A synthesis of animal-mediated seed dispersal of palms reveals distinct biogeographic differences in species interactions

Disclaimer/Complaints regulations If you believe that digital publication of certain material infringes any of your rights or (privacy) interests, please let the Library know, stating your reasons. In case of a legitimate complaint, the Library will make the material inaccessible and/or remove it from the website. Please Ask the Library: https://uba.uva.nl/en/contact, or a letter to: Library of the University of Amsterdam, Secretariat, Singel 425, 1012 WP Amsterdam, The Netherlands. You will be contacted as soon as possible.


| INTRODUCTION
Quantifying knowledge gaps in digitally accessible information is a priority for advancing biodiversity science (Hortal et al., 2015;Meyer, Kreft, Guralnick, & Jetz, 2015). One of the key gaps is the lack of knowledge about interactions among species or groups of species (Hortal et al., 2015). Although interest in studying species interactions over broad spatial extents is increasing (Araújo & Rozenfeld, 2014;Schleuning et al., 2014;Trøjelsgaard & Olesen, 2013;Zanata et al., 2017), comprehensive datasets on ecological networks are still restricted to a few study sites, limiting many applications in biogeography and macroecology (Kissling & Schleuning, 2015;Poisot et al., 2016). Nevertheless, a considerable-yet underutilized-portion of our current knowledge on species interactions can be found within the published scientific literature (de Almeida & Mikich, 2018;Poelen, Simons, & Mungall, 2014). This information can be extracted from the text, but is usually not readily accessible in a digital format (Skusa, Rüegg, & Köhler, 2005;Thessen & Parr, 2014). Once extracted, such interaction data might provide an avenue for a deeper integration of macroecology and network research, for example by analysing the structure and functional composition of aggregated meta-networks (de Almeida & Mikich, 2018;Hagen et al., 2012;Kissling & Schleuning, 2015).
One of the key mutualistic interactions among plants and animals is frugivory, that is when animals, especially vertebrates, consume fleshy fruits and subsequently disperse the seeds (Fleming, Breitwisch, & Whitesides, 1987;Fleming & Kress, 2013;Kissling, Böhning-Gaese, & Jetz, 2009). Animal-mediated seed dispersal is particularly ubiquitous in the tropics where up to 90% of the plant species in a particular community might rely on fruit-eating vertebrates for seed dispersal (Fleming et al., 1987). Recent research on fleshy-fruited plants and frugivores has revealed that interactions among these mutualistic partners are often constrained by functional traits (Bender et al., 2018;Dehling, Jordano, Schaefer, Böhning-Gaese, & Schleuning, 2016). For example, fruit size limits the fruit ingestion by relatively small-sized seed dispersers, and the size of the ingested fruits therefore tends to be positively correlated with body sizes and gape widths of consumers (Jordano, 2000;Lord, 2004;Onstein et al., 2017). Nevertheless, there is considerable variation in functional traits among biogeographical regions (Fleming & Kress, 2013;Mack, 1993) and it remains an open question whether functional trait matching in plantfrugivore interactions can be consistently observed across regions (Bender et al., 2018;Onstein et al., 2017).
With >2,500 species, the palms (Arecaceae) are a major plant family and characteristic of tropical and subtropical regions across the world (Dransfield et al., 2008). Due to decades of research, palms have received a comprehensive taxonomic scrutiny (Govaerts & Dransfield, 2005) and many aspects of their ecology, biogeography and evolution have been widely studied (Baker & Couvreur, 2013;Eiserhardt, Svenning, Kissling, & Balslev, 2011;Henderson, 2002;Kissling et al., 2012). Moreover, palms are a keystone resource for frugivorous animals in the tropics because they provide a large amount and wide variety of fruits to animal consumers (Fleming & Kress, 2013;Henderson, 2002;Onstein et al., 2017;Zona & Henderson, 1989). Seed dispersers of palms include many frugivorous animals, including birds, bats, non-flying mammals, reptiles, insects, and even fishes (Zona & Henderson, 1989). Zona and Henderson (1989) provided the most comprehensive review of animal seed dispersers of palms, and despite updates of this work (http://www.virtualherbarium.org/palms/psdispe rsal.html), we still know little about how palm-frugivore interactions differ among biogeographical regions and where knowledge gaps are most pronounced. Key questions are to what extent the strong differences in palm species richness between the Afrotropics (e.g., mainland Africa with 65 species, and Madagascar with 175 species) and the Neotropics (>700 species)  as well as the regional differences in the taxonomic and functional trait composition of avian and mammalian frugivores (Kissling et al., 2009(Kissling et al., , 2014 are reflected in macroecological patterns of palm-frugivore interactions (Kissling, 2017). Moreover, identifying knowledge gaps could help to prioritize targeted efforts for new and more intensive data collection.
Here, we synthesize published knowledge on frugivore seed dispersal of palms in the Afrotropics and the Neotropics and quantify biogeographical differences in palm-frugivore interactions between both regions. We focus on those two biogeographical regions because they differ strongly in both species richness and traits of palms and frugivores (e.g., Kissling, 2017;Kissling et al., 2009Kissling et al., , 2012Onstein et al., 2017;Sandom et al., 2013) and because biogeographers have a great interest in explaining the diversity anomalies among these regions (Fleming & Kress, 2013;Fleming et al., 1987;Mack, 1993;Richards, 1973). Using interactions recorded in the primary literature, we aggregate information on species interactions to quantify the geographic, taxonomic, and functional variation in palmfrugivore interactions. Specifically, we explore (a) to what extent this information is incomplete and unevenly distributed in geographic space (Hortal et al., 2015;Meyer et al., 2015), (b) whether taxonomic and biogeographical differences in palms and frugivores drive MUÑOZ ET AL. | 467 differences in the structure and composition of the aggregated meta-networks (Kissling et al., 2009(Kissling et al., , 2014, and (c) whether functional trait relationships between palms and frugivores are consistent across both biogeographical regions. Furthermore, we also aim to enhance the digital availability of species interaction data (Poelen et al., 2014) by increasing accessibility, interoperability, and reusability (Wilkinson et al., 2016).

