Myrmecophagy has evolved multiple times across mammals (Luo, 2007), reptiles (Pianka and Pianka, 1970; Pianka and Parker, 1975), and amphibians (Toft, 1980, 1981; Simon and Toft, 1991; Daly et al., 1997; Savitsky et al., 2012). The repeated evolution of ant-eating is relatively unsurprising given that ants are conspicuous in both terrestrial and arboreal environments, may constitute 20% of total tropical animal biomass (Fittkau and Klinge, 1973), and perform a myriad of ecological and ecosystem functions (Del Toro et al., 2012; Roslin et al., 2017; Tiede et al., 2017). While some predator–ant interactions evolve towards highly specialized interactions (e.g., dart frogs, whose aposematic colors may reflect ant-harvested toxins [Caldwell, 1996; Santos et al., 2016]), most predators are thought to select ants opportunistically (Redford, 1986; Griffiths et al., 1990). However, this consensus may stem from the lack of diet studies that identify prey taxa to species level (but see Konopik et al., 2014). Though it is well established that ant-eating anurans partition prey based on size (Toft, 1980; Pimentel, 1998; Menendez, 2001), it is unknown if frogs can choose for specific species of ants or for certain ant traits that are available in the surrounding ant community. This uncertainty is particularly evident for tropical anurans whose diets are potentially derived from highly diverse and taxonomically unresolved ant communities.

Tropical leaf litter ant communities are part of a complex ‘brown food web’ (BFW, Coleman and Crossley, 2003). The nutrients in BFWs are derived from detritus, although the processes are poorly understood (Kaspari et al., 2017). In BFWs, most nematodes, mites, and collembolans harvest wood-decomposing microbes, and in turn these groups are fed upon by invertebrates higher in the food web such as ants and millipedes. The abundance and diversity of litter animals can reach astounding numbers in areas as small as 1 m², are determined by environmental (Donoso and Ramón, 2009; Donoso et al., 2013), biogeochemical (Kaspari et al., 2017), and ecological factors (Moore et al., 2004; Donoso et al., 2010; Donoso, 2017), and are major drivers of ecosystem productivity (Endara et al., 2017; Tiede et al., 2017; Schuldt et al., 2018). It is in this context that many animals including anurans feed upon BFW productivity. However, few animal ecologists have incorporated BFW principles into their research, and we therefore know little of how BFWs regulate animal biomass (bottom-up regulation; Redford, 1986), and in turn how animals control BFW communities (top-down control; Morrison, 2018).

Ants are a morphologically, taxonomically, and ecologically diverse group (Del Toro et al., 2012). They provide anurans with potentially different nutrients (Kaspari et al., 2012) and aposematic alkaloids (Daly et al., 1994, 1997; Saporito et al., 2009; Santos et al., 2016; Moskowitz et al., 2018), and yet few studies have looked at the specific identities or traits of ants in frog diets. The first of such studies by Deyrup et al. (2013) found that the Narrow-Mouthed Toad, Gastrophryne carolinensis, fed on up to 43 species of mostly nocturnal ants. These results gave no evidence that frogs exhibit species-level specialization in the ants they eat. Importantly, the diet of G. carolinensis included the largest ants in the area (in the genera Odontomachus and Camponotus, known for their stings and formic acid, respectively), although in relatively low abundance. Konopik et al. (2014) compared the prey (ant) preference of the Borneo River Toad, Phrynoidis juxtaspera, between primary forest and oil plantations in Malaysia and found that prey species remained consistent across sites despite differences in ant communities composition. Finally, McGugan et al. (2016) used mitochondrial barcodes to identify ants and mites present in the diets of the dendrobatid frog Oophaga sylvatica from three localities in coastal Ecuador. Invertebrate alkaloids differed across geography and underlied the alkaloids that were sequestered by O. sylvatica. However, no study has looked at predator preference for specific ant traits.

