Foraging predicts the evolution of warning coloration and mimicry in snakes

Significance

Warning coloration and Batesian mimicry have been thoroughly studied since Wallace’s time, but concomitant foraging costs have been largely overlooked. We evaluated the evolution of such striking coloration in snakes using a phylogenetic comparative approach. Our results indicated frequent gains and losses of conspicuous coloration, correlations between conspicuous coloration and secretive ecology, and recurrent evolution of mimicking inconspicuous models, all of which were well explained by the trade-offs between warning and foraging. These results demonstrate that additional selection factors and selection regime make the evolution and diversity of warning coloration and mimicry less confusing. Our findings help answer long-lasting questions in the classical systems and also raise many new questions about how multiple selection factors interact in shaping animal coloration.

Abstract

Warning coloration and Batesian mimicry are classic examples of Darwinian evolution, but empirical evolutionary patterns are often paradoxical. We test whether foraging costs predict the evolution of striking coloration by integrating genetic and ecological data for aposematic and mimetic snakes (Elapidae and Dipsadidae). Our phylogenetic comparison on a total of 432 species demonstrated that dramatic changes in coloration were well predicted by foraging strategy. Multiple tests consistently indicated that warning coloration and conspicuous mimicry were more likely to evolve in species where foraging costs of conspicuous appearance were relaxed by poor vision of their prey, concealed habitat, or nocturnal activity. Reversion to crypsis was also well predicted by ecology for elapids but not for dipsadids. In contrast to a theoretical prediction and general trends, snakes’ conspicuous coloration was correlated with secretive ecology, suggesting that a selection regime underlies evolutionary patterns. We also found evidence that mimicry of inconspicuous models (pitvipers) may have evolved in association with foraging demand for crypsis. These findings demonstrate that foraging is an important factor necessary to understand the evolution, persistence, and diversity of warning coloration and mimicry of snakes, highlighting the significance of additional selective factors in solving the warning coloration paradox.

The theory of warning coloration predicts the evolution of bright, contrasting coloration in unprofitable prey because such coloration promotes innate and learned aversion by predators (1, 2). In terrestrial systems, warning coloration is commonly red, yellow, orange, and/or white combined with black, which yields strong contrast against natural backgrounds and within the body (1–4). Palatable prey may benefit by using deceptive warning signals, which leads to the evolution of resemblance to the local warning coloration (Batesian mimicry) (1, 5).

Although their anti-predator function is well understood, some aspects of warning coloration and mimicry remain confusing. First, despite the advantage of strong warning signals, many unprofitable prey animals are inconspicuous (2, 4, 6). Second, there is inconsistency in the evolutionary patterns among animal lineages even though warning coloration or mimicry works by a common mechanism. For example, the transition from ancestral crypsis to mimicry is usually one-directional in butterflies (7–9), but reversion to crypsis is frequent in snakes (10). Finally, the theory predicts the convergence of warning coloration because warning signals are more effective at educating predators when they are similar to each other, but warning coloration in nature is strikingly diverse (11).

Because animal coloration serves a variety of functions, multiple selection pressures simultaneously act on coloration. Empirical and theoretical studies indicate that selection agents other than predators might play important roles in the diversification of warning coloration (1, 4, 11). Although the effects of sexual selection and thermoregulation have been investigated (1, 4, 11), other selection factors have been largely overlooked (11). Warning coloration and Batesian mimicry have evolved in animals that are themselves predators as well as prey (e.g., snakes, spiders, frogs, and some fishes). In these intermediate predators, conspicuous anti-predator coloration may reduce foraging efficiency, as is sometimes the case for sexual coloration (12, 13). Given the fitness significance of foraging efficacy, foraging-induced conspicuousness cost (FC cost) may have a significant impact on the evolution of conspicuous anti-predator coloration.

Snake radiations provide an ideal opportunity to address such issues. Many species of venomous snakes in the family Elapidae exhibit strong warning coloration, although those that do not are usually rather cryptic (Fig. 1 A and B). In the Neotropics, approximately 18% of snake species are putative Batesian mimics of elapid coralsnakes (10, 14). Harmless snakes of the family Dipsadidae include especially large numbers of such mimetic species, although crypsis is also common in this family of >800 species. Also, some dipsadids are suggested to mimic another group of highly venomous snakes, pitvipers (15, 16). In contrast to coralsnakes and their mimics, pitvipers and their mimics are cryptic (Fig. 1C).