| Data compilation
We first extracted interaction data from the literature (Figure 1, green boxes). Candidate articles were selected from a comprehensive literature search in Thomson's Web of Science (WoS) in January 2016. We used the following combinations of English search terms: "seed disper" and/or "africa" and/or "palm" and/or "southameric" and/or "neotropic" and/or "afrotropic." Since the WoS has not all titles and full-text available in English, we repeated the same search by replacing the terms "seed disper" and "southameric" with the corresponding translations in Portuguese, Spanish, and French (i.e., languages widely used in the study regions). For instance, we included the search terms "dispersao," "sementes," "dispersion," "semillas," and "graines" as translations of "seed dispersal." From this literature search, we compiled an initial list of 2,232 articles ( Figure 1). Each reference was labelled with a random alphanumerical code using the ZOTERO software (www.zotero.org).
Since full text scans (by manual reading) are very time consuming, we screened the abstracts and titles of all articles from this initial list and pre-selected those articles for further consideration that mentioned potential information on palm-frugivore interactions and/or seed dispersal in the title or abstract. We did this pre-selection screening using the abstract_screener() function from the R package 'metagear' (Lajeunesse, 2015). This R package provides a graphical user interface that allows visualizing abstracts and titles and creating a database with references without seeing the names of journals and authors (thereby avoiding potential publication selection biases based on author name, journal, or year of study). The pre-selection screening resulted in 692 references for a manual full text scan (Figure 1). This resulted in extracting pairwise seed dispersal interactions between palms and frugivores from 162 articles ( Figure 1).
F I G U R E 1 Workflow to extract and analyse pairwise palm-frugivore interactions from the literature. Left (green boxes): extraction of interaction data from articles through literature search, title and abstract screening (with R package 'Metagear'), full text scans and crossreferences to other articles. Increasing arrow sizes reflect larger number of articles. Right (blue boxes): aggregation of pairwise interaction data into meta-networks and analyses of trait matching using additional data on frugivore body size and palm fruit size [Colour figure can be viewed at wileyonlinelibrary.com] Appendix S1 in Supporting Information lists the data sources (except those that are already cited in the main text).
For the data extraction, we recorded an interaction when an article mentioned the fruit or the seed of a palm being dispersed, carried or defaecated by a frugivorous animal. Hence, we did not only focus on endozoochorous seed dispersal but also on seed dispersers that do not swallow the seeds, for example bats and scatter-hoarding rodents that usually carry large fruits. We further aimed to only record effective seed dispersal interactions, that is avoiding interactions that reflect seed predation. In most cases, a clear difference between seed disperser and seed predator was made in the examined articles because seed predators typically destroy the seeds during consumption. However, for some taxa (e.g., parrots) the relative importance of seed predation versus seed dispersal remains debated (Tella et al., 2015). We also included secondary dispersers such as scatter-hoarding rodents which take the fruits from the ground.
Interactions were only recorded if the specific palm species occurred in the Neotropics or Afrotropics, not in other biogeographic regions.
Most observations of pairwise species interactions could be extracted from the article text or from tables. Interactions were not included if species-level information was unavailable for either the palm or the frugivore (e.g., general statements such as "hornbills and primates consume fruits of palm A" were not included).
The dataset obtained from our literature search included the palm-frugivore seed dispersal data from Zona and Henderson (1989).
We additionally included the latest update of this dataset (from July 2006) which is only available online (http://www.virtualherbarium. org/palms/psdispersal.html). From all examined articles, we further extracted basic meta-data information, including the year of publication, journal name, and location where the interaction was observed.
To standardize the taxonomic names of palms, we followed the World Checklist of palms (Govaerts & Dransfield, 2005), using an updated version (downloaded July 2015). For birds, we followed the BirdLife Taxonomic Checklist v8 (BirdLife International, 2015). For mammals, we followed the taxonomy from the IUCN Red List (IUCN et al., 2008). All other species names of animal seed dispersers were standardized according to the Integrated Taxonomic Information System (ITIS, https://www.itis.gov/). Besides species names, we also extracted information on the taxonomic class and order for each palm and frugivore using the tax_name() function from the R package 'taxize' 0.7.4. (Chamberlain & Szöcs, 2013).
In the second part of the workflow (Figure 1, blue boxes), we aggregated all pairwise interaction data extracted from the 162 articles and combined it with location and trait data of palms and frugivores. To extract information on functional traits (see below), we used external data sources (i.e., not from the articles from which interaction data were extracted). We focused on measures of fruit size for palms and body size for frugivores (a proxy of gape width) because size matching is the most commonly reported morphological trait matching relationship in plant-frugivore interactions (Bender et al., 2018;Burns & Lake, 2009;Dehling et al., 2016;Donatti et al., 2011;Eklöf et al., 2013;Jordano, 2000;Lord, 2004;Mack, 1993; Onstein et al., 2017).

