Digital Commons@Georgia Southern Predator-based Selection and the Impact of Edge Sympatry on Components of Coralsnake Mimicry

Studying warning coloration and mimicry is an effective way to understand predator-driven selection and phenotypic diversity. The presence (sympatry) or absence (allopatry) of a toxic model plays a role in shaping mimetic phenotypes. However, the impact of edge sympatry and allopatry on predation of mimetic phenotypes is not well understood. We studied coralsnake mimicry to test how edge sympatry and allopatry affect predation on mimetic phenotypes. Specifically, we tested 1) if overall attack rates varied with edge sympatry of coralsnakes 2) which color patterns conferred a fitness advantage 3) which specific mimetic signal components are important in driving predatory attacks and 4) whether selection patterns varied temporally. We deployed clay replicas that utilized a cryptic pattern, two different signal components (red and white), and a tricolor pattern that included both signal components. We found that overall attacks did differ between edge sympatry and allopatry, with higher attack rates in allopatry. All mimetic phenotypes in 2019 had higher attack rates than cryptic phenotypes in edge sympatry, with a similar but nonspecific pattern in allopatry. Replicas with red and banded patterns received more predatory attacks in edge sympatry than those without, once again with a similar pattern in allopatry. There was also a difference in attack rates and patterns between years, indicating temporally variable selection such as frequency-dependent selection. These results suggest that mimetic phenotypes may not have a fitness advantage in areas of edge sympatry or allopatry. This suggests the role of sympatry and allopatry may be more complex than previously thought, particularly in how sympatry may interact with extraneous factors such as behavior and frequency of phenotypes.

A fundamental goal of biology is understanding the forces that drive the evolution of phenotypic diversity. In order for evolution by natural selection to occur in a population, the population must have both variation in phenotype and a selective force (such as predation) acting upon the population (Endler 1986;Freeman and Herron 2004). Studying predator-driven selection is an effective method to understand phenotypic diversity because the selective agent is known, and much has been learned by analyzing the evolutionary drivers of antipredator traits such as warning coloration (e.g. aposematism) and mimicry (Brodie III 1993;Brodie III and Janzen 1995;Ham et al. 2006;Kleisner and Saribay 2018;Kuchta 2005;Pfennig et al. 2001;Pfennig et al. 2007;Quicke 2017). For aposematism, conspicuous coloration acts as a signal to predators that the prey is unpalatable and thus promotes avoidance of inedible or dangerous prey (Arbuckle and Speed 2015;Kuchta 2005;Leimar et al. 1986;Mappes et al. 2005;Ruxton et al. 2004). Mimics then honestly (Müllerian) or deceitfully (Batesian) replicate this signal to deter predators (Quicke 2017). Therefore, aposematism and mimicry provide a direct link between phenotype and predator-driven selection and can offer strong evidence for how evolution by natural selection operates in nature (Davis Rabosky et al. 2016;Endler 1986;Harper and Pfennig 2007;Kleisner and Saribay 2018;Kuchta 2005;Leimar et al. 1986;Stevens and Ruxton 2012).
For example, the abundant literature on mimicry has revealed that mimetic phenotypes can be subject to multiple types of selection, including directional, stabilizing, and frequency dependent selection (Akcali et al. 2018;Akcali and Pfennig 2014;Cox and Davis Rabosky 2013;Holmes et al. 2017;Lindstedt et al. 2011;Mappes and Alatalo 1997;Ruxton et al. 2004). One factor that can influence the form of selection on mimetic phenotypes is the presence (occurring in sympatry) or absence (occurring in allopatry) of the toxic model species in the habitats of the mimetic species (Finkbeiner et al. 2018;Greene and McDiarmid 1981;Harper and Pfennig 2007;Kikuchi and Pfennig 2013;Pfennig 2016;Pfennig et al. 2001;Pfennig et al. 2007;Pfennig and Mullen 2010;Ruxton et al. 2004). However, the effect of this co-occurrence, or lack thereof, is complex and is only well studied in some well characterized systems (Akcali and Pfennig 2017;Cox and Davis Rabosky 2013;Harper and Pfennig 2007;Holmes et al. 2017;Pfennig et al. 2001;Pfennig et al. 2007;Pfennig and Mullen 2010).