Fig. 1.

Diversity in snake coloration. These species all have extreme venoms but radically differ in coloration. Many species of the family Elapidae are cryptic (A), but some elapids use strong warning coloration (B). Venomous pitvipers (Viperidae) often have characteristic geometric patterns that may aid in both camouflage and warning (C). Brightly colored elapids and cryptic pitvipers are mimicked by many other snakes. (A) Eastern brown snake (Pseudonaja textilis), courtesy of Bridget Lunn. (B) Arizona coralsnake (Micruroides euryxanthus), courtesy of Andrew Durso. (C) Terciopelo (Bothrops asper), courtesy of Ryobu Fukuyama.

We hypothesized that foraging ecology might predict the evolution of warning coloration and mimicry and made the following predictions. First, conspicuous coloration correlates with foraging traits that mitigate FC cost, such as concealed habitat, nocturnal activity, active foraging, and feeding on animals with poor vision. Because a conspicuous species need not have all of these traits simultaneously, correlations between coloration and individual ecological traits may not necessarily be strong. Second, complex patterns of correlations can be interpreted in terms of an integrated FC cost. Conspicuous coloration will be strongly correlated with low FC cost, reflecting a direct association. Here we test these predictions by investigating macroevolutionary patterns within elapids and dipsadids using a phylogenetic comparative approach. To test the first prediction, we employed a Bayes factor test widely used for testing correlated evolution of discrete traits (Pagel’s test, ref. 17) and another more conservative test, concentrated changes test (CCT). For the second prediction, we defined a new composite variable, foraging type, as an indicator of FC cost. Since the foraging type and associated FC cost is assumed to be ordinal, we also performed regression analysis that accounts for the ordinality (see below).

Results

We accumulated genetic and phenotypic data (coloration and ecology) for 194 species of elapids and 238 species of dipsadids through an extensive literature survey. Our survey clarified that 68 of the 194 elapids have a contrasting dorsal pattern, such as bands and head/tail-cap, consisting of red/yellow/white and black or uniform red, which is typical warning coloration. We here refer to such coloration as “bold coloration.” For Dipsadidae, 46 of the 238 species were listed as having bold coloration. Eight genera were indicated to include putative pitviper mimics (Dataset S2).

Regarding ecology, we compiled data on habitat, diel activity, foraging mode, and diet. Foraging ecology of elapids and dipsadids is diverse and incompletely known (Dataset S3). We categorized species into five foraging types in terms of their FC cost (Dataset S3). Foraging type A indicates feeding on animals with poor vision. Type B is characterized by concealed habitats (fossorial, litter) or specialization to feed on hiding prey. Type C indicates nocturnal, active foraging in open habitats. Type D indicates diurnal, active foraging in open habitats. Type E indicates sit-and-wait foraging in open habitats. We presumed that FC cost increases from type A to E. Based on the phenotypic dataset and time-calibrated trees (SI Appendix, Fig. S1), we investigate the evolution of warning coloration and mimicry and its association with foraging ecology.

Warning Coloration in Elapidae.

Our ancestral state reconstruction (ASR) indicated that bold coloration has evolved 27 times independently, and losses of such coloration have occurred six times (Fig. 2A). The Bayes factor test and CCT indicated broad correlations between coloration and ecological traits (Table 1, Fig. 2A, and SI Appendix, Figs. S2–S5). Specifically, both or either test(s) indicated that bold coloration correlates positively with active foraging, fossorial and aquatic habitats, nocturnal activity, and feeding on snakes, on wormlizards/caecilians, and on fishes, while it correlates negatively with terrestrial/arboreal habitat, diets of general vertebrates/anurans, and mammals/birds. Both of the tests provided enormous support for the correlation between bold coloration and foraging types A+B or A–C. As a result of our regression analysis using Bayesian generalized linear mixed models (GLMMs) explaining coloration, the model that has foraging type as an ordinal explanatory variable (M2) yielded much larger marginal likelihood than the null model (M0) (BF = 49.8, SI Appendix, Table S3), indicating that foraging type, an index of FC cost, predicts coloration. The probability of having bold coloration is highest for the foraging type A and decreases toward type E (Fig. 3A).

