Study site
The study was conducted in the Brasília National Park (PNB), Federal District, Brazil (15º39′57″ S; 47º59′38″ W), a 42.355 ha Protected Area with a typical vegetation configuration found in the Cerrado of the central highlands of Brazil, i.e., a mosaic of gallery forest patches along rivers surrounded by a matrix of savannas and grasslands34. The climate in the region falls into the Aw category in the Köppen scale, categorizing a tropical wet savanna, with marked rainy (October to March) and dry (April to September) seasons.
We carried out the study in eight fixed sampling sites scattered evenly throughout the PNB and separated by at least two kilometers from one another (Supplementary Fig. S1). The sites consisted of four cerrado sensu stricto sites (bushy savanna containing low stature trees); two gallery forest edges sites (ca. 5 m from forest edges, containing a transitional community), and two gallery forest interior sites. These three types reflect the overall availability of habitat types in the reserve (excluding grasslands) and are the most appropriate foraging areas to sample interactions as bat-visited plants are either bushes, trees, or epiphytes, but rarely herbs35.
Bat and interaction samplings
We sampled bat-plant interactions using pollen loads collected from bat individuals captured in the course of one phenological year, thus configuring an animal-centered sampling. We carried out monthly field campaigns to capture bats from October 2019 to February 2020, from August to September 2020, and from March to July 2021. In each month, we carried out eight sampling nights during periods of low moonlight intensity, each associated with one of the eight sites. Each night, we set 10 mist nets (2.6 × 12 m, polyester, denier 75/2, 36 mm mesh size, Avinet NET-PTX, Japan) at ground level randomly within the site, which were opened at sunset and closed after six hours. We accumulated a total sampling effort of 552 net-hours, 28,704 m2 of net area, or 172,224 m2h sensu Straube and Bianconi36.
All captured bats were sampled for pollen, irrespective of family or feeding guild. We used glycerinated and stained gelatin cubes to collect pollen grains from the external body of bats (head, torso, wings, and uropatagium). Samples were stored individually, and care was taken not to cross-contaminate samples. Pollen types were identified by light microscopy, and palynomorphs were identified to the lowest-possible taxonomical level using an extensive personal reference pollen collection from plants from the PNB (details in next section). Palynomorphs were sometimes classified to the genus or family level or grouped in entities representing more than one species. Any palynomorph numbering five or fewer grains in one sample was considered contamination, alongside any anemophilous species irrespective of pollen number.
Bats were identified using a specialized key37 and four ecomorphological variables were measured for each individual. (i) Forearm length and (ii) body mass were used to calculate the body condition index (BCI), a proxy of body robustness38, where higher BCI values indicate larger and heavier bats, which are less effective in interacting with flowers in general due to a lack of hovering behavior, the incapability of interacting with delicate flowers that cannot sustain them, a lower maneuverability and higher energetic requirements39. Moreover, we measured (iii) longest skull length (distance from the edge of the occipital region to the anterior edge of the lower lip) and (iv) rostrum length (distance from the anterior edge of the eye to the anterior edge of the lower lip) to calculate the rostrum-skull ratio (RSR), a proxy of morphological specialization to nectar consumption23. Higher RSR values indicate bats with proportionally longer rostra in relation to total skull length. Longer rostra in bats are associated with a weaker bite force and thus less effective in consuming harder food items such as fruits and insects, thus suggesting a higher adaptation to towards nectar40,41. Bats were then tagged with aluminum bands for individualization and released afterward. To evaluate the sampling completeness of the bat community and of the pollen types found on bats, we employed the Chao1 asymptotic species richness estimator and an individual-based sampling effort to estimate and plot rarefaction curves, calculating sampling completeness according to Chacoff et al.42.
All methods were carried out in accordance with relevant guidelines and regulations. The permits to capture, handle and collect bats were granted by the Ethical Council for the Usage of Animals (CEUA) of the University of Brasília (permit 23106.119660/2019-07) and the Instituto Chico Mendes de Conservação da Biodiversidade (ICMBio) (permit: SISBIO 70268). Vouchers of each species, when the collection was possible, were deposited in the Mammal Collection of the University of Brasília.