The effectiveness and precision of mimicry varies with presence and abundance of models (Akcali and Pfennig 2017;Finkbeiner et al. 2018;Harper and Pfennig 2007;Kikuchi and Pfennig 2013;Lindström et al. 1997;Owen and Owen 1984;Pfennig et al. 2001;Ruxton et al. 2004;Yamauchi 1993). For example, mimetic precision decreases in areas with denser populations of models and increases where the model is rare (Akcali and Pfennig 2014;Harper and Pfennig 2007). Because population densities often decrease in geographic range margins (Hengeveld and Haeck 1982), the relative distance from the sympatry-allopatry border (i.e. edge sympatry vs deep sympatry) can affect mimetic precision. This link between selection and sympatry could be even more important when the model is deadly (Lindström et al. 1997;Pough 1988;Quicke 2017). For example, in the harmless snakes that mimic deadly coralsnakes (family Elapidae, which also includes mambas and cobras), components of color mimicry patterns are precise in edge sympatry, may be relaxed in deep sympatry, and may not convey a fitness advantage at all in allopatry in a temperate mimicry system ). In addition, in areas of allopatry where coralsnakes only recently went extinct, their mimics became more precise, supporting the idea that mimicry must be more precise in areas where the model is rare (Akcali and Pfennig 2014). These results imply directional selection towards precise mimicry in a newly allopatric region. However, this research has largely been focused on temperate zones, with few studies conducted in diverse tropical ecosystems and even fewer focused on montane tropical environments. The diversity in these systems yields a vast range of aposematic patterns on both coralsnakes and mimics (Savage and Slowinski 1992).
Because aposematic coloration generally leads to stabilizing selection on models, this may indicate that selection acts differently in areas with high biodiversity (Joron and Mallet 1998;Mallet and Turner 1997). The montane tropics not only have higher biodiversity for both coralsnakes and mimics than temperate systems, but also provide habitats both in allopatry and edge sympatry of coralsnakes and their mimics in close geographic range as well as under a variety of different predator assemblages (Figure 1.1) (Townsend and Wilson 2008). Thus, more research is needed to determine how selection on mimetic phenotypes in the montane tropics might act differently than in temperate habitats. Research that focuses on mimicry dynamics in the poorly studied montane tropics would allow us to see how selection on mimicry could change with habitat.
We studied the impact of edge sympatry and allopatry on predation rates on coralsnake banding patterns in the montane tropics of Honduras. Highly venomous, elapid coralsnakes have evolved aposematism via a brightly colored and banded phenotype, which has subsequently been mimicked by a variety of nonvenomous colubroid snakes in a classic example of Batesian mimicry (Davis Rabosky et al. 2016;Greene and McDiarmid 1981;Hinman et al. 1997;Pfennig et al. 2007;Pough 1988;Quicke 2017). This mimicry system provides an excellent model to study the evolutionary drivers of phenotypic diversity because it involves multiple species of both models and mimics and spans multiple habitats (Quicke 2017;Savage and Slowinski 1992).
However, most research on coralsnake mimicry has focused on species-poor temperate systems that often have only one model and one to three mimics. In these systems, there are few local phenotypes, and so it may be more difficult to determine which aspects of these phenotypes are important in deterring predation. In systems with many mimics, we can tease out precisely which . By analyzing predation rates on mimicry patterns in edge sympatry and allopatry in the montane tropics, we can achieve a greater understanding of which mimetic phenotypes are most successful in areas with a relatively high diversity of phenotypes.
We conducted a field experiment using clay replicas of snakes along an elevational gradient to determine how attack rates on aposematic snakes change with edge sympatry or allopatry with coralsnakes in the montane tropics of northwestern Honduras. Specifically, we studied whether predation rates differ with level of sympatry (edge sympatry or allopatry) and mimetic signal components (i.e. presence of red, white, or bands) within the color pattern. First, we tested whether overall attack rates on models varied between edge sympatry and allopatry.