Rhinella alata is a medium-sized, diurnal, and locally abundant leaf-litter toad in Panama. The species is an exceptionally polymorphic and cryptically colored toad (McElroy, 2016), with a diet almost entirely comprised of ants (Toft, 1980, 1981; Parmelee, 1999; Menendez, 2001; Fajardo-Martínez et al., 2013; Astwood-Romero et al., 2016). The toad is chemically defended, but unlike aposematic frogs, predation of ants by R. alata should not be influenced by toxin sequestration from their diet because R. alata synthesizes its own toxins (bufodianaloids) in paratoid glands (Lyttle et al., 1996). Therefore, R. alata provides a simplified opportunity to test how prey species identity, morphology, and nutritional traits influence predation. Specifically, we answer the following questions: 1) Which ant species does R. alata preferentially consume and avoid? and 2) How do prey species morphological and nutritional traits help explain their rates of consumption? We conducted this study in the Panama Canal, a tropical seasonal forest that harbors a ∼400 species ant community.

MATERIALS AND METHODS

Toad diet collections.—Individual toads were sampled from August to September 2010, on Barro Colorado Island (BCI; 09°09°N, 79°51°W) in Panama, which receives approximately 2600 mm of annual rainfall, with nearly 90% of it falling between May and November (Leigh, 1999). We opportunistically sampled toads and recorded GPS localities at the point of capture (McElroy, 2016). We recorded snout–vent length (SVL) and removed each individual's stomach contents using non-lethal gastric lavage (Solé et al., 2005). Stomach contents were stored in 95% ethanol, and individual toads were released at the point of capture the following day. We identified individual ant prey to species level based on morphology. Non-ant prey items (e.g., mites and Coleoptera) were identified to class or family level.

Sampling the BCI ant community.—We collected 216 leaf litter samples (1 m²) from six sites at BCI. We brought samples to the laboratory and used Berlese funnels to separate invertebrates from the leaf litter. The sampling was designed and implemented as part of broader studies looking at patterns of morphological trait and phylogenetic dispersion across spatial scales in the island (Donoso, 2014) and the world (Gibb et al., 2017; Parr et al., 2017). More information on methods to survey the leaf litter ant community can be found in Donoso (2014). To reduce noise in our analyses, we excluded species that were captured in fewer than five Berlese funnels and that were eaten by fewer than two toads. We retained ant species that were eaten frequently but that were not captured in traps, or vice versa, as these potentially represent prey species that R. alata highly prefers or avoids. We also excluded army ant species, such as those in the genera Labidus and Eciton, because they are swarming species with high spatial and temporal variation in abundance. We used generalized linear models (GLMs) with Poisson error to relate the abundance and diversity of consumed ants to toad sex and SVL.

Prey trait datasets.—Our final dataset comprised 84 ant species and included five continuous and two morphological variables: Weber's length, head width, head length, pilosity (i.e., hairiness), number of spines, sculpture (smooth, moderate, rough), and head color (yellow, yellow-brown, red-brown, brown, black, blue). We chose these morphological traits because ants are likely to rely on them (size, speed, camouflage) as defense (Parr et al., 2017), as well as because frogs could use them as visual cues. Specific information about morphological traits can be found in table 1 of Parr et al. (2017). To test the hypothesis that prey nutritional content and trophic position are associated with predator preference, we compiled a smaller dataset of 40 species for which nutritional data are available (Kaspari et al., 2012). Nitrogen composition (high %N = low C:N ratio) is a general proxy for the nutritional value of ants because %N is positively correlated with protein content and negatively correlated with chitin, an indigestible molecule for most vertebrates (Sullivan et al., 2014). Stable isotope ratios of nitrogen represent a measure of ant trophic position. We report our isotopic data using standard delta (δ) notation, where δ values represent the ratio of the rare to common isotopes. The abundance of δ15N of a consumer is typically enriched by 3.4% relative to its diet.