| Quantification of knowledge gaps
We quantified knowledge gaps in digitally accessible palm-frugivore interactions by (a) estimating sampling completeness, and (b) assessing geographic coverage of interaction data.
Sampling completeness was estimated for each biogeographical region (Neotropics, Afrotropics) as the ratio between the total observed frugivore richness and the expected asymptotic value as derived from a species richness estimator (i.e., the expected richness of frugivorous dispersers for all palm species) (Rivera-Hutinel, Bustamante, Marín, & Medel, 2012). Given the heterogeneous nature of our sampling, we selected the nonparametric incidence-based Chao estimator because it is among the best estimators for low sample sizes and it accounts for frequencies of rare species (Chao, Colwell, Lin, & Gotelli, 2009). The Chao estimator was calculated with the specpool() function from the R package 'vegan' 2.3.5 (Oksanen et al., 2007) which assumes a lognormal distribution of the variance to estimate the 95% confidence intervals around the expected asymptotic value. We further calculated accumulation curves with 100 random permutations of the sampling units (i.e., palm species) using the function specaccum() (with "method = random") from the R package 'vegan' 2.3.5 (Oksanen et al., 2007). We used palm species as the sampling units because they were the focal taxa for which interactions were extracted from the literature. Our implementations generally followed the approach suggested by Gotelli and Colwell (2001) and Rivera-Hutinel et al. (2012).
We also estimated sampling completeness of frugivore interaction information for each individual palm species. To do this, we quantified accumulation curves for each individual palm species by randomizing the unique records of frugivores using articles with interaction information as sampling units. Individual palm sampling completeness was estimated as the ratio between the observed palm degree and the expected degree (i.e., the asymptotic value of frugivores calculated with Chao). Sampling completeness was only quantified for palm species which interacted with at least two different frugivore species and for which interaction information was available from at least two articles. Frugivore records from reviews such as the one from Zona and Henderson (1989) were excluded because it was not possible to relate every single interaction record to the original article source. Subsequently, we investigated (with Spearman rank correlations and single predictor regressions) whether the number of articles in which a palm has been recorded was a good predictor of (a) palm degree (i.e., the total number of frugivores the palm interacted with), and (b) sampling completeness of individual palms.
This was done to quantify potential publication bias, that is to test whether the amount of knowledge on interaction partners is related to the number of published articles.
To assess the geographic coverage of interaction data, we used location data from the articles (i.e., where the interaction was observed) and standardized them by assigning the location information to geographic units as defined by the Taxonomic Databases Working Group (TDWG) (Brummitt, 2001). We used the TDWG level 3 units ("botanical countries,") which mostly represent MUÑOZ ET AL.

