Feeding habits influence species habitat associations at the landscape scale in a diverse clade of Neotropical fishes

A primary goal of community ecology is to understand the mechanisms that drive species' spatial distribution and habitat associations. Species' geographic distribution can be influenced by the distribution of their prey partly because consumers' behaviour is oriented to optimal energy use during foraging. We analysed how differences in dietary preferences influence the spatial distribution and habitat associations of species at the landscape scale. We hypothesized that differences in feeding guilds will lead to divergent habitat association patterns among species.


| INTRODUC TI ON
Elucidating the mechanisms that drive species' spatial distribution and habitat associations is a primary goal of community ecology (Heino et al., 2015;Jackson et al., 2001;King et al., 2021;Mittelbach & McGill, 2019). Strong relationships between species' geographic distribution and the distribution of their prey demonstrate the influence of food availability in shaping the habitat association of consumers (Doublet et al., 2019;Johnson & Sherry, 2001;Tableau et al., 2016). Consumer behaviours that seek to optimize energy use relative to foraging could explain part of this influence on spatial patterns (Tableau et al., 2016). According to optimal foraging theory, to enhance fitness, animals favour foraging strategies that provide the most benefit for the least cost, thus maximizing the net energy gained by individuals (Perry & Pianka, 1997). This implies that animal diets will adapt to fluctuations in food resource accessibility. For instance, in the Amazon, the diet breadth of frugivore fish species changes according to seasonal variation in food availability, consuming a higher amount of fruits during the flooding season (Correa & Winemiller, 2014). Similarly, mammals like northern Australian quolls exhibit plasticity in diet according to variations in the landscape and habitat around them (Dunlop et al., 2017).
Food resource use, one of the axes of the multidimensional niche of a species, plays a fundamental role in the relationship between species and their use of the environment (Pianka, 2000). Divergent use of food resources usually leads to niche partitioning among species, allowing them to coexist in a shared niche space; however, niche overlap can also occur along other niche axes (i.e. space and time; Chesson, 2000;Kraft et al., 2015;Mittelbach & Schemske, 2015;Pianka, 1973Pianka, , 2000. Environmental characteristics associated with resource availability, such as landscape heterogeneity, can significantly influence the partitioning of resources among species for different stages in a species' life cycle (Pérez-Crespo et al., 2013).
Although different definitions have been used to talk about environmental heterogeneity, referring to measures of diversity and structure of the environment (Ben-Hur & Kadmon, 2020;Stein et al., 2014), in the context of this study, we refer to landscape heterogeneity as areas containing several dissimilar habitat types or land cover types. Areas with higher levels of landscape heterogeneity are expected to contain more species and individuals than more homogeneous areas of the same habitat type (Turner & Gardner, 2015).
Variability in environmental conditions, like topography and microclimate, increases with patch size and offers more opportunities for organisms with different preferences and tolerances to find optimal conditions within the patch (Turner & Garner, 2015). Landscape variability, for instance, has been shown to be important in determining the distribution, abundance and diversity of several mammal, bird and fish species (Arantes et al., 2019;Lee & Martin, 2017;Thornton et al., 2011). For instance, differences in avian diversity were found when comparing landscapes dominated by agriculture versus non-crop vegetation cover, showing that species richness was lower when there were more agricultural fields in a landscape (Lee & Martin, 2017).
The Neotropical fish family Serrasalmidae (pacus and piranhas) offers an ideal model for studying the influence of feeding specialization on species habitat associations, given that dietary composition and level of specialization vary considerably within clades. This family is composed of c. 100 species and has developed trophically specialized clades ranging from frugivory to piscivory (Correa et al., 2007). Rheophilic species (e.g. Ossubtus xinguense, Tometes ancylorhynchus and Mylesinus paucisquamatus) specialize in periphytic bryophytes and vascularized plants (e.g. Podostemaceae) that grow on rocks in rapids (Andrade et al., 2019;Vitorino et al., 2016).
and Pygocentrus spp. are piscivorous and feed mainly from biting off pieces of flesh from other fishes (Goulding, 1980;Nico & Taphorn, 1988).
The diversity of feeding habits within Serrasalmidae is related to morphological traits for food acquisition (Huby et al., 2019). The species that feed on fruits and seeds have molariform teeth, while species that graze on leaves or stems have high-crowned incisiformlike teeth (Huie et al., 2020). In contrast, predators like piranhas that feed on other fishes have sharp, multicuspid, blade-like teeth. Lastly, scale-feeders have specialized stouter, conical or spatulate dentitions (Correa et al., 2007;Goulding, 1980;Kolmann et al., 2018).