Greater intensity of predation in one habitat over the other could suggest selection is stronger there. Second, we tested which mimetic color patterns conferred a fitness advantage in edge sympatry or allopatry. Empirical research suggests that mimetic phenotypes should convey a fitness advantage in edge sympatry but should have a fitness cost in allopatry (Akcali and Pfennig 2017;Harper and Pfennig 2007;Ries and Mullen 2008). However, if mimetic phenotypes convey a fitness advantage in both habitats, this implies that the fitness advantage of mimetic signals is maintained allopatry close to the range edge. Conversely, if mimetic phenotypes impose a fitness cost in both edge sympatry and allopatry, then this implies that the mimetic signals could be maladaptive in range edges and allopatry. Third, we tested which mimetic signal components would be important in driving predatory attacks. Previous research has suggested that both the presence of red and banding are critical signal components of coralsnake mimicry (Akcali and Pfennig 2014;Pfennig and Mullen 2010). Finally, we examined whether attack rates varied temporally, which could suggest negative frequency dependent selection. Negative frequency-dependent selection occurs when the most common phenotypes in a population are attacked disproportionally more by predators. Over time, this allows rarer phenotypes to become more prominent in the population, thus creating temporal variation (Holmes et al. 2017). Our results will contribute to understanding the effect edge sympatry and allopatry have on predation rates on coralsnakes and their mimics in the tropics and so will provide insight into the broader evolutionary processes that drive the evolution of phenotypic diversity. Honduras. Dotted line at 1800m represents range edge of coralsnakes. Neither coralsnake species in the park is found above 1800m, but mimics can be found up to the highest elevations in the park.

Study site
The study location in Cusuco National Park, Honduras, houses a tropical, montane environment that includes both sympatric and allopatric populations of coral snakes and coral snake mimics (Townsend and Wilson 2008). The park is a protected area located in the

Field experiment
In order to assess predation rates on various mimetic and non-mimetic color patterns, we analyzed the number of attacks on clay replicas. We constructed the replicas using non-toxic, pre-colored clay, and they displayed one of four color patterns. These patterns are based on native models and mimics: fully mimetic (tri-color), black and red banded, black and white banded, or brown (Figure 2.1).
We placed replicas in groups of 4, hereafter referred to as a tetrad, which included one of each replica type. Over a span of 8 weeks from June to August in 2018 and 2019, we arranged the tetrads along 9 or18 transects, respectively, at varying elevations within the park. Transects material, figure S1a). We placed each replica within a tetrad at least 3 meters apart from other replicas in the tetrad. The order of the replicas within each tetrad was randomly determined using a random number generator. We calculated an ideal sample size for 2019 through a power analysis based on the attack data from 2018. We left replicas out on transects for a period of two weeks and checked them every 3-4 days for marks of predation (for example, distinct beak, bite, or claw marks). Once checked, replicas were smoothed over and placed in the same spot. If a replica was attacked more than once during the two weeks, only one attack was counted. Each instance of attack was scored for type of predator (bird, mammal, or unidentifiable) and recorded. Different observers checked and scored the replicas from year to year.

Statistical Analyses
We analyzed the data using a combination of model comparison of mixed effect models and contingency analyses. To avoid missing patterns of selection that could occur by discarding ambiguous attack marks, we combined bird and mammal attacks into an overall "attacked" category. For analysis, replicas were classified according to five color variables. These variables included one categorical variable of replica color pattern (brown, white & black, red & black, or fully mimetic; hereafter referred to as individual pattern), and four binary variables: the presence of red, presence of white, presence of bands, and whether it is a mimetic color pattern. We created logistic mixed effect models in R and compared them using Akaike information criterion (AIC) values to determine which models were most effective in predicting attack rates (Bates et al. 2015). The response variable for these models was whether a replica was attacked or not attacked during the two week period. The null models for each year included only transect as a random effect. The remaining models included transect as a random effect as well as all combinations of: level of sympatry, one of the five color variables, and interaction terms. We performed contingency analyses in JMP.