Selectivity analysis.—To determine the preference of R. alata for each ant species, we used a linear selectivity index (Strauss, 1979; Palkovacs and Post, 2008; Zandonà et al., 2011), which is defined as:

for a prey species i, where L_i is the measure of selectivity, r_i is the relative abundance in the stomach, and p_i is the relative abundance in the environment. To calculate selectivity, we pooled the data for stomach contents and subtracted the proportion of a prey species in the environment from the proportion of that prey species in the stomachs. One benefit of the metric is that it enabled us to retain prey species that were absent from either the stomachs or from the environment. After using simulations to determine a range of L_i values that represent neutral preference, linear selectivity is easy to interpret; prey species with values greater than a positive threshold are considered ‘preferred’ and species with values below a negative threshold are considered ‘avoided.’ To assess significance of linear selectivity values, we replicated our analysis through simulation. We started with the environmental proportions of ant species and generated 1000 matrices by randomly sampling between 3000–5000 ants with replacement. From the simulated matrices, we calculated linear selectivity and generated a null distribution of values for each prey species. The null distributions are species-specific, as they depend on each species’ proportion in the environment. We classified ants as preferred if their selectivity values were greater than the null distribution and avoided if their selectivity values were less than the null distribution. We considered species with real selectivity values that fell within the null distribution to be predated upon at a rate that is proportional with their environmental abundance (i.e., ‘neutral’).

Principal components analysis.—We characterized the phenotypes of ant species by conducting principal component analysis (PCA) of continuous and categorical trait data. We used the function dudi.mix from the package ade4 because it allows for analysis of both continuous and ordered categorical variables (Dray and Dufour, 2007). The PCA was done with a correlation matrix of centered and normalized variables. The function dudi.mix transformed the ordinal variables “trait.L” and “trait.Q” into linear and quadratic variables. In the PCA plot, the “trait.L” vector points toward increasing values of the trait, while the “trait.Q” vector points in the direction of moderate values for the trait. We grouped ant species in the PCA by their preference category to explore which prey traits influence predation.

Stepwise model selection.—Prior to conducting stepwise model selection, we used z-scores to standardize the continuous traits. Head width, head length, and Weber's length were >95% correlated with each other, so we only retained Weber's length as our proxy for size. We used backwards model selection to determine the best model that explains linear selectivity. We started from a full 5-parameter model and used the function ‘dropterm’ in the R-package MASS to iteratively remove parameters from our model. We then compared AICc values for each model, calculated Akaike weights to determine our support for each model, and compared coefficients to assess the importance of each trait. Because Ectatomma ruidum was an extremely highly preferred prey item (i.e., represents an outlier; see below), we performed a second analysis excluding E. ruidum and checked that its inclusion did not skew the results.

Linear selectivity is a continuous response variable that does not incorporate our null model framework. As such, large selectivity values (e.g., E. ruidum) may influence model selection. To solve this, we performed ordered logistic regression with ant traits as the independent variables and preference category as the response variable. We used backwards model selection and AICc values to determine the best model that explains preference category. To explore model accuracy, we trained the model on 70% of the ant species and tested it with the remaining species. We performed 1,000 bootstraps and calculated the percent of the time our model predictions matched the empirical dataset. When our model predictions for a species did not match the empirical dataset, we distinguished between ‘misses’ (i.e., mismatches incorporating a neutral preference) and ‘fails’ (i.e., mismatches between avoid–prefer or prefer–avoid).

Nutrition analysis.—We analyzed a subset of 40 ant species for which we had %N and δN data to determine whether prey nutrition represents an important but overlooked trait. We conducted ANOVA to determine whether toad preference related to ant size (WL), nitrogen content (%N), or trophic position (δN). We used a Tukey's post hoc test to explore significant differences between preferred, avoided, and neutral ants.

RESULTS

Toad diets.—We found 3,645 prey items from 61 ant species sampled from 105 individual R. alata. Of these prey items, 3,462 were ants and 183 (5%) were other taxa. Scarabaeidae (n = 47), Curculionidae (n = 36), and Aranea (n = 34) were the most abundant non-ant taxa. We removed 28 ant species because they were either army ants or were consumed by fewer than two toads, resulting in a final dataset of 3,424 individual ant prey items. The most consumed ant species included Ectatomma ruidum (n = 2,064), Pachycondyla harpax (n = 245), Odontomachus bauri (n = 236), and Solenopsis sp. “lash4” (n = 213).

The community dataset comprised 26,234 ant specimens from 98 species. The most abundant species in the environment were S. sp. “lash4” (n = 4,121), Wasmannia auropunctata (n = 3,647), and Solenopsis sp. ‘JTsp1° (n = 1988). More details of the survey can be found in Donoso (2014). After removing army ants and species found in fewer than five traps, our final community dataset consisted of 25,645 specimens from 84 ant species.