Other characteristics that differentiate clades within this family are bite force and digestive tract length. Carnivorous species deliver more forceful bites than their herbivorous counterparts (Huby et al., 2019). Herbivorous species have longer guts than carnivorous and omnivorous relatives to cope with less digestible plant material (Pelster et al., 2015). Independent of family, planktivorous fishes have long and thin gill rakers for filter feeding so they are considered a specialized group (Burton & Burton, 2017). The morphological characteristics of the alimentary tract are thus helpful to infer trophic guild and level of specialization of fish taxa.
The geographic distribution of the Serrasalmidae family is restricted to tropical South America, mainly in the lowland (i.e. <500 m above sea level) areas of the Amazon drainage basin (Dagosta & de Pinna, 2017;Jézéquel et al., 2020). Amazon lowlands are characterized by a mosaic of habitats composed of evergreen forests, periodically flooded forests (e.g. várzea or igapó), savannas, lakes and extensive floodplains (Junk et al., 2010). The flooding dynamic is an essential ecological driver in floodplain ecosystems. According to the Flood Pulse Concept (FPC), the flood pulse influences the presence and distribution of organisms, determines life-history traits, affects primary and secondary production and influences decomposition and nutrient recycling (Junk & Wantzen, 2004). This is particularly important for fish species because the flood pulse enhances lateral connectivity between the river channel and floodplain, thereby enlarging aquatic habitats and allowing access to different feeding resources not available in the main channel (Junk, 1997).
The high landscape heterogeneity present in the Amazon lowlands and the role of the flood pulse in the expansion and connectivity of aquatic habitats are related to the distribution and availability of food resources for fish. They, therefore, are expected to influence fish abundance and distribution. For example, frugivorous fish species feed mainly on fruits from the flooded forest (Correa et al., 2007). The access and permanence of frugivorous fishes in the flooded forest depend on the flood pulse and the flood duration (Correa et al., 2015). The extent and diversity of floodplain habitats also are flood-pulse dependent. Wider floodplains will likely include more diverse habitats such as oxbow lakes. In these lentic systems, low water velocity enhances sedimentation and thereby present higher sunlight penetration. Such conditions support plankton growth (Bogotá-Gregory et al., 2020) which provides food for planktivorous species (e.g. Metynnis) and juvenile frugivores (e.g. Colossoma; Oliveira et al., 2006). Areas with a greater flooding magnitude would likely inundate more extensive floodplains and thereby increase food resource availability for flooded forest-dependent species, like frugivores.
Here, we used geospatially explicit fish occurrences from the most comprehensive dataset available (the AmazonFish project database, Jézéquel et al., 2020), diet data and satellite-derived landscape variables to explore how differences in feeding guilds influence the habitat association of serrasalmid species in the Amazon basin. We tested the hypothesis that serrasalmid species specializing in different food resources will use different habitat types. If serrasalmids optimize their foraging, as predicted by theory (Perry & Pianka, 1997), then the most specialized species should have distributions that closely track the habitats harbouring their preferred food resource. In the case of frugivores, for instance, their primary food source is restricted to floodplain forests (Correa & Winemiller, 2014;Goulding, 1980) and thus are expected to inhabit areas with a broad floodplain extent and long inundation time but low habitat heterogeneity (mostly forest). Likewise, planktivores follow the distribution of plankton, which in Amazonia is restricted to floodplain lakes because slow-moving and less turbid waters promote plankton production Forsberg et al., 2017) and thus are expected to inhabit areas with a broad floodplain extent and long inundation time but low habitat heterogeneity (mostly open-water floodplain lakes). In contrast, herbivores feed on a wide range of plant material (leaves, stems, flowers) and the distributions of those food types are spatially broad (Arantes et al., 2019;Silva et al., 2021). Thus, herbivores are expected to inhabit areas with high floodplain extent and landscape heterogeneity irrespective of flood duration. Lastly, for piscivores and species feeding on flesh, scales or fins of other fishes, their food base is broadly distributed (Martelo et al., 2008;Siqueira-Souza et al., 2016). Thus, these species are expected to inhabit areas with high landscape heterogeneity irrespective of floodplain extent and flood duration given that consumers can track prey movement into the floodplain for short or long periods of time ( Table 1). . Basin-wide, about three-quarters of these wetlands, are covered by forest, woodland or shrubland . The flooding regime for the Amazon River and its tributaries presents a monomodal flood pulse related to dry and rainy seasons (Junk & Wantzen, 2004). Therefore, these area estimates can change with the flooding regime; for instance, the flooded portion of the wetland area can vary from 34% to 75% from the dry to the flood season .

