed as the importance given to the null model across all layers, taking into account the links connecting nodes in the different 38 24,38 layers (see Supplementary Information for details). It expresses the total number compared to the empirical data , and we used the default resolution γ = 1 . The of shared interactions of a dispersers species with all its neighboring species across interlayer coupling quantifies the strength of the connection between layers of a all habitats. Versatility and multistrength were calculated using muxViz 1.0 . network, i.e., the effect that species have in connecting the different layers. However, measuring such interlayer strength is intrinsically challenging and there is no absolute method to do so. To explore the importance of interlayer Code availability. MATLAB scripts for the estimation of modularity and gen- strength for the detection of modules, we calculated modularity along a range of erating the interlayer null model are available from https://figshare.com/s/ values of interlayer strength (from 0 to 10), assuming uniform interlayer strength cb5723b5adf640ec0451, and R code for generating the intralayer null model, across all species, i.e., all species connecting any two pair of habitats have the same analysis of the modular structure, and generating the files to be used with muxViz effect in the interlayer process . The stochastic nature of this algorithm means that are available from https://figshare.com/s/197888d27ec3d4a0b9d0. a different maximum is reached on each run, thus we run the modularity function 12,25 100 times, and averaged the results obtained . We compared the results Data availability. The seed–dispersal interaction network matrices are available on obtained with a multilayer network to that of two different representations of the reasonable request. same network: (a) an aggregated network (Q ), where all interactions across aggregated the different layers were pooled to create one overall aggregated network, with the frequency of interactions that occur in multiple habitats being summed, and (b) a Received: 18 October 2016 Accepted: 18 December 2017 disconnected network where habitats are considered fully independent from each other, i.e., interlayer strength is set to zero, and thus modularity is calculated for each of them . We test the modular structure of the observed network comparing it to the 12,25 structure of networks built under the assumptions of two null models : (1) an “intralayer null model” that keeps the number of links (i.e., connectance), while References redistributing the individual interactions across the whole matrix as implemented by r00_both in vegan , was used to assess the influence of structure within each 1. Thebault, E. & Fontaine, C. Stability of ecological communities and the layer/habitat in the modular structure of the multilayer network ; and (2) an architecture of mutualistic and trophic networks. Science 329, 853–856 (2010). “interlayer model” following the same rationale of the nodal model , i.e., keeping 2. Tylianakis, J. M., Laliberté, E., Nielsen, A. & Bascompte, J. Conservation of the same matrix, but redistributing species identities independently in each layer, to species interaction networks. Biol. Conserv. 143, 2270–2279 (2010). assess if the modular structure is dependent on the identity of the species 3. Jordano, P. Chasing ecological interactions. PLoS Biol. 14, e1002559 (2016). connecting the habitats. We run each null model 100 times along the same range of interlayer strength (0–10). The significance of the observed modularity was compared against the distribution of the modularity of the null networks, and NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications 9 | | | ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y 4. Timóteo, S., Ramos, J. A., Vaughan, I. P. & Memmott, J. High resilience of seed 38. Bazzi, M. et al. Community detection in temporal multilayer networks, with an dispersal webs highlighted by the experimental removal of the dominant application to correlation networks. Multiscale Model. Simul. 14,1–41 (2016). disperser. Curr. Biol. 26, 910–915 (2016). 