Fig. 2.

The evolution of foraging and coloration in (A) Elapidae, (B) Dipsadidae, and (C) a detailed view of Xenodon. The branch color indicates reconstructed ancestral character state. A dotted line indicates that the state is uncertain (<60% probability). The snake pictures were illustrated by Kanon Tanaka.

Table 1.

Summary results of the Bayes factor test and CCT for correlated evolution

Trait1	Trait2	Bayes factor test, Log Bayes factor	CCT, P-value	Direction
Elapidae
Bold coloration	Foraging mode (active)	6.5	0.0356	+
Bold coloration	Habitat (terrestrial/arboreal)	15.6	0.0025	–
Bold coloration	Habitat (fossorial)	5.7	0.5211	+
Bold coloration	Habitat (aquatic)	4.0	0.0028	+
Bold coloration	Activity (nocturnal)	3.7	0.0025	+
Bold coloration	Diet (general/anurans)	11.8	0.0035	–
Bold coloration	Diet (fishes)	10.2	0.0022	+
Bold coloration	Diet (snakes)	20.0	0.1369	+
Bold coloration	Diet (mammals/birds)	4.9	0.4357	–
Bold coloration	Diet (wormlizards/caecilians)	16.0	0.5324	+
Bold coloration	Foraging type (A+B)	20.8	0.0058	+
Bold coloration	Foraging type (A–C)	21.6	0.0001	+
Dipsadidae
Bold coloration	Foraging mode (active)	2.3	0.3850	+
Bold coloration	Habitat (terrestrial/arboreal)	–1.5	0.0259	–
Bold coloration	Habitat (fossorial)	–0.4	0.0076	+
Bold coloration	Activity (nocturnal)	0.4	0.0263	+
Bold coloration	Diet (invertebrates)	5.7	0.0008	+
Bold coloration	Diet (anurans)	–5.9	0.0005	–
Bold coloration	Diet (wormlizards)	2.2	0.2673	+
Bold coloration	Diet (snakes/wormlizards)	3.3	0.7056	+
Bold coloration	Foraging type (A+B)	–0.4	0.0000	+
Bold coloration	Foraging type (A–C)	4.0	0.0001	+
Pitviper mimicry	Habitat (terrestrial/arboreal)	6.6	0.0988	+
Pitviper mimicry	Habitat (fossorial)	2.5	0.1920	–

Only correlations supported by both or either test(s) (Log Bayes factor > 2.0, P < 0.5) are shown here (see SI Appendix, Tables S1 and S2 for full results).

Fig. 3.

Probabilities of evolution of bold coloration for elapids (A) and dipsadids (B) and pitviper mimicry (C) by foraging type estimated by regression analysis. Odds ratios were calculated from the coefficient estimates of the largest marginal likelihood model (M2). The coloring corresponds with Fig. 2. A box, a bold line, and bars denote the interquartile range, median, and the range excluding outliers, respectively. The points below the first quartile – 1.5× interquartile range or above the third quartile + 1.5× interquartile range were regarded as outliers.

Analysis using the dependent model implemented in BayesTraits (18) indicated that warning coloration was 15.7 times more likely to evolve when given foraging types A+B (low FC cost) than when given foraging types C–E (high or moderate FC cost) (Fig. 4A). Conversely, losses of bold coloration were 8.0 times more likely when given high or moderate FC cost. Gains of bold coloration were 8.6 times more likely than losses when given low FC cost, whereas losses of bold coloration were 14.6 times more likely than gains when given high or moderate FC cost. Therefore, color transitions are rather one-directional, where the direction corresponds to their foraging type. The same patterns were found when foraging type C was grouped with A+B instead of with D+E (SI Appendix, Fig. S6). The results with this grouping also indicated that coloration had a 2.9-fold effect on the probability of gains of foraging type A–C (low or moderate FC cost), indicating that coloration may cause the evolution of ecology.

Fig. 4.

Probabilities of evolutionary transitions among the four combinations of states resulting from two binary variables (coloration and foraging type). A numeral by an arrow indicates the rate coefficient estimated by fitting the dependent model implemented in BayesTraits. Arrow thickness corresponds to the value.

Mimicry of Conspicuous Models in Dipsadidae.