| Habitat association proxies
We used occurrence data, from the AmazonFish project database , covering all the major sub-basins within the Amazon drainage basin such as Xingú, Tapajós, Madeira, Purus, Juruá, Uacayali, Putumayo, Japurá, and Negro ( Figure 1). This database contains the most complete information currently available on freshwater fish species distribution for the Amazon drainage basin.
For the Serrasalmidae family, there are 14,269 occurrences of 81 species available, representing 83% of the valid species in the family.
We, however, selected a subset of species based on the availability of diet data and phylogenetic information, yielding a final number of 61 species with 13,667 occurrences (i.e. 62.24% of the valid species in the family; Table S1). We assumed that all the occurrences corresponded to adults with complete ontogeny.
Considering the relevance of scale on habitat association studies (Jackson et al., 2001), we used a multiscale approach meaning that our response variables are landscape characteristics measured at different extents. We analysed three habitat association proxies (i.e. floodplain extent, landscape heterogeneity and flood duration) with three different buffer sizes around each species occurrence: (1) small scale: 5 km for floodplain extent and 300 m for landscape heterogeneity and flood duration; (2) intermediate scale: F I G U R E 1 Spatial distribution of occurrence data in the Amazon River basin for the fish family Serrasalmidae retrieved from the Amazon fish project  C. macropomum, Piaractus brachypomus and Mylossoma duriventre; Duponchelle et al., 2021).

| Floodplain extent
To

| Landscape heterogeneity
The habitat types in the Amazon drainage basin were obtained from the satellite-derived product LBA-ECO LC-07 Wetland Extent, Vegetation and Inundation: Lowland Amazon Basin . (SAR). Shrub represents woody vegetation that is partitioned between tree and shrub using a height limit of 5 m. Woodland represents areas with 20%-70% of tree canopy cover. Forest represents areas with tree canopy cover >70% . To maintain standardized buffer areas, we considered all pixels falling within the buffer areas.
We calculated the number of occurrences per dominant land cover type (land cover type with higher proportion) within three buffer distances (300 m, 5 km and 25 km). A diversity index for the landscape was calculated within the three buffers, considering the proportional abundance of each land cover type, based on the Shannon's diversity index: where Pi is the proportion of class i. The values of this index range from SHDI = 0 when the landscape contains only one land cover type (i.e. no diversity) to SHDI >0, without limit; SHDI increases as the number of different land cover types (i.e. land cover richness) increases and/or the proportional distribution of area among land cover types becomes more equitable (McGarigal et al., 2012).

| Flood duration
We calculated the average flood duration (in number of months) in buffers of 300 m, 5 km and 25 km around occurrence points per species. We used the GIS product Surface WAter Fraction High Resolution (SWAF-HR) for 2021 (Parrens et al., 2019) which was the only publicly available dataset. This database contains monthly inundation areas at a high spatial resolution (1 km). Average flood duration represents the mean of the number of months each 1 km pixel, within each buffer, was flooded in 2012.

| Dietary classification
We used diet data from analyses of stomach contents to group species according to feeding habits. Diet data were compiled through a literature review (Kolmann et al., 2021)  composition. Because our dataset contained qualitative and quantitative data, we used a quartiles system to assess the specialization degree per species. For instance, if a species had >75% of fish remains in the reported stomach contents, the species was classified as a highly specialized piscivore and a species with <25% of any food item was classified as a low specialized species (Table S1). In addition, because fish frugivory is a unique feeding habit and fruit has a restricted distribution (fruits mainly occur in forested areas; Correa & Winemiller, 2014), we classified all species according to their level of frugivory, considering the presence and proportion of fruits and seeds in the stomach contents, using the rule described above (Table S1).
For species where different references used the same method to estimate diet composition, the records were averaged, and then, the quartile rule was applied. In cases where different methods were used, we could not calculate an average but considered the highest value from each reference. Based on food items consumed (Table S2), species were grouped into five functional feeding guilds: frugivores, herbivores, planktivores, piscivores and fin & scale feeders (Table S3). Note that insects were not reported in serrasalmid diets.