| 469
countries, but some of the very large countries such as Brazil are subdivided into states or provinces. This is the finest spatial resolution at which global palm distribution data are currently available . We calculated four metrics to assess geographic coverage because each of them gives different insights into the coverage of interaction data. First, we calculated the ratio of palm species with interaction data relative to the total number of palm species recorded in each botanical country (based on the World Checklist of palms, see above). This indicated how many palm species within a botanical country have at least some information on interaction partners (i.e., a minimum of one recorded interaction).
Second, we calculated for each botanical country the total number of unique interactions (i.e., the number of pairwise seed dispersal interactions between a particular palm species and a particular frugivore species). This indicated how many unique interactions have been recorded in total for a particular botanical country. Third, we calculated for each botanical country the mean number of interactions recorded per palm species using all palm species present in a botanical country (i.e., including also palms without interaction data).
Fourth, we calculated the mean sampling completeness of individual palm species for each botanical country. Here, we also used all palm species present in a botanical country (based on the World Checklist of palms) and assigned a sampling completeness of zero to palm species for which sampling completeness could not be estimated. This metric provides an estimate of how well, on average, palms are sampled (in terms of frugivore interactions) within a botanical country.

| Comparison of networks
The pairwise interaction data extracted from the literature were aggregated into two meta-networks: one for the Neotropics and one for the Afrotropics. Both meta-networks were binary two-mode matrices with P palm species in rows and F frugivores in columns. Each element (P i, F j ) with species i and j in the binary adjacency matrices had a value of 1 if frugivore F i had been observed to interact with palm species P j , and zero otherwise. To compare the diversity and organization of species interactions within the two regional meta-networks, we calculated a number of commonly used network-level indices. The most simple indices included the number of palm and frugivore species, their ratio, the size of the interaction matrices (P×F), the total number of species (P+F), the total number of interactions (I), connectance (i.e., the ratio between realized and potential interactions given as C=I/(P×F)), and the mean number of interactions per species (Table 1). These indices describe the size and diversity of the networks in terms of species and interactions (Trøjelsgaard & Olesen, 2013). We further calculated nestedness and modularity, two of the most commonly used network-level indices that describe the organization of species interactions within the entire network (Bascompte, Jordano, Melian, & Olesen, 2003;Fortuna et al., 2010;Olesen, Bascompte, Dupont, & Jordano, 2007).
Nestedness measures to what extent interactions of specialist species are proper subsets of the interactions of generalist species (Almeida-Neto, Guimarães, Guimarães, Loyola, & Ulrich, 2008;Bascompte et al., 2003). We calculated nestedness with the NODF metric proposed by Almeida-Neto et al. (2008) which is included in the bipartite package (Dormann, Fründ, Blüthgen, & Gruber, 2009) for R. The NODF metric ranges from 0 (not nested) to 1 (perfectly nested). To compare nestedness between networks, we standardized the empirical NODF values as Z-scores: where NODF null is the average NODF-value of 999 random matrices using a null model that re-ordered interactions while maintaining species richness and palm degree (i.e., the number of interactions).
Hence, in this null model the probability of allocating an individual palm-frugivore interaction is ultimately dependent on the total number of interactions a specific frugivore has, with generalist frugivores having a higher probability of being selected than specialists. This null model corresponds to the 'r1' null model from the R package 'vegan' (Oksanen et al., 2007). The SD null value represents the standard deviation. The NODF Z-score therefore measures the difference between empirical and random nestedness in numbers of standard deviations. Values at least 1.96 standard deviations away from the mean are considered to be statistically significant.
Modularity quantifies the degree to which a network is structured into independent or semi-independent subgroups (modules) of species within the network (Donatti et al., 2011;Olesen et al., 2007). We evaluated differences in modularity between networks by measuring modularity Q for each meta-network as defined by Barber (2007): where m is the total number of interactions when considering an adjacency matrix A of P×F dimensions where interactions are only allowed between F i and P j . H ij is a null model matrix that describes the probability of occurrence of an interaction between i and j based on the degree distributions of A ij . The δ refers to the Kronecker's delta function and equals 1 when individual species P i and F j are sorted into the same module. Species i and j are assigned to a community group or module, denoted by g i and h j . Modularity Q then measures the extent to which interactions are formed within modules instead of between modules (Barber, 2007;Olesen et al., 2007).
Q ranges from 0 to 1, where higher values represent a stronger division of a network into modules of closely interacting species (Donatti et al., 2011;Olesen et al., 2007).
We used a Label Propagation Algorithm (LPAwb+) to calculate Q (Beckett, 2016). LPAwb+ works well for calculating modularity in binary bipartite networks and it can be integrated with the R software (Beckett, 2016). Since the LPAwb+ algorithm is sensitive to node label initialization, we analysed each network 100 times and