Our ASR indicated 23 gains and seven losses of bold coloration (putative mimicry of coralsnakes, Fig. 2B). The Bayes factor test and/or CCT indicated positive correlations between bold coloration and active foraging, fossorial habitat, nocturnal activity, and feeding on invertebrates, on snakes/wormlizards, and on wormlizards, and negative correlations between bold coloration and terrestrial/arboreal habitats and anuran diet (Table 1). CCT strongly supports correlation between bold coloration and foraging types A+B or A–C, while the Bayes factor test did not do so (Table 1). In the regression analysis, none of the models received strong support, but M2 yielded the largest marginal likelihood (SI Appendix, Table S3). The coefficient estimates indicated that the probability of bold coloration decreases from foraging type A to E, replicating the trend in elapids (Fig. 3).

Dipsadids were also remarkably similar to elapids in the patterns of correlated evolution between bold coloration and foraging type shown in Fig. 4. Gains of bold coloration were 6.4 times more likely when given foraging type A+B (Fig. 3B). A notable difference from elapids was that dipsadids were more likely to lose bold coloration with the condition of foraging type A+B. Also, the probability of transitions to bold coloration was relatively low even with low FC cost. Thus, transitions between cryptic and bold coloration were bidirectional even when FC cost is alleviated.

Mimicry of Inconspicuous Models in Dipsadidae.

Our ASR indicated that pitviper mimicry has independently evolved seven times (Fig. 2B). The foraging types at the nodes of mimicry gains were estimated to have been type A (two lineages of snail/slug specialists), type C (one anuran-feeding species), type D (three lineages of lizard/anuran-feeders), or type E (one lineage of snakes which ambush lizards and other vertebrates). Losses of pitviper mimicry were indicated to have occurred four times, and these were invariably transitions to bold coloration. In contrast to bold coloration, pitviper mimicry was not correlated with most of the ecological traits tested here. Nonetheless, it correlated positively with terrestrial/arboreal habitats and negatively with fossoriality (Table 1). Our regression analysis indicated that the probability of pitviper mimicry decreased sequentially from foraging type E toward B, which is an opposite trend from bold coloration, but increased again for type A (Fig. 3C).

Fig. 4 also illustrates that pitviper mimicry is distinct from bold coloration in the patterns of correlated evolution. Important differences include that losses of pitviper mimicry were much more likely than gains, by 6.5-fold, for foraging type A+B. This may be related to the tendency of transitioning to bold coloration when given such a foraging strategy. Such putative model shifts were seen in two lineages, including Xenodon (Fig. 2B). Pitviper mimicry in Xenodon was indicated to have evolved once in their most recent common ancestor (Fig. 2C), which was estimated to be diurnal (99%), terrestrial (99%), anuran-eating (84%), and an active forager (100%) having type D foraging (93%). Subsequently, type B foraging (burrowing) has evolved at node 2 or 3 in Fig. 2C. The extant species of this lineage are psammophilic, have up-turned nasal scales, and may dig up food, including reptile eggs (19). This ecological transition was followed by the shift to bold coloration at node 4 or 5 (Fig. 2C).

We examined the effects of phylogenetic uncertainty by independently running analyses using alternative trees. The Bayes factor test, CCT, and regression analysis using a previous caenophidian tree (20) and the regression analysis using polytomic trees, where low-support nodes were collapsed, recovered very similar macroevolutionary patterns as shown above (SI Appendix, Fig. S8 and Tables S4–S6), indicating that the differences in phylogenetic trees had only limited effects on the results. Also, we examined the robustness of the regression analysis by identifying outlier species, which have the greatest impact on the results, based on the CPO statistic. Outlier species tended to be concentrated in certain genera, namely, Hydrophis for Elapidae and Leptodeira and Siphlophis for Dipsadidae. Analysis excluding these genera provided similar results, indicating that the patterns do not originate solely from a particular lineage.

Discussion

There is growing evidence for coloration-ecology correlations in snakes, implying that coloration and ecology have been integrated via some adaptive mechanisms (e.g., refs. 21–25). We found a broad and consistent correlation between bold coloration and foraging ecology for two snake radiations where bold coloration is widespread. For elapids, the evolution of bold coloration (warning coloration) was very well predicted by their foraging ecology. For dipsadids, we found notably similar patterns, but the evolution of bold coloration (mimicry of coralsnakes) was evidently more difficult to predict. Pitviper mimicry was radically different from coralsnake mimicry in terms of its association with ecology. Our results are also indicative of cause-consequence relationships of coloration and ecology. In the following discussion, we first interpret the observed patterns and then discuss implications of the findings.