Model comparison
Sympatry and presence of red were the most important factors impacting attack rates on models. In 2018, the null model had the lowest AIC score (Table 3.1). All of the models that fell within 2 AIC units of the best model had only one additional term after transect: sympatry, the presence of white, red, or bands, or whether it was mimetic. In 2019, the best model included transect, sympatry, and the presence of red (no interaction) (Table 3.2). Three models fell within 2 AIC units of the best of model, and they all contained sympatry as a term in the model. Those models also included individual color pattern both with and without an interaction, and the presence of red with an interaction.

Contingency analyses
The level of sympatry impacted attack rates: we found a difference in attack rates on all replica colors combined in allopatric and edge sympatric habitats with a contingency analysis in data from 2019 (p < .0001) but not 2018 (p = 0.2869) (Figure 3.1).
When taking color pattern of replicas into account, we found differences between edge sympatric and allopatric habitats in the specific patterns and signal components that were important in affecting attack rates (Figure 3.2). Specifically, we found that in 2019 individual pattern of the replica was not independent of getting attacked in edge sympatric habitats, but this was not found to be significant in allopatric habitats (Table 3.3). Additionally, we did not find this pattern in 2018 in either edge sympatric or allopatric habitats.
When analyzing within allopatric and edge sympatric habitats, the presence of red and bands had an impact on attack rates. We did not find any differences between edge sympatric and allopatric habitats in 2018 (Table 3.3). However, in 2019 we found that in edge sympatric habitats both replicas with both red and bands were attacked more (Table 3.3, Figure 3.3). In allopatric habitats, only the presence of red affected attack rates (Table 3.3, Figure 3.3). Table 3.1 Results of mixed effect model comparison using Akaike Information Criterion (AIC). Linear models were created to predict whether a clay replica in 2018 was attacked using the following terms: the transect on which it was located, if it was in sympatry, a color variable (the pattern, presence of white, presence of red, presence of bands, or whether it is mimetic), and an interaction between sympatry and the color variable. K is the number of parameters and ΔAIC is the difference between that model and the best model.  Results of mixed effect model comparison using Akaike Information Criterion (AIC). Linear models were created to predict whether a clay replica in 2019 was attacked using the following terms: the transect on which it was located, if it was in sympatry, a color variable (the pattern, presence of white, presence of red, presence of bands, or whether it is mimetic), and an interaction between sympatry and the color variable. K is the number of parameters and ΔAIC is the difference between that model and the best model.

DISCUSSION
Contrary to expectations, we found that while sympatry does impact predation, coralsnake patterns do not have a fitness advantage in either edge sympatry or allopatry.
Specifically, replicas without mimicry components almost always had lower attack rates than replicas with the three mimetic patterns, regardless of both year and whether they were placed in an edge sympatric habitat or an allopatric habitat. We also found that the presence of red and bands were particularly important regarding attack rates in 2019 in areas of edge sympatryreplicas with red or bands had higher attack rates than those without. Attack rates in allopatry showed similar patterns but likely do not have enough statistical power due to a lower sample size. This is in contrast to what has previously been found, where both brown replicas and replicas with imprecise mimetic patterns were attacked more than accurate mimetic replicas in sympatric regions (including edge sympatric) (Brodie III 1993;Brodie III and Janzen 1995;Kikuchi and Pfennig 2010;Pfennig et al. 2001). In addition, we found patterns of attacks on models varied temporally, which in a stable, tropical system, could be caused by negative frequency-dependent selection (Akcali et al. 2018;Cox and Davis Rabosky 2013;Holmes et al. 2017). Our research implies that mimicry could operate differently in the montane tropics than in temperate habitats and may not be entirely advantageous.