| Functional trait matching
To test for functional trait matching, we explored the relationships between palm fruit size and frugivore body size using generalized linear models. As a measure of fruit size, we extracted fruit length of palms (in cm) using available information from books (e.g., Henderson, 2002) and other sources, including monographs and species descriptions (e.g., references and data sources listed in appendix of Göldel, Kissling, & Svenning, 2015) as well as the palmweb database from the Royal Botanical Gardens, Kew (http://palmweb.org). For each mammalian and avian frugivore, we extracted average body mass (in g) for birds from Dunning (2008) and for mammals from Sandom et al.
(2013). Since we did not have morphological measurements of gape width (which usually have to be obtained from museum specimens or individuals captured in the field), we used body mass as a proxy for gape width to describe the size matching relationship between frugivores and palms (Donatti et al., 2011;Mack, 1993;Onstein et al., 2017). Hence, we used palm fruit length (log-transformed) as the predictor variable and the median value of body masses of all frugivores eating, dispersing, carrying, or defaecating a particular palm species as response variable (log-transformed). We performed analyses for all frugivore species (mammals, reptiles, and birds) as well as separate analyses for the most commonly recorded disperser group (bird and mammal, respectively). There were no palms being mainly dispersed by reptiles. We further did these analyses separately for the Neotropics and the Afrotropics. Because palm species can be dispersed by frugivores from different animal classes, we assigned a particular palm species as being dispersed by either birds or mammals depending on the most commonly recorded taxonomic class of frugivore dispersers.

| Data compilation
Our literature search resulted in an initial list of 2,232 articles. After screening of titles and abstracts, a total of 692 articles were selected for a full, manual text scan. From these articles, pairwise seed dispersal interaction data for palms and frugivores were found in 162 articles.
T A B L E 1 Network-level indices that describe the diversity and organization of Neotropical and Afrotropical palm-frugivore networks. For the Neotropics, indices for both the full meta-network as well the mean (±SD) from subsampled networks (n = 100) are provided. The Neotropical network was subsampled to have the same number of palm species as the Afrotropical network Average indices for the subsampled Neotropical networks which contained the same number of palm species (n = 29) as the full Afrotropical meta-network. The subsampling was repeated 100 times and the mean (±SD) of network indices is provided (see Methods for details).
c Standardized modularity and standardized nestedness were calculated as Z-scores against a null model calculating modularity (Q) and nestedness (NODF) from a set of 999 random matrices (see Methods for details).
The assembled species interaction dataset comprised a total of 1,122 interactions, with 750 being unique pairwise interaction records. A total of 340 frugivore species and 126 palm species were involved. The vast majority of interactions was recorded from the Neotropics (1,008 in total, containing 660 unique interactions), involving a total of 98 palm species and 283 frugivore species. The Afrotropics had considerably less interaction data (114 in total, of which 90 were unique), and only involved 29 palm species and 57 frugivore species.

| Quantification of knowledge gaps
Accumulation curves of interactions for both biogeographical regions did not approach asymptotes ( Figure 2a). This suggested that knowledge on palm-frugivore interactions in both regions is highly incomplete. Sampling completeness (i.e., the percentage of the total number of frugivores that are estimated to disperse all palm species in a region) was 47% for the Neotropics and 40% for the Afrotropics (Figure 2a).
At the level of individual palm species, estimated sampling completeness ranged from 14-100%, with an average of 57 ± 21% (±SD) across all palm species for which asymptotic values of the frugivore accumulation curves could be estimated (n = 49 species; Table S1).
Hence, accumulation curves for individual palm species rarely approached asymptotes (Appendix S2 Figure S1), even not for palm species such as Euterpe edulis and Syagrus romanzoffiana that had the largest number of interaction partners recorded (i.e., 56 and 38 frugivores, respectively, see Table S1)