Assessment of the plant community
In each of the eight sampling sites, we delimited a 1000 × 10 m transect, each of which was walked monthly for one phenological year (January and February 2020, August to December 2020, and March to July 2021) to build a floristic inventory of plants of interest and to estimate their monthly abundance of flowering individuals. Plant species of interest were any potential partner for bats, which included species already known to be pollinated by bats, presenting chiropterophilous traits sensu Faegri and Van Der Pijl43, or any plant that could be accessed by and reward bats, whose flowers passes all the three following criteria:
(i) Nectar or pollen is presented as the primary reward to visitors. (ii) Corolla diameter of 1 cm or more. This criterion excludes small generalist and insect-pollinated flowers where the visitation by bats is mechanically unlikely. It applies to the corolla diameter in non-tubular flowers or the diameter of the tube opening. Exceptions were small and actinomorphic flowers aggregated in one larger pollination unit (pseudanthia) where the 1 cm threshold was applied to inflorescence diameter. (iii) Reward must be promptly available for bats. This criterion excludes species with selective morphological mechanisms, such as quill-shaped bee-pollinated flowers or flowers with long and narrow calcars.
All flowering individuals of interest species found in the transects were registered. A variable number of flowers/inflorescences (n = 5–18) were collected per species for morphometric analysis. For each species, we calculated floral tube length (FTL), corresponding to the distance between the base of the corolla, calyx, or hypanthium (depending on the species) to its opening, and the corolla’s outermost diameter (COD), which corresponds to the diameter of the corolla opening (tubular flowers) or simply the corolla diameter (non-tubular flowers). For pseudanthia-forming species, inflorescence width was measured. Pseudanthia and non-tubular flowers received a dummy FTL value of 0.1 mm to represent low restriction and enable later calculations. Finally, we collected reference pollen samples from all species from anthers of open flowers, which were used to identify pollen types found on bats. For plant species found in pollen loads but not in the PNB, measures were taken from plants found either on the outskirts of the site (Inga spp.) or from dried material in an online database (Ceiba pentandra, in https://specieslink.net/) using the ImageJ software44. Vouchers were deposited in the Herbarium of the Botany Department, University of Brasília.
Data analysis
Network macrostructure
We built a weighted adjacency matrix i x j, where cells corresponded to the number of individuals of bat species i that interacted with plant species or morphotype j. All edges corresponding to legitimate interactions were included. With this matrix, we calculated three structural metrics to describe the network’s macrostructure. First, weighted modularity (Qw), calculated by the DIRTLPAwb + algorithm45. A modular network comprises subgroups of species in which interactions are stronger and more frequent than species out of these subgroups10, which may reveal functional groups in the network9. Qw varies from zero to one, the latter representing a perfectly modular network.
Second, complementary specialization through the H2′ metric46. It quantifies how unique, on average, are the interactions made by species in the network, considering interaction weights and correcting for network size. It varies from zero to one, the latter corresponding to a specialized network where interactions perfectly complement each other because species do not share partners.
Lastly, nestedness, using the weighted WNODA metric25. Nested networks are characterized by interaction asymmetries, where peripheral species are only a subset of the pool of species with which generalists interact47. The index was normalized to vary from zero to one, with one representing a perfectly nested network. Given that the network has a modular structure, we also tested for a compound topology, i.e., the existence of distinct network patterns within network modules, by calculating intra-module WNODA and between-module WNODA36. Internally nested modules appear in networks in which consumers specialize in groups of dissimilar or clustered resources and suggest the existence of distinct functional groups of consumers25,48. Metric significance (Qw, H2′, and WNODA) was assessed using a Monte Carlo procedure based on a null model. We used the vaznull model3, where random matrices are created by preserving the connectance of the observed matrix but allowing marginal totals to vary. One thousand matrices were generated and metrics were calculated for each of them. Metric significance (p) corresponded to the number of times the null model delivered a value equal to or higher than the observed metric, divided by the number of matrices. The significance threshold was considered p ≤ 0.05.
Given a modular structure, we followed the framework of Phillips et al.49 that correlates network concepts (especially modularity) with the distribution of morphological variables of pollinators to unveil patterns of niche divergence in pollination networks. Given the most parsimonious module configuration suggested by the algorithm, we compared modules in terms of the distribution of morphological variables of the bat (RCR and BCI) and plant (FTL and COD) species that composed the module. Differences between modules means were tested with one-way ANOVAs.
Drivers of network microstructure
The role of different ecological variables in determining pairwise interaction frequencies was assessed using a probability matrices approach3. This framework considers that an interaction matrix Y is a product of several probability matrices of the same size as Y, with each matrix representing the probability of species interacting based on an ecological mechanism. Thus, adapting it to our objectives, we have Eq. (1):
$$mathrm{Y}=mathrm{f}(mathrm{A},mathrm{ M },mathrm{P},mathrm{ S})$$
(1)
where Y is the observed interaction matrix, and a function of interaction probability matrices based on species relative abundances (A), representing neutrality as species interact by chance; species morphological specialization (M), phenological overlap (P), and spatial overlap (S). We built models containing each of these matrices in the following ways:
Relative abundance (A): matrix cells were the products of the relative abundances of bat and plant species. The relative abundances of bats were determined through capture frequencies (each species’ capture frequency divided by all captures, excluding recaptures) and the relative abundances of plants were determined by the number of flowering individuals recorded in transections (each species’ summed abundance in all transects and all months divided by the pooled abundance of all species in the network). Cell values were normalized to sum one.