39. Pilosof, S., Morand, S., Krasnov, B. R. & Nunn, C. L. Potential parasite 5. Heleno, R. H., Ramos, J. A. & Memmott, J. Integration of exotic seeds into an transmission in multi-host networks based on parasite sharing. PLoS ONE 10, Azorean seed dispersal network. Biol. Invasions 15, 1143–1154 (2013). 1–19 (2015). 6. Allen, C. R. et al. Quantifying spatial resilience. J. Appl. Ecol. 53, 625–635 40. Mello, M. A. R. et al. Keystone species in seed dispersal networks are mainly (2016). determined by dietary specialization. Oikos 124, 1031–1039 (2015). 7. Bascompte, J. & Jordano, P. Mutualistic Networks 206 (Princeton University 41. Yachi, S. & Loreau, M. Biodiversity and ecosystem productivity in a fluctuating Press, Princeton, NJ, 2014). environment: the insurance hypothesis. Proc. Natl Acad. Sci. USA 96, 8. Hagen, M. et al. Biodiversity, species interactions and ecological networks in a 1463–1468 (1999). fragmented world. Adv. Ecol. Res. 46,89–120 (2012). 42. Opsahl, T. Projecting Two-mode Networks Onto Weighted One-mode Networks 9. García-Callejas, D., Molowny-Horas, R. & Araújo, M. B. Multiple interactions https://toreopsahl.com/2009/05/01/projecting-two-mode-networks-onto- networks: towards more realistic descriptions of the web of life. Oikos 125, weighted-one-mode-networks/ (2009). 336–342 (2017). 43. Newman, M. E. J. Scientific collaboration networks. II. Shortest paths, weighted 10. Lundberg, J. & Moberg, F. Mobile link organisms and ecosystem functioning: networks, and centrality. Phys. Rev. E 64, 16132 (2001). implications for ecosystem resilience and management. Ecosystems 6, 44. Opsahl, T. Structure and Evolution of Weighted Networks 104–122 (University 0087–0098 (2003). of London, Queen Mary College, London, 2009). 11. Knight, T. M., McCoy, M. W., Chase, J. M., McCoy, K. A. & Holt, R. D. Trophic 45. Turner, M. G. Landscape ecology: the effect of pattern on process. Annu. Rev. cascades across ecosystems. Nature 437, 880–883 (2005). Ecol. Syst. 20, 171–197 (1989). 12. Pilosof, S., Porter, M. A., Pascual, M. & Kéfi, S. The multilayer nature of 46. Hampe, A. Plants on the move: the role of seed dispersal and initial population ecological networks. Nat. Ecol. Evol. 1, 101 (2017). establishment for climate-driven range expansions. Acta Oecol. 37, 666–673 13. Kivela, M. et al. Multilayer networks. J. Complex Netw. 2, 203–271 (2014). (2011). 14. Olesen, J. M. et al. From broadstone to Zackenberg: space, time and hierarchies 47. Bélisle, M. Measuring landscape connectivity: the challange of behavioral in ecological networks. Adv. Ecol. Res. 42,1–69 (2010). landscape ecology. Ecology 86, 1988–1995 (2005). 15. Frost, C. M. et al. Apparent competition drives community-wide parasitism 48. Bassett, D. S. et al. Robust detection of dynamic community structure in rates and changes in host abundance across ecosystem boundaries. Nat. networks. Chaos 23, 13142 (2013). Commun. 7, 12644 (2016). 49. Levine, J. & Murrell, D. The community-level consequences of seed dispersal 16. Poisot, T., Canard, E., Mouillot, D., Mouquet, N. & Gravel, D. The dissimilarity patterns. Annu. Rev. Ecol. Evol. Syst. 34, 549–574 (2003). of species interaction networks. Ecol. Lett. 15, 1353–1361 (2012). 50. Rodríguez-Echeverría, S. et al. Arbuscular mycorrhizal fungi communities from 17. Bodin, Ö. & Norberg, J. A network approach for analyzing spatially structured tropical Africa reveal strong ecological structure. New Phytol. 213, 380–390 populations in fragmented landscape. Landsc. Ecol. 22,31–44 (2007). (2017). 18. Trojelsgaard, K., Jordano, P., Carstensen, D. W. & Olesen, J. M. Geographical 51. Jordán, F., Liu, W. & Davis, A. J. Topological keystone species: measures of variation in mutualistic networks: similarity, turnover and partner fidelity. Proc. positional importance in food webs. Oikos 112, 535–546 (2006). R. Soc. B Biol. Sci. 282, 20142925–20142925 (2015). 52. Ray, J. C. & Sunquist, M. E. Trophic relations in a community of African 19. Boccaletti, S. et al. The structure and dynamics of multilayer networks. Phys. rainforest carnivores. Oecologia 127, 395–408 (2001). Rep. 544,1–122 (2014). 