| Comparison of networks
The taxonomic composition of the two regional meta-networks differed strongly (Figure 4). The Afrotropical meta-network contained fewer palms and frugivores than the Neotropical one, and was taxonomically also less diverse because only birds and mammals appeared as seed dispersers of palms ( Figure 4). Seven Afrotropical palm tribes were included in the network, and Areceae was the most species-rich tribe (10 palm species). The palm species Elaeis guinensis and Phoenix reclinata had the highest number of interactions as they were recorded to interact with 15 and 14 frugivores, respectively. Afrotropical seed dispersers of palms included parrots, hornbills, elephants, baboons, guenons, mangabeys and chimpanzees, with no single frugivore species standing out as dispersing a particular high number of palm species. In Madagascar, lemurs were the most important palm seed dispersers, especially for species in the palm genus Dypsis.
Most network-level indices showed considerable differences in the diversity and organization of species interactions between the two regions, even when the Neotropical network was subsampled to an equal number of palm species (Table 1). Specifically, the Neotropical meta-network included more frugivore species, more pairwise interactions, and a larger interaction matrix than the Afrotropics, and this remained true when subsampling the Neotropical networks (Table 1). After subsampling, connectance (i.e., the proportion of all potential interactions that are realized) was similar in the Neotropics and the Afrotropics (Table 1). Both regions also showed significantly modular networks (Z-scores > 1.96; Table 1). After subsampling, the mean standardized modularity (Q Z-score ) remained slightly larger for the Neotropics than the Afrotropics, although it had a large variance (

| Functional trait matching
Overall, frugivores with larger body sizes tended to disperse palm species with larger fruits ( Figure 5), but the relationship was only statistically significant in the Afrotropics (Table S2). In both regions, mammalian frugivores tended to disperse medium to large palm fruits whereas birds tended to disperse small to medium-sized fruits.
However, considering birds and mammals separately, positive functional trait matching relationships between palm fruit size and frugivore body size were only statistically significant in the Afrotropics ( Figure 5, Table S2). This suggests that biogeographical differences in traits of interacting species can determine the broad-scale structure of ecological networks.

| DISCUSSION
By aggregating pairwise interaction data from the scientific literature, we compiled a comprehensive palm-frugivore seed dispersal dataset for the Neotropics and the Afrotropics. Even when accounting for differences in palm species richness, the available data revealed considerable differences in palm-frugivore interactions between the two biogeographical regions, with the Neotropics having a broader taxonomic range, larger interaction diversity and stronger modularity than the Afrotropics. Functional trait analyses between palm fruit size and frugivore body size revealed that large-scale trait-matching relationships are only observed in the Afrotropics, indicating that larger palm fruits are dispersed by larger frugivores (both birds and mammals). Nevertheless, major knowledge gaps in interaction diversity remain in tropical regions where palms and vertebrate frugivores are particular diverse (Kissling et al., 2009(Kissling et al., , 2014.
Our analyses suggest that the amount of knowledge of observed palm-frugivore interactions is not merely driven by publication bias.
Sampling completeness of individual palm species did not correlate with the total number of articles in which they were present. Hence, the most well-studied palms (which also have many interaction partners) tend to be the most abundant and widespread ones, and the wide geographic distribution allows them to interact with a larger diversity of frugivores. In both regions, the majority of the seed dispersers are mammals and birds, but the Neotropical meta-network also included reptiles, a land crab, one beetle, and ten fish species.