We found strong correlation between bold coloration and diets both for elapids and dipsadids. As we predicted, bold coloration was negatively correlated with feeding on animals with acute vision and escape ability, such as anurans, and on vertebrates in general, whereas it was positively correlated with specialized diets of snakes and wormlizards. Many snakes rely primarily on chemical senses (26), and wormlizards are practically blind. Also, feeding on these limbless elongated reptiles is often related to fossorial habits. Although snakes as a group occupy various habitats, many bold-colored, ophiophagous snakes (e.g., Vermicella, Calliophis, and some Micrurus) feed predominantly on fossorial snakes and other specialized burrowers, such as wormlizards, limbless lizards, and caecilians (22, 27, 28). For elapids, another clear trend was the correlation between fish diet and bold coloration. This may be partly attributable to the fact that turbid water makes coloration of hunting snakes less visible to their fish prey. Seasnakes often inhabit clear seas, but these snakes generally prey upon hiding or sedentary prey and many species forage in a burrow or crevice (29, 30). For dipsadids, bold coloration was strongly correlated with specialized diets of invertebrates (mostly earthworms or gastropods). These correlations invariably agree with our prediction and provide evidence that warning coloration is associated with foraging ecology that mitigates the cost of being conspicuous.

Notably, bold coloration was positively correlated with nocturnality and fossoriality. Predator-induced selection is expected to produce a correlation between warning coloration and diurnality (31) and open habitat (32). This is because under such conditions, alternative strategies (e.g., camouflage) are relatively difficult to employ, prey animals are more likely to encounter diurnal, visually hunting predators (intended signal receivers), and color signals are transmitted more effectively (31–33). On the other hand, prey-induced selection predicts a correlation between warning coloration and nocturnality and concealed habitat. This is because conspicuous coloration of predators is less visible to their prey under such conditions. Our results match the latter prediction, indicating that foraging costs are a better predictor of the evolution of warning coloration and conspicuous mimicry than benefits of predation avoidance, although both likely play an essential role.

Dipsadids differ from elapids in that the probability of bold coloration gains was relatively low and reversion to crypsis was likely even when given low FC cost. Concordantly, correlations between bold coloration and foraging ecology were relatively weak when compared to Elapidae. The important difference between the two families is that elapids are models but dipsadids are mimics. The benefits of Batesian mimicry depend on a number of factors, including the relative abundance of models and mimics, unprofitability of models, and the accuracy of signals (1, 5, 34, 35). We consider that the complexity of survival by Batesian mimicry may make the evolution of bold coloration intermittent and relatively difficult to predict.

Our results also exhibited recurrent evolution of pitviper mimicry in dipsadids. The benefit of strong signals is an essential assumption of warning coloration theory and has received considerable empirical support (1, 2). Then, why do some dipsadids mimic inconspicuous models? Among the seven lineages of pitviper mimics, two are gastropod specialists. Several authors have suggested that feeding adaptations for eating slugs and snails enhance behavioral pitviper mimicry (head triangulation), and resemblance in coloration may work in tandem with such behavior (e.g., ref. 36). Although the remaining four lineages were rather diverse in ecology, each has at least one characteristic that is negatively correlated with bold coloration, such as anuran diet, terrestrial habitat, diurnal activity, and/or sit-and-wait foraging. Also, pitviper mimicry was correlated with habitat in the opposite direction from bold coloration. These results suggest that pitviper mimicry might be an alternative strategy when coralsnake mimicry is unlikely. For example, pitviper mimicry in Xenodon is indicated to have evolved at their ancestor, which is estimated as foraging type D and thus unlikely to have bold coloration. Within this lineage, a dramatic shift to arboreal pitviper mimicry has occurred (15), but conspicuous coralsnake mimicry has not evolved prior to the transition to foraging type B. We propose that either morphological preadaptations or ecological trade-offs may result in the evolution of inconspicuous mimicry. Four independent shifts from pitviper mimicry to bold coloration (coralsnake mimicry) were also indicated. Since all of these are estimated to have occurred under conditions of foraging type A or B, mitigated foraging costs may be a requirement, and the benefit of strong signals might drive model shifts under such conditions.