| Data analysis
First, we used principal component analysis (PCA) to visually explore spatial associations among species and land cover types based on their feeding guilds. We created buffers of 300 m, 5 km and 25 km, per occurrence record, to characterize the surrounding landscape. We conducted a covariance-based PCA of dominant land cover types within buffers (i.e. land cover type with the highest proportion) around each occurrence record and classified species according to their feeding guilds. The percentage of variance explained was used to inform the number of axes retained for interpretation (Jackson, 1993). PCAs were implemented in R 4.1.1 (R Core Team, 2021) and visualized using the package 'ggbiplot' (Vu, 2011).
Next, we used phylogenetic generalized least squares (PGLS) regressions to test for relationships between each habitat association response variables (i.e. floodplain extent, landscape heterogeneity and flood duration) and feeding guild predictor variables (i.e. frugivore, herbivore, planktivore, piscivore, fin and scale feeder) in a phylogenetic context. Given that closely related species are assumed to have more similar traits because of their shared ancestry, they are expected to produce more similar residuals from the least square's regression line than distantly related species. PGLS accounts for the interspecific autocorrelation due to phylogenic relatedness (Garamszegi, 2014;Martins & Hansen, 1997). PGLS models were implemented in R (R Core Team, 2021) using the function 'gls' from the package 'nmle' (Pinheiro et al., 2021) with the maximum-likelihood transformation of branch length optimized for the data ('method = ML'). We used a phylogenetic tree that includes 61 species (62.24% of the species in the family; Table S1). The tree was generated by increasing the taxon sampling of the most recent comprehensive phylogeny for serrasalmids (i.e. containing 36 species for which we have diet and occurrence data; Kolmann et al., 2021) with representatives from all recognized serrasalmid genera (See Appendix S1 for method details, Appendix S2 for accessions used for the legacy markers, Appendix S3 for the phylogeny). Previous to model implementation, we tested the assumption of homogeneity of variances which was met by all except one of the models (Landscape heterogeneity at 5 km). Given the closeness of the significance value to 0.05 alpha level, we decided not to change the model or transform the data to make all models comparable (

| RE SULTS
Habitat association proxies were correlated. The mean Pearson's R values for the pairwise correlation between the habitat association proxies are as follows: floodplain extent/landscape heterogeneity = 0.66; floodplain extent/flood duration = 0.37; and flood duration/landscape heterogeneity = 0.52.

| Land cover types
Irrespective of the buffer size (i.e. 300 m, 5 km or 25 km) around the occurrence record, ordination analyses failed to detect clear associations among feeding guilds and the dominant land cover type within buffers ( Figure 2). The degree of feeding guild overlap was buffer size dependent, from complete overlap at 300 m to partial overlap at the largest scales ( Figure 2).

| Floodplain extent
Our PGLS models revealed that feeding guild influences habitat association ( Table 2). Pairwise comparisons showed that frugivorous, F I G U R E 2 Principal components analysis (PCA) representing feeding guilds of the fish family Serrasalmidae in each dominant land cover type at buffers of (a) 300 m, (b) 5 km and (c) 25 km in the Amazon River basin. Concentric ellipses represent a 95% confidence ellipse interval. Arrows represent the main land cover types within each buffer. Abbreviated labels on arrows correspond to land covers: AMP, aquatic macrophyte; EP, elevation above 500 masl; FFP, flooded forest; FWP, flooded woodland; NFFP, no flooded forest; NWP, no wetland within Amazon Basin; OWP, open water. Numbers next to abbreviated labels represent the buffer size 3=300 m, 5=5 km, 25=25 km piscivorous and fin and scale feeders species use habitats associated with a broader mean floodplain extent than herbivores and plantktivores at all spatial scales studied (Table 3; Figure S1). However, frugivorous species' associations with floodplain extent were held even when there were small floodplain areas inside the buffers (represented by the first quartile values) (Table S5). Unexpectedly, piscivores and frugivores are not different in terms of floodplain mean extent (Table 3). Based on the Pseudo R 2 , feeding guilds explained more mean floodplain extent used at all buffers' radius than landscape heterogeneity and flood duration (Table S6). When considering the frugivory level of all species, our analyses failed to detect habitat association patterns related to floodplain extent (Table S7).

| Landscape heterogeneity
At all scales, some feeding guilds inhabit areas with greater landscape heterogeneity than others (based on the Shannon diversity index for buffers 5 and 25 km, Tables 2 and 3, Figure S1). At small buffer scales, frugivores are different from plantktivores, piscivores and fin and scale feeders by inhabiting areas with greater landscape heterogeneity (Table 3, Figure S1). Frugivores, piscivores and fin and scale feeders are associated with areas with greater landscape heterogeneity than herbivores and planktivores at the largest scale (25 km, Table 2). With the increase in buffer scale, the number of feeding guild pairs that exhibit significant differences also increased (Table 3). This trend is unexpected as one expects that the higher the scale, the higher the homogenization, so the feeding guilds should be more similar in their habitat associations.
In terms of frugivory level, mid-low frugivores were related to lower levels of landscape heterogeneity at the 25 km buffer size (Table S8).