Neotropics Afrotropics
F I G U R E 2 Sampling completeness of palm-frugivore interaction data in the Neotropics and Afrotropics. (a) Rarefied species accumulation curves obtained from randomizing palm species and estimating the asymptote value with the Chao species richness estimator. The stippled lines represent the variance around the expected asymptote. The vertical lines represent estimates of the total number of palm species actually recorded in each biogeographical region (x-axis) versus the estimated asymptote and its associated variance for the total number of frugivore species dispersing palm species in each region (y-axis). (b) Relationships between estimated sampling completeness of palm-frugivore interactions and the number of articles in which individual palm species have been recorded. Only palm species for which interaction information is available from at least two articles and two unique interactions are included (n = 49 species). Sampling completeness was quantified as the ratio between the expected number of frugivores and the observed number of frugivores for each palm species (see Appendix S2 Figure S1 for accumulation curves of individual palm species). The symbol size represents the decadic logarithm of the number of frugivores reported for each palm species [Colour figure can be viewed at wileyonlinelibrary.com] Fishes are important seed dispersers of Neotropical plants, including palms (Astrocaryum, Bactris, Mauritiella, and Iriartella) . In the Afrotropics, fishes have been recorded as fruit-eaters in the Congo basin for other plants (Beaune et al., 2013), suggesting that palm-fish interactions could in principle occur in this region.
The most widely recorded mammalian seed dispersers of palms in the Neotropics were tapirs, peccaries, primates, scatter-hoarding rodents, and bats, whereas in the Afrotropics mostly primates and bats were recorded (including lemurs in Madagascar). The birds were mostly non-passerine species (~80%), with toucans, cracids, and parrots being commonly recorded in the Neotropics and hornbills and turacos in the Afrotropics. At the meta-network level, an important group of generalist species in the Neotropics were rodents, which were present in approximately one-third of the modules in this region. In the Afrotropics, primates were generalist species and represented in about half of all modules. In both regions, passerine birds tended to be specialists, interacting only with one or few palm species. Parrots were the most generalist bird group, being represented in about one-fifth of the modules. While it is generally assumed that parrots are seed predators and thus do not participate in seed dispersal mutualisms (Fleming & Kress, 2013), there is increasing evidence that many parrots do indeed disperse the seeds of plants, including palms (Tella et al., 2015). This indicates that parrots can serve as efficient long-distance seed dispersers, although their net effect on the population dynamics of their food plants will rely on the negative impact of seed predation versus the benefits derived from long-distance seed dispersal (Tella et al., 2015).
Network indices can be affected by network size (Rivera-Hutinel et al., 2012). For instance, connectance correlates negatively with network size and our subsampling showed that Neotropical networks had a similar connectance to the Afrotropical meta-network when subsampled to the same number of palm species. However, for modularity the standardized Q Z-score showed a statistically significant modular structure for all networks, that is the Neotropical meta-network, the subsampled Neotropical networks, and the Afrotropical meta-network. This suggests that seed dispersal  We found a positive relationship between palm fruit size and frugivore body size in the Afrotropics, suggesting morphological trait matching between these mutualistic partners (Bender et al., 2018).
This relationship was not statistically significant in the Neotropics. This contrasts with results from a locally sampled, hyperdiverse seed dispersal network in the Neotropics where the relationship between frugivore body size and plant fruit size is driven by differences in body size between birds and mammals (i.e., birds predominantly disperse small fruits whereas mammals predominantly disperse large fruits) (Donatti et al., 2011). The lack of a broad-scale, Neotropical sizematching relationship in our study could be driven, at least partly, by the late Quaternary loss of mammalian megafauna, which was particularly pronounced in the Neotropics (Janzen & Martin, 1982;Pires, Guimarães, Galetti, & Jordano, 2018;Svenning et al., 2016). This could have resulted in an increased extinction rate of palms with megafaunal fruits (Onstein et al., 2018). After these Pleistocene megafauna extinctions, Neotropical scatter-hoarding rodents seem to have substituted to some degree the extinct megafauna seed dispersers of large-seeded trees such as palms (Guimarães, Galetti, & Jordano, 2008; Jansen et al.,