53. Slater, K. & du Toit, J. T. Seed dispersal by chacma baboons and syntopic 20. Guimarães, P. R., Pires, M. M., Marquitti, F. M. & Raimundo, R. L. eLS1–9 ungulates in southern African savannas. South Afr. J. Wildl. Res. 32,75–79 (Wiley, Hoboken, NJ, 2016). (2002). 21. Fontaine, C. et al. The ecological and evolutionary implications of merging 54. Campos-Arceiz, A. & Blake, S. Megagardeners of the forest - the role of different types of networks. Ecol. Lett. 14, 1170–1181 (2011). elephants in seed dispersal. Acta Oecol. 37, 542–553 (2011). 22. Guimerà, R. & Amaral, L. A. N. Functional cartography of complex metabolic 55. Bunney, K., Bond, W. J. & Henley, M. Seed dispersal kernel of the largest networks. Nature 433, 895–900 (2005). surviving megaherbivore-the African savanna elephant. Biotropica 49, 395–401 23. Donatti, C. I. et al. Analysis of a hyper-diverse seed dispersal network: (2017). modularity and underlying mechanisms. Ecol. Lett. 14, 773–781 (2011). 56. Schupp, E. W., Jordano, P. & Gómez, J. M. Seed dispersal effectiveness revisited: 24. Mucha, P. J., Richardson, T., Macon, K., Porter, M. A. & Onnela, J.-P. a conceptual review. New Phytol. 188, 333–353 (2010). Community structure in time-dependent, multiscale, and multiplex networks. 57. Guimarães, P. R., Galetti, M. & Jordano, P. Seed dispersal anachronisms: Science 328, 876–878 (2010). rethinking the fruits extinct megafauna ate. PLoS ONE 3, e1745 (2008). 25. Bassett, D. S. et al. Dynamic reconfiguration of human brain networks during 58. Howe, H. F. & Smallwood, J. Ecology of seed dispersal. Annu. Rev. Ecol. Syst. learning. Proc. Natl Acad. Sci. USA 108, 7641–7646 (2011). 13, 201–228 (1982). 26. Emer, C., Memmott, J., Vaughan, I. P., Montoya, D. & Tylianakis, J. M. Species 59. Escribando-Avila, G., Lara-Romero, C., Heleno, R. & Traveset, A. in Seed roles in plant–pollinator communities are conserved across native and alien Dispersal Networks in the Tropic (eds Datillo, W. & Rico-Gray, V.) (Springer, ranges. Divers. Distrib. 22, 841–852 (2016). New York). 27. Jordan, F. Keystone species and food webs. Philos. Trans. R. Soc. B Biol. Sci. 60. Bascompte, J. & Jordano, P. Mutualistic Networks: Monographs in Population 364, 1733–1741 (2009). Biology (Princeton University Press, Princeton, NJ, 2014). 28. Martín González, A. M., Dalsgaard, B. & Olesen, J. M. Centrality measures and 61. Jordano, P. Sampling networks of ecological interactions. Funct. Ecol. 30, the importance of generalist species in pollination networks. Ecol. Complex 7, 1883–1893 (2016). 36–43 (2010). 62. Wilson, E. O. & Naskrecki, P. A Window on Eternity: A Biologist’s Walk 29. De Domenico, M., Sole-Ribalta, A., Omodei, E., Gomez, S. & Arenas, A. Through Gorongosa National Park (Simon and Schuster, New York, 2014). Ranking in interconnected multilayer networks reveals versatile nodes. Nat. 63. Cillie, B. Pocket Guide to Mammals of Southern Africa (Sunbird Publishers, Commun. 6, 6868 (2015). Verlag, 2010). 30. McDonald-Madden, E. et al. Using food-web theory to conserve ecosystems. 64. Murray, K. D. Scatalog: Quick ID Guide to Southern African Animal Droppings Nat. Commun. 7, 10245 (2016). (Struik Nature, Cape Town, 2011). 31. Schleuning, M., Blüthgen, N., Flörchinger, M., Braun, J. & HM Specialization 65. Heleno, R. H., Ross, G., Everard, A., Memmott, J. & Ramos, J. A. The role of and interaction strength in a tropical plant-frugivore network differ among avian ‘seed predators’ as seed dispersers. Ibis 153, 199–203 (2011). forest strata. Ecology 92,26–36 (2011). 66. Chao, A. Estimating the population size for capture-recapture data with 32. Stalmans, M., Peel, M. & Massad, T. Aerial Wildlife Count of the Parque unequal catchability. Biometrics 43, 783 (1987). Nacional da Gorongosa. (2014). 67. Oksanen, J. et al. Vegan: community ecology package R package version 2. 4–1. 33. Correia, M., Timóteo, S., Rodríguez-Echeverría, S., Mazars-Simon, A. & https://CRAN.R-project.org/package=vegan (2013). Heleno, R. Refaunation and the reinstatement of the seed-dispersal function in 68. R Core Team. R: a language and environment for statistical computing (R Gorongosa National Park. Conserv. Biol. 31,76–85 (2017). Foundation for Statistical Computing, Vienna, 2015). 