| Flood duration
Only at the smallest scale (300 m), some feeding guilds are influenced by flood duration in the areas they inhabit ( Table 2). Specifically, fin and scale feeders, piscivores and frugivores are associated with areas of longer flood duration than planktivores (Table 3, Figure S1).

| DISCUSS ION
Our results indicated that the distribution of preferred food resources on habitats influences habitat association by consumers (i.e. feeding guilds) at the landscape scale in the Amazon River basin.
This pattern was mainly supported when considering the variation in floodplain extent and landscape heterogeneity that characterizes large rivers and their floodplains. Moreover, our results TA B L E 2 Results of PGLS for the response variables mean floodplain extent at the scales of 5 km, 10 km and 25 km and landscape heterogeneity (Shannon diversity index) and flood duration at the scales of 300 m, 5 km and 25 km Taken together, our results suggest that feeding guilds do influence species habitat associations. Our predictions, however, were only partially supported. We expected that herbivores would be associated with high floodplain extent and landscape heteroge-  (Correa et al., 2022).
Frugivores and piscivores appear to be similar in terms of how much floodplain extent they occupy but are different when we consider the landscape heterogeneity, with frugivores occupying areas with higher levels of heterogeneity at the smallest buffer size (300 m).
For frugivores, the observed association with a wider floodplain TA B L E 3 Pairwise comparison results for three habitat association proxy variables at three spatial scales extent is supported by the fact that they feed primarily on fruits from the flooded forest and serve as seed dispersers for numerous plant species (Correa et al., 2007(Correa et al., , 2015  . Moreover, land cover types such as wetlands and shrubs greatly influence fish assemblages diversity in Amazonian floodplain lakes . However, the relationship between land cover and fish assemblage structure seems to be scale-dependent, with stronger landscape effects at larger spatial scales Lobón-Cerviá et al., 2015). Our results show that habitat associations per feeding guilds are scale-dependent (i.e. buffer size) in landscape heterogeneity and flood duration, but not floodplain extent. We observed a trend of more feeding guilds occupying areas with similar heterogeneity at the small buffer size, but divergent heterogeneity at intermediate and large buffer sizes. This could be explained because a small buffer size includes fewer pixels, so the landscape heterogeneity is lower than the intermediate buffer size (5 km) and large buffer size (25 km), where more habitat types are likely included. This also implies that species feeding on different resources require different levels of landscape heterogeneity.
Our study provides insights on patterns of species habitat associations at a large scale. This is possible because we used data for the entire Amazon River basin and focused on perhaps one of the most trophically diverse fish families within the Amazon region. Previous studies relating landscape variables and patterns of freshwater fish diversity are mainly localized to central Amazonia (Arantes et al., , 2019Castello et al., 2018;Lobón-Cerviá et al., 2015).
An important caveat is that the microhabitat variability that influences species habitat use, and associations is not detectable with the land cover types and occurrence data that we used due to limitations on spatial resolution. However, our study provides a valuable first step towards assessing general relationships among species that use similar resources (i.e. feeding guilds) and their surrounding environment at the scale of the entire Amazon basin. High-resolution land cover data are available for Brazil (Mapbiomas https://mapbi omas. org) (Souza et al., 2020); however, the rest of the Amazon region lacks this type of information, making it challenging to conduct highresolution regional assessments and preventing us from providing a more detailed regional assessment.
Understanding species habitat associations by fish, through food resource dynamics and floodplain dependence, is pivotal to assessing the impact of anthropogenic activities, such as water regulation projects, pollution and climate change, on the processes that affect ecological patterns. Species in feeding guilds that are associated with wider extensions of floodplain and that depend on allochthonous food resources provided by floodplain habitats (i.e. frugivores) are very likely to be affected by the modification of the flow regime due to hydropower damming of large rivers (Arantes et al., 2019;Correa et al., 2022;de Bem et al., 2021) and climate change (Herrera-R et al., 2020). That is due to the disruption of the connectivity between the river and floodplains, which impedes the lateral exchange of nutrients and organisms. The homogenization caused by the land-use change in the Amazon (Souza et al., 2020) can affect species that require different levels of landscape heterogeneity (Tuomisto et al., 2003) because they need different habitats during their life cycle. Ultimately, our results can be useful to identify which guilds and species could be more sensitive to anthropogenic impacts affecting Amazonian freshwater ecosystems. Examples are frugivorous, piscivorous and fin and scale feeder fishes that seem largely dependent on extensive floodplain habitat.

ACK N OWLED G EM ENTS
We are grateful to Céline Jézéquel (Laboratoire Evolution et

CO N FLI C T O F I NTE R E S T
We do not have any conflict of interest.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that supports the findings of this study are available in the supplementary material of this article.