We found that mimetic traits do not convey a fitness advantage of protection from predators and might even confer a fitness cost in both edge sympatry and allopatry in the montane tropics. In temperate regions, mimetic patterns can incur a fitness cost that varies between allopatric regions with different local mimics, which implies that even between allopatric populations in similar habitats there can be differences in patterns of selection ). Predators in the southeast United States differentially attacked replicas depending on whether they were in deep sympatry or edge sympatry . Those predators avoided both poor and accurate mimetic phenotypes in areas of deep sympatry, whereas they avoided only accurate mimics in edge sympatry. Furthermore, avian predators in the lowland tropics attacked brown replicas more than tricolor or red & black banded replicas (Brodie III 1993;Brodie III and Janzen 1995). However, while we did find overall attack rates were lower in areas of edge sympatry, we did not find that predators avoided the more accurate mimetic phenotypes. Indeed, the replicas with phenotypes most closely resembling those of coralsnakes (tricolor and red & black banded) were often those attacked most both in areas of edge sympatry and allopatry. These results support the idea that in areas where coralsnakes are rare or absent, mimetic phenotypes may not confer as much an advantage as in areas where coralsnakes are common (such as the lowland tropics), which is consistent with classic Batesian mimicry theory (Brodie III and Janzen 1995;Pfennig and Mullen 2010;Quicke 2017;Ruxton et al. 2004). Given the fitness cost of mimetic phenotypes in the montane tropics, what permits the persistence of multiple species of models and mimics in these areas? There are three primary reasons that mimicry could be maintained in the face of apparent maladaptation: 1) there is an interplay of aposematic coloration and behavior and thus mimetic phenotypes are not actually maladaptive 2) mimetic phenotypes are indeed maladaptive but may be maintained by gene flow 3) the apparent maladaptation of mimetic phenotypes fluctuates temporally, likely due to negative frequency dependent selection (Pfennig and Mullen 2010).
First, the behavioral component of coralsnake mimicry might reinforce and strengthen the avoidance of mimetic patterns. For example, brown replicas of frogs were attacked significantly more when moving than when stationary, whereas red replicas of aposematic frogs were attacked significantly more when stationary than when moving (Paluh et al. 2014). Coralsnakes display stereotyped coiling and thrashing behavior when threatened, and thus behavior is an integrated part of their antipredator coloration (Davis Rabosky et al. 2020). These findings suggest that movement and behavior may play an important role in selection for both cryptic and aposematic individuals (Bateman et al. 2017). Hence, while the use of clay replicas provides a useful tool for studying predation, a limitation of the method is that it is unable to capture the effects of behavior (Bateman et al. 2017). Our results suggest that there may also be variation in how important movement is in relation to selection behavior. In tropical, sympatric regions where coralsnakes are relatively common, stationary mimetic replicas were avoided as predicted by mimicry theory (Brodie III 1993;Brodie III and Janzen 1995). However, in range edges where the models may be rare, selection for avoidance of aposematic phenotypes may be weaker and thus behavior could play an important role in reinforcing color signals (Charlesworth and Charlesworth 1975). Whether mimetic color signals coupled with behavior would confer a fitness advantage in the montane tropics would be a fruitful area for future research.
Second, mimetic phenotypes could indeed be maladaptive in areas of edge sympatry and allopatry, but migration and gene flow from regions where mimicry is adaptive (areas of deep sympatry at lower elevations) maintains the presence of mimetic phenotypes (Ries and Mullen 2008). Coralsnake mimics in temperate systems migrate from areas of sympatry to allopatry based upon indirect DNA evidence (Harper and Pfennig 2008). Because our study was conducted on a limited geographic scale, high-elevation allopatric sites were within a few kilometers of low elevation areas, so immigration from higher coralsnake density in lower elevations to higher elevations with lower coralsnake density is plausible. Similarly, range expansion of mimics may create newer areas of edge sympatry or allopatry where predators have not yet evolved or learned avoidance (Pfennig and Mullen 2010), although this is less common in the climatically stable tropics. Thus, there could be enough immigration and gene flow of mimics between high elevations without coralsnakes and low elevations with coralsnakes that allow mimicry to persist at high elevations.