Neither the sex (Poisson GLM, Est. = –0.64, P = 0.65) nor size (Poisson GLM, Est. = 0.01, P = 0.46; Table 1) of individual toads related to the consumed ant diversity. Larger toads ate more ants (Poisson GLM, Est. = 0.05, P ≤ 0.001). Though females were larger than males (males: x̄ = 40.6±0.6 mm, females = 46.1±0.9 mm, P ≤ 0.001), they did not eat significantly more ants than males (Poisson GLM, Est. = –1.05, P = 0.06; Table 1).

Table 1

Poisson GLM analysis of counts of prey diversity and abundance according to frog gender, frog size (snout–vent length, SVL), and their interactions. Significant values are depicted in bold.

Selectivity analysis.—Based on our linear selectivity (L_i) analysis and simulations, we categorized 23 ant species as preferred, 27 as neutral, and 34 as avoided. Larger ant size was necessary but not sufficient to be preferred as a food by toads (Fig. S1; see Data Accessibility). Ectatomma ruidum was overwhelmingly the most preferred ant species (L_i = 0.60), while P. harpax (L_i = 0.07), O. bauri (L_i = 0.07), and S. sp. “lash4” (L_i = 0.06) represented a second tier of preferred prey items (Fig. 1). The first and second most avoided ant species were Solenopsis sp. “yellow” (L_i = –0.16; Fig. 1) and W. auropunctata (L_i = –0.13; Fig. 1).

Fig. 1

The 84 ant species ordered by their linear selectivity value. Preference categories of ant species (preferred [blue], neutral [black], and avoided [red]) were established relative to a null model distribution (gray area). The cutoffs for null distributions are species-specific, which explains why preference categories are not perfectly ordered. Photographs of the top four and bottom four ants are size-scaled in relation to each other.

Principal components analysis.—The first principal component accounted for 38.4% of the variation and represented size traits (Weber's length, head length, and head width), pilosity, and headcolor.L. The second principal component accounted for 16.8% of the variation and was primarily characterized by traits relating to texture (sculpture.L, spines, and headcolor.Q). Overlaying the preference categories on our PCA indicated that neutral and avoided ants largely overlap in trait space, but preferred ant species were larger, darker, and hairier (Fig. 2).

Fig. 2

Principal component analysis of seven ant traits. Ant species were plotted in the PCA and colored by preference categories (preferred [blue], neutral [black], and avoided [red]). The inset depicts the correlations between principal components and each ant trait. “trait.L” (linear) and “trait.Q” (quadratic) are representations of ordinal variables. The linear representation indicates the direction in PC space of increasing trait values, while the quadratic representation indicates the direction in PC space of moderate trait values.

Stepwise model selection.—The AIC model ‘Weber's length + pilosity’ was almost three times more informative than the second best model (‘Weber's length + pilosity + sculpture’) at explaining linear selectivity (Akaike weights = 0.57 vs. 0.21). We found a steep drop off in the ability of subsequent models to explain linear selectivity (Table 2). Removing E. ruidum from the analysis did not influence the top models. However, the one parameter model (‘pilosity’) dropped from 3^rd (Akaike weight = 0.09; Table S1; see Data Accessibility) to 5^th (Akaike weight = 0.01), likely because E. ruidum is very pilose. The model rankings for the ordinal logistic regression analysis were consistent with the linear regression analyses. The top three models explaining preference category (Table 3) and linear selectivity (Appendix S1; see Data Accessibility; excluding E. ruidum) include the ant traits Weber's length, pilosity, and sculpture, and have combined probabilities >90%. We utilized the second best model (‘Weber's length + pilosity + sculpture’) for predicting preference category from ant traits because it includes all the variables in the 90% confidence set. The chosen model correctly predicted the preference category in our training set 70% of the time (x̄ = 0.70; s.d. = 0.10; Fig. S2A; see Data Accessibility). When the model incorrectly predicted preference category it was due to ‘misses’ 93% of the time (x̄ = 0.93; s.d. = 0.10; Fig. S2B; see Data Accessibility) and ‘fails’ only 7% of the time (x̄ = 0.07; s.d. = 0.10; Fig. S2C; see Data Accessibility). Moderately sized ants that were textured were more likely to be preferred, and small ants that were textured were more likely to be neutral (Fig. S2; see Data Accessibility).