All species
Log[ fruit length (cm) ] Log[ median body mass (g) ] F I G U R E 5 Relationships between frugivore body size (y-axis) and palm fruit size (x-axis) for the Neotropics (upper row) and the Afrotropics (lower row). The relationships are shown for all palm species dispersed by mammals, reptiles, and birds (left), palm species only dispersed by birds (middle), and palm species only dispersed by mammals (right). Fruit size of each palm species is represented by log-transformed fruit length (cm). Frugivore body size is represented by log-transformed median body mass (g) of all frugivore species that have been recorded to disperse a particular palm species. The median body mass values are given in colours (purple and orange) whereas grey points represent species-level variation of frugivore body mass. Lines represent statistically significant relationships between median frugivore body size and palm fruit size (see Table S2 for results from linear regressions) [Colour figure can be viewed at wileyonlinelibrary.com] 2012; Pires et al., 2018). In the Afrotropics, a positive relationship between palm fruit size and frugivore body size was retained when analysing birds and mammals separately, which corresponds to the few megafauna extinctions in this regions (Svenning et al., 2016). The morphological (size) matching relationship between frugivorous birds and fleshy-fruited plants may be caused by avian body mass being correlated with avian gape size which imposes a strong selective pressure on fruit ingestability (Burns & Lake, 2009;Jordano, 1995;Lord, 2004).
Studies of plant-frugivore networks across the Neotropical Andes further show that trait matching between birds and fleshy-fruited plants can also be found for other trait combinations such as plant crop mass and avian body mass (Bender et al., 2018). However, only a few palm species occur at high elevations and data on crop masses of palms are rare, making it difficult to specifically test this trait matching relationship for palms.
Our study exemplifies the Eltonian shortfall, that is, the incomplete and limited knowledge of species interactions we currently have (Hortal et al., 2015). Despite aggregating 750 unique pairwise interaction records for 340 frugivores and 126 palms, extensive knowledge gaps remain in terms of taxonomic and geographic coverage of palm-frugivore interactions (Figure 3). The low coverage and sampling completeness of available data further provide an incentive to increase our baseline knowledge of palm-frugivore seed dispersal interactions, for example, through targeted field work as well as further data mining.
For instance, additional interaction data could be retrieved by (a) expanding empirical field data through mobilizing citizen science projects (Poelen et al., 2014); (b) utilizing more advanced text mining tools that can find biological information (e.g., on species interactions) through searches of taxon names in machine-readable texts (Thessen & Parr, 2014); and (c) broaden the literature search by querying not only the WoS, but also other databases (e.g., Google Scholar, Scopus) or unpublished information (e.g., theses that are only available in local university repositories, interaction data from dedicated homepages). A bottleneck of our current data extraction framework (Figure 1) is that the title and abstract screening (via the graphical user interface of the R package 'metagear') as well as the manual full text scans of articles are costly and labour-intensive. Hence, a promising next step could be to develop text mining tools that allow automated content analysis to extract ecologically relevant information from high volumes of literature (Nunez-Mir, Iannone, Pijanowski, Kong, & Fei, 2016). However, in the context of species interactions this requires not only to automatically identify entities in the text (i.e., species names), but also how they are related (e.g., species A disperses species B) (Skusa et al., 2005;Thessen & Parr, 2014). To our knowledge, user-friendly software that allows such text mining of species interaction data is not yet available.
As a contribution towards increasing the digital accessibility of species interaction data, following the FAIR (findability, accessibility, interoperability, and reusability) principles (Wilkinson et al., 2016), we make our palm-frugivore dataset not only available in the DRYAD digital repository (https://doi.org/10.5061/dryad.rd46vq3), but also through the Global Biotic Interactions (GloBI) infrastructure (see Data accessibility). This allows to standardize the extracted interaction data and to integrate them with available taxonomies, ontologies, and vocabularies (Poelen et al., 2014). For instance, the GloBI infrastructure applies the Open Biomedical Ontologies (OBO) Relations Ontology (Smith et al., 2005) which uses defined interaction terms to describe how species interact with each other (e.g., the Brazilian Tapir Tapirus terrestris "eats" the palm Oenocarpus mapora, the palm Attalea humilis is "eatenBy" the collared peccary Pecari tajacu). This standardizes not only the terms for recording species interaction data, but also allows to link and cross-reference these interaction records with several taxonomies and name services, with other ontologies that describe environments and habitats, and with standardized information on locations and geographic names (Poelen et al., 2014).

| CONCLUSION S
Our synthesis of palm-frugivore interactions reveals distinct biogeographical differences in animal-mediated seed dispersal of palms. In the Afrotropics, the lower interaction diversity relative to the Neotropics parallels the low taxonomic diversity previously reported for the Afrotropical flora and fauna (Fleming & Kress, 2013;Fleming et al., 1987;Kissling et al., 2009Kissling et al., , 2012Richards, 1973). Network indices such as modularity and nestedness suggest a non-random arrangement of seed dispersal meta-networks in both biogeographical regions, with stronger modularity in the Neotropics than the Afrotropics. We hypothesize that the lack of broad-scale morphological trait matching among Neotropical palms and avian and mammalian frugivores, respectively, could be partly driven by the extinction of megafauna Janzen & Martin, 1982) and a subsequent replacement of primary seed dispersal through secondary seed dispersal by scatter-hoarding rodents (Jansen et al., 2012). We suggest that deeper insights into the biogeography of species interactions could be obtained through additional data mining and targeted field work. We therefore urge ecologists and biogeographers to collect and aggregate more species interaction data, and to make them findable, accessible, and reusable through open-data repositories and interoperable web services.

DATA ACCESSIBILI TY
All palm-frugivore interaction data extracted for this study are avail-