Finally, mimicry might be maintained despite a fitness cost through temporally fluctuating selection, such as negative frequency-dependent selection. Indeed, we found evidence of temporal variation in predation rates on coral snake color patterns, yet we cannot rule out variation in experimental design between years that could contribute to temporal variation. Many mimicry systems are influenced by frequency-dependent selection, including frogs, snails, butterflies, and snakes (Holmes et al. 2017;Ries and Mullen 2008). When this selection allows rarer phenotypes to increase in population, over time this could change the ratio of which patterns are most common, thus altering the pattern that is targeted by predators (Holmes et al. 2017;Pfennig et al. 2007). Frequency-dependent selection has been shown to play an important role in Batesian mimicry (Akcali et al. 2018;Cox and Davis Rabosky 2013;Finkbeiner et al. 2018;Holmes et al. 2017;Pfennig et al. 2001;Pfennig and Mullen 2010). In the context of the coralsnake mimicry system, in areas where models are uncommon or absent and mimics are abundant, the high abundance of harmless mimics might override predator avoidance, thus leading to predators learning or evolving a preference for conspicuous prey Pfennig and Mullen 2010). This in turn reduces the population of mimics and potentially induces polymorphic mimics to shift morph frequencies to non-mimetic morphs. The decrease in mimics then increases the ratio of models to mimics, which exerts stronger selection on predators to avoid mimetic phenotypes. Mimics then have a fitness advantage in survivorship. Therefore, negative frequency-dependent selection could result in a dynamic relationship between the frequencies of different mimetic phenotypes, causing temporal heterogeneity in the phenotypes most selected against (Cox and Davis Rabosky 2013). Our study may have taken place at the point in the cycle where mimetic phenotypes are maladaptive, yet they are able to persist because the phenotype is favored when rare. The presence of both learned and innate avoidance would likely impact the dynamics of frequency dependent cycles (Akcali et al. 2018). For example, naïve juveniles of motmots and great kiskadees innately avoid coralsnake patterns, despite no experience with snakes (Smith 1975;Smith 1977). However, tropical birds such as puffbirds and adult motmots have been documented to consume coralsnakes (Smith 1969). While predators likely have a combination of innate and learned avoidance, learned avoidance would lead to shorter periods of negative frequency-dependent cycles than innate avoidance because innate avoidance relies on intergenerational processes.
While we found that the mimetic phenotype did not confer a fitness advantage, and may impose a fitness cost of decreased survivorship, our results were consistent with previous research on coralsnake mimicry. Replicas that included red as a signal component were attacked more than those without. In addition, replicas with just red and black bands were attacked more than the tricolor replicas in areas of edge sympatry during both years. In allopatric habitats, however, tricolor replicas were attacked more than red and black replicas. This could indicate that it is beneficial to be a precise mimic in areas where the model is rare but present, but that protection breaks down farther from the model's range. These results are consistent with previous findings where predators avoided good mimics over poor mimics in edge . Observational studies using museum specimen have also found differences in mimetic precision in relation to sympatry and allopatry (Akcali and Pfennig 2014;Akcali and Pfennig 2017;Harper and Pfennig 2007). Within a single species, the most precise mimics to a local model were found at the edge of the model's range rather than in deep sympatry . However, other studies have found that mimetic precision differs between species, and in some cases the most precise mimics have been found in areas of allopatry (Akcali and Pfennig 2014;Akcali and Pfennig 2017). A potential future study could use museum specimen from the tropics to link our findings on selection with patterns of mimetic precision to determine if these patterns are similar between temperate and tropical habitats.
Our data suggest that coralsnake mimicry may not provide an advantage in areas where coralsnakes are rare or absent, such as in edge sympatry and allopatry. Despite this, mimicry persists. This may be because of gene flow and immigration of both mimics and predators, frequency-dependent selection, or a combination. Frequency-dependent selection may also play a role in the temporal heterogeneity seen in our results. In addition, behavior is an important aspect of antipredator defenses in coralsnakes that is unable to be captured using stationary clay replicas yet may play an integral role in selection. Mimicry may even operate differently between temperate and tropical habitats. To further understand the complex nature of how sympatry and allopatry effect selection on coralsnake mimics, studies comprising multiple years and locations ranging from deep sympatry to deep allopatry would be ideal.