Table 2

Model rankings for general linear regression where model variables are ant traits and the response variable is linear selectivity (L_i). For each model, we report the number of parameters (K), AICc, differential AIC (Δ_i), and the Akaike weight (w_i). The Akaike weight can be interpreted as the likelihood for a particular model.

Table 3

Model rankings for logistic regression where model variables are ant traits and the response variable is predator preference (e.g., “preferred”). For each model we report the number of parameters (K), AICc, differential AIC (Δ_i), and the Akaike weight (w_i). The Akaike weight can be interpreted as the likelihood for a particular model.

Nutrition.—Weber's length differed between preference categories (ANOVA: F_(2,37) = 128, P < 0.001). Preferred ants were two and a half standard deviations larger than avoided ants (Tukey HSD: P < 0.001) and two standard deviations larger than neutral ants (Tukey HSD: P < 0.001). Neutral ants were slightly larger than avoided ants (Tukey HSD: P = 0.03). Neither %N (ANOVA: F_(2,37) = 0.028, P = 0.97) nor δN (ANOVA: F_(2,37)=1.011, P=0.37) differed between preference categories (Fig. 3).

Fig. 3

Species trait values for size (WL) and nutrition (%N, δN) grouped and colored by predator preference. Line graphs represent mean and standard error. The relationship between trait values and predator preference were tested with ANOVA.

DISCUSSION

While it is well established that R. alata eats ants, we extend previous studies by identifying prey to species-level and comparing prey abundances with species abundances in the leaf litter habitat. We found that R. alata prefers to eat relatively rarer large ants and avoids eating common smaller ants. Generally speaking, optimal foraging theory predicts that an organism will behave and forage in a manner that maximizes fitness by reducing the energetic and opportunity costs associated with consuming poor-quality prey (Emlen, 1966). In this context, that R. alata prefers the largest (and more rare) ants suggests that the nutritional gain from consuming large ants outweighs the energetic cost of searching for relatively rare prey. However, it is also plausible that the apparent avoidance of small ants by R. alata is a result of their visual system and inability to see the smallest ants. Furthermore, we found no relationship between predator preference and nutrition content of prey. Thus, R. alata does not select for individual prey items that are high in nutritional value; instead they appear to increase energy intake by choosing large prey over more nutritious prey.

The best models explaining selectivity and preference include elements of texture, namely hairiness (pilosity) and rugosity (sculpture). This is the first anuran field study to find that prey traits other than size can explain predation rates. However, it is difficult to envision a scenario in which R. alata discerns between—and differentially preys upon—textured and untextured ants. Instead, we propose that R. alata preferentially selects for large ants and that predation success increases with increasing prey texture. Frog tongues are highly specialized for fast and reliable adhesion to prey items. Salivary glands located within the tongue increase the production and availability of highly viscous saliva and are instrumental in the tongues ability to adhere to prey items (Kleinteich and Gorb, 2015a; Noel and Hu, 2018). Micro-structures on the tongue are thought to increase the adaptability of the tongue to attach to uneven prey surfaces and facilitate the development of mucous fibrils which sustain prey adhesion during tongue retraction (Kleinteich and Gorb, 2015b; Kleinteich and Gorb, 2016). How variable prey texture (e.g., fur, hair, feathers) may influence tongue adhesion remains poorly understood (Kleinteich and Gorb, 2015a). However, microstructures of prey (pilosity and rugosity) may be analogous to microstructures on the tongue by promoting mucous fibrils and increasing the strength or length of tongue adhesion. Thus, we hypothesize that adhesion between a frog's tongue and a prey item is greater for textured ants than it is for untextured ants, which increases the probability of successfully handling and consuming pilose and rugous ant species. Understanding more about the general functional benefits of texture traits (e.g., pilosity and sensory capability; sculpture and desiccation resistance) and the potential trade-offs associated with predation deserves investigation.