34. Matlack, G. R. Plant species migration in a mixed-history forest landscape in 69. Dormann, C. F., Frund, J., Bluthgen, N. & Gruber, B. Indices, graphs and null Eastern North America. Ecology 75, 1491–1502 (1994). models: analyzing bipartite ecological networks. Open Ecol. J. 2,7–24 (2009). 35. Hanski, I. Metapopulation dynamics. Nature 396,41–49 (1998). 70. Vázquez, D. P., Morris, W. F. & Jordano, P. Interaction frequency as a 36. Guimerà, R. & Amaral, L. A. N. Cartography of complex networks: modules surrogate for the total effect of animal mutualists on plants. Ecol. Lett. 8, and universal roles. J. Stat. Mech. Theory Exp. 2005, P02001 (2005). 1088–1094 (2005). 37. Jutlla, I. S., Jeub, L. G. & Mucha, P. J. A Generalized Louvain Method for 71. Shea, K. How the wood moves.pdf. Science 1231, 1231–1233 (2007). Community Detection Implemented in MATLAB http://netwiki.amath.unc.edu/ GenLouvain (2014). 10 NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-017-02658-y ARTICLE 72. Hervé, M. RVAideMemoire: diverse basic statistical and graphical functions R Author contributions Package version 0.9-68. https://CRAN.R-project.org/package=RVAideMemoire S.T., R.H., H.F., and S.R.-E. designed the study; S.T. and M.C. collected the data; S.T. led (2016). the analysis and wrote the first draft; and all authors contributed to manuscript revisions. 73. Bioinformatics toolbox: user’s guide R2016b. Version 9.1.0.441655 (The MathWorks Inc., Natick, Massachusetts, 2015). Additional information 74. Blüthgen, N., Menzel, F. & Blüthgen, N. Measuring specialization in species Supplementary Information accompanies this paper at https://doi.org/10.1038/s41467- interaction networks. BMC Ecol. 6, 9 (2006). 017-02658-y. 75. Poisot, T., Canard, E., Mouquet, N. & Hochberg, M. E. A comparative study of ecological specialization estimators. Methods Ecol. Evol. 3, 537–544 (2012). Competing interests: The authors declare no competing financial interests. 76. Bates, D., Mächler, M., Bolker, B. & Walker, S. Fitting linear mixed-effects models using lme4. J. Stat. Softw. 67,1–48 (2015). Reprints and permission information is available online at http://npg.nature.com/ 77. Fox, J. & Weisberg, S. An R Companion to Applied Regression (Sage reprintsandpermissions/ Publications Inc., Thousand Oaks, CA, 2011). 78. Hothorn, T., Bretz, F. & Westfall, P. Simultaneous inference in general Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in parametric models. Biom. J. 50, 346–363 (2008). published maps and institutional affiliations. 79. De Domenico, M., Porter, M. A. & Arenas, A. MuxViz: a tool for multilayer analysis and visualization of networks. J. Complex Netw. 3, 159–176 (2015). 80. Brin, S. & Page, L. The anatomy of a large-scale hypertextual web search engine. Comput. Netw. 56, 3825–3833 (2012). Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Acknowledgements Commons license, and indicate if changes were made. The images or other third party This work was supported by FEDER funds from the “Programa Operacional Factores de material in this article are included in the article’s Creative Commons license, unless Competitividade – COMPETE” and by national funds from the Portuguese Foundation for Science and Technology – FCT through the research project PTDC/BIA-BIC/4019/ indicated otherwise in a credit line to the material. If material is not included in the 2012. S.R.-E., R.H., and M.C. were also supported by FCT (Grants IF/00462/2013, IF/ article’s Creative Commons license and your intended use is not permitted by statutory 00441/2013, and SFRH/BD/96050/2013, respectively). We thank Greg Carr, and the Greg regulation or exceeds the permitted use, you will need to obtain permission directly from Carr Foundation – Gorongosa Restoration Project, Dr Marc Stalmans and all the staff the copyright holder. To view a copy of this license, visit http://creativecommons.org/ from Gorongosa National Park for their logistic assistance during fieldwork, and for licenses/by/4.0/. sharing their knowledge and passion about Gorongosa. © The Author(s) 2017 NATURE COMMUNICATIONS (2018) 9:140 DOI: 10.1038/s41467-017-02658-y www.nature.com/naturecommunications 11 | | |

Journal

Nature CommunicationsSpringer Journals

Published: Jan 10, 2018

There are no references for this article.