Morphological specialization (M): cells were the probability of species interacting based on their matching degree of morphological specialization. Morphologically specialized bats (i.e., longer rostra and smaller size) are more likely to interact with morphologically specialized flowers (i.e., longer tubes and narrower corollas), while unspecialized bats are more likely to interact with unspecialized, accessible flowers. For this purpose, we calculated a bat specialization index (BSI) as the ratio between RCR and BCI, where higher BSI values indicate overall lower body robustness and longer snout length. Likewise, the flower specialization index (FSI) was calculated for plants as the ratio between FTL and COD, where higher values indicate smaller, narrower, long-tubed flowers that require specialized morphology and behavior from bats for visitation. BSI and FTL were normalized to range between zero and one and were averaged between individuals of each species of bat or plant. Therefore, interaction probabilities were calculated as in Eq. (2):
$${P}_{i,j}=1-|{BSI}_{i}-{FSI}_{j}|$$
(2)
where Pi,j is the interaction probability between bat species i and plant species j and |BSIi – FSIj| is the absolute difference between bat and plant specialization indexes. Similar index values (two morphologically specialized or unspecialized species interacting) lead to a low difference in specialization and thus to a high probability of interaction (Pi,j → 1), whereas the interaction between a morphologically specialized and a morphologically unspecialized species leads to a high absolute difference and thus lower probability of interaction (Pi,j → 0). Cell values of the resulting matrix were normalized to sum one.
Phenological overlap (P): cells were the probability of species interacting based on temporal synchrony, calculated as the number of months that individuals of bat species i and flowering individuals of plant species j co-occurred in the research site, pooling all capture sites/transections. Cell values were normalized to sum one.
Spatial overlap (S): cells were the probability of species interacting based on their co-occurrence over small-scale distances and vegetation types, calculated as the number of individuals from a bat species i captured in sampling sites where the plant species j was registered in the transection, considering all capture months. Cell values were normalized to sum one.
Because more than one ecological mechanism may simultaneously drive interactions3,9, we built an additional set of seven models resultant from the element-wise multiplication of individual probability matrices:
SP: The spatial and temporal distribution of species work simultaneously in driving a resource turnover in the community, driving interactions.
AS: Abundance drives interactions between bats and plants, but within spatially clustered resources in the landscape caused by a turnover in species distributions.
AP: Abundance drives interactions between bats and plants, but within temporally clustered resources caused by a seasonal distribution of resources.
APS: Abundance drives interactions between bats and plants, but within resource clusters that emerge by a simultaneous temporal and spatial aggregation.
MS: Similar to AS, but morphology drives interactions within spatial clusters.
MP: Similar to MP, but morphology drives interactions within temporal clusters.
MPS: Similar to APS, but morphology drives interactions within spatiotemporal clusters.
Finally, we created a benchmark null model in which all cells in the matrix had the same probability value. All the compound matrices and the null model were also normalized to sum one.
To compare the fit of these probability models with the real data, we conducted a maximum likelihood analysis3,9. We calculated the likelihood of each of these models in predicting the observed interaction matrix, assuming a multinomial distribution for the probability of interaction between species12. To compare model fit, we calculated the Akaike Information Criterion (AIC) for each model and their variation in AIC (ΔAIC) in relation to the best-fitting model. The number of species used in the probability matrices was considered the number of model parameters to penalize model complexity. Intending to assess whether nectarivorous bats and non-nectarivorous bats assembly sub-networks with different assembly rules, we created two partial networks from the observed matrix. One contained nectarivores only (subfamilies Glossophaginae and Lonchophyllinae) and their interactions, and the other contained frugivore and insectivore bats and their interactions. We repeated the likelihood procedure for these two partial networks.
To conduct the likelihood analysis, we excluded plant species from the network that could not have their interaction probabilities measured, such as species found in pollen samples but not registered in the park or pollen types that could not be identified to the species level. Therefore, the interaction network Y and probability matrices did not include these species (details in Supplementary Table S1).
Software
Analyses were performed in R 3.6.050. Network metrics and null models were generated with the bipartite package51, and the sampling completeness analysis was performed with the vegan package52. Gephi 0.9.253 was used to draw the graph.
Source: Ecology - nature.com