Ectatomma ruidum, a common item in Neotropical frog diet studies (Weber, 1938; Lopez et al., 2007), had the highest selectivity value and was the most consumed species in our study representing 60% of all stomach items, despite it only representing 0.32% of the ant community. An underlying assumption while calculating linear selectivity is that the environmental community has been accurately sampled. As such, any trapping bias against certain species or phenotypes (e.g., size) may skew our findings. Ectatomma ruidum, while low in terms of abundance, was collected in 27% of the samples, indicating that it is well sampled and commonly encountered in the leaf litter, which gives us confidence in our environmental estimation. In fact, the next three most preferred ant species were all well represented in our sampling procedure and occurred in 21%, 12%, and 11% of traps, indicating that elevated selectivity values for those species is unlikely to result from environmental sampling bias (Fig. S1; see Data Accessibility). Ectatomma ruidum has the characteristics we expect in a preferred species: it is one of the largest ants on the island, and it is among the most textured species in our dataset. Even though nutrition and trophic position were not good predictors of preference of R. alata, it is intriguing that E. ruidum is one of the more nutritious ant species in terms of %N and δN. Importantly, size and pilosity predict predation of R. alata even when E. ruidum is removed from the analyses or when we analyze preference categories (as opposed to linear selectivity). Therefore, the extraordinary presence of E. ruidum in the diet of R. alata provides a natural validation of our result.

Though we did not formally test for avoidance of defensive traits in our analysis, we were unable to identify any patterns relating predator avoidance to other hypothesized defensive traits. While W. auropunctata and other weedy ants in the genus Solenopsis were the most avoided ant species, we attribute this primarily to their small size. Certainly, these species can be noxious; W. auropunctata is known to be invasive and to displace ants (Arnan et al., 2018) and other animals as well (reviewed in Wetterer and Porter, 2003). Solenopsis includes the fire ants with powerful stings. Nonetheless, Solenopsis ants are the preferred food items for the poison frog Oophaga histrionica in Colombia (Osorio et al., 2015), and in our analysis Solenopsis were preferred if they were large enough (e.g., Solenopsis sp. “lash4”). Furthermore, despite being removed from the analysis, we found army ants (subfamily Dorylinae) in stomachs of R. alata, indicating that aggressive ants with pincers were not necessarily a predator deterrent. Rhinella alata also preferred large ponerine ants (e.g., P. harpax and O. bauri) known for their painful stings (see Deyrup et al., 2013), and they preferred Trachymyrmex ants (e.g., T. isthmica and T. zeteki) that had the most spines (Parr et al., 2017). Taken together, these results indicate that ant traits that are traditionally thought of as anti-predator defenses are unlikely to play a large role in deterring Rhinella alata. Instead, predation is more likely to be influenced by prey size and foraging behavior than anti-predator defenses.

Arthropod declines can initiate bottom-up trophic cascades that affect insectivorous vertebrate predators, potentially resulting in localized extirpation at higher trophic levels (Lister and Garcia, 2018). Understanding the pheno-typic traits that underlie trophic interactions allows for nuanced insights into food web dynamics (Spitz et al., 2014). Rhinella alata prefers large ants, which in general are expected to be negatively impacted by global climate change (Gibb et al., 2015, 2018). If large ants are negatively affected by global change, vertebrate predators like R. alata may be forced to shift their diets towards smaller ant species. This study complements previous work showing that ant communities on BCI are structured by sets of phylogenetically similar ants of small size (Donoso, 2014). The preferential predation of large ant species may be partially responsible for producing this pattern (Abrams and Rowe, 1996; Roslin et al., 2017). It is difficult to draw general conclusions on the impact that prey traits have on myrmecophagous predator preference until more species-level diet studies accumulate. We provide one example for a non-alkaloid-sequestering myrmecophagous frog that clearly selects for large ants despite their scarcity in the environment. Our intriguing finding that pilosity and rugosity influence predation highlights that prey texture may be an overlooked factor in studies on the biomechanics of prey capture.

DATA ACCESSIBILITY

Data are available at  https://github.com/mattmcelroy/ant_buffet. Supplemental information is available at  https://www.copeiajournal.org/ch-18-162.