The analysis with dependent models has made it possible to interpret cause-consequence relationships of correlated traits. Bold coloration was more likely to evolve when given low FC cost foraging, indicating that foraging causes the evolution of coloration. This is exemplified by Xenodon; the gain of bold coloration was preceded by the transition to foraging ecology with low FC cost. These results support our hypothesis that relaxed foraging costs are an important condition that facilitates the evolution of conspicuous anti-predator coloration. On the other hand, it was also shown that coloration may in turn affect the evolution of foraging. This implies that warning coloration may result in niche specialization. For example, conspicuous kraits (Bungarus) are strictly nocturnal, in contrast to more diverse diel activity in dorsally inconspicuous cobras (Naja). These results suggest that coevolution may have shaped a variety of traits, sometimes causing dramatic phenotypic changes.

The Bayes factor test and the CCT were sometimes inconsistent in supporting correlations. Our hypothesis predicts that changes to warning coloration will be concentrated on branches with low FC cost, but not necessarily on branches with a relevant condition for an individual ecological trait, because multiple factors are involved in determining FC cost. Therefore, it is not surprising that the CCT immensely supported the correlation between warning coloration and foraging type, but did not do so for all correlations between coloration and individual ecological traits. There is theoretical indication that the Bayes factor test tends to report a significant correlation even when the pattern is derived from a single or very few evolutionary events (37). Although many of the correlations tested here were supported by relatively numerous evolutionary changes (e.g., bold coloration has evolved >20 times), this argument may apply regarding rare ecological traits, such as snake and wormlizard diets. However, we found very similar patterns across the two families, including the correlations between bold coloration and snake and wormlizard diets, and therefore, we suggest that the correlations indicated by the Bayes factor test here likely have biological significance. More unexpected was that some correlations indicated by the CCT were not supported by the Bayes factor test. Notably, these cases were seen solely for dipsadids. In the mimetic snakes, the probability of bold coloration gains was relatively low and losses were likely even when FC cost is low. Such a lower predictability might be the reason why some correlations were not supported by the likelihood-based test.

Although the correlations shown here were fully consistent with our predictions, it is important to consider alternative explanations. In some animals that sequester chemical defense from their diet, warning coloration is associated with dietary specialization because diet determines the toxicity (1, 38). This is not the case for snakes, however, because they synthesize rather than sequester venom. Also, it is unlikely that snakes feeding on animals with poor vision are more unprofitable to their visually hunting predators (e.g., birds) because snake venoms are generally most effective against their prey animals (e.g., ref. 39). Another possible confounding factor is body size. However, body size itself can hardly explain the broad and complex patterns of correlations because body size was usually highly variable among species classified into the same category in our study. The possibility that other unknown lineage-specific factors (40) play an important role is weakened by the convergent patterns seen between the distantly related clades. We suggest that, until a plausible alternative explanation is provided, the trade-off between foraging and warning is regarded as the factor producing the macroevolutionary patterns shown here.

Overall, our results demonstrate a close association between warning coloration/mimicry and foraging ecology and provide evidence that the trade-off between foraging and predator avoidance is responsible for such macroevolutionary patterns. This indicates that prey are an important agent in the evolution of anti-predator coloration. Clearly, much remains to be studied in this area. Investigations into prey behavior are awaited since we know very little about how prey animals utilize predators’ coloration to avoid attack and how they are variable in this regard. Detailed ecological and physiological data, which allow quantitative analyses, and an understanding of the genetic basis of traits will also be needed to move the field forward. Our findings promote such studies by highlighting the role of prey in the evolution of predators’ coloration. We below discuss the implications of the present findings.

Evolution and Persistence.

Considerable attempts have been made to identify the factors responsible for the evolution of warning coloration (1, 33). Some ecological traits, including aggregation, facilitate initial evolution of warning coloration by making it easier for rare, conspicuous mutants to survive (41–44). Chemical defenses and signaling environments, which underpin the function of warning coloration, may also predict the evolution of warning coloration (33, 38). In addition to these, our results indicate that ecological costs may be a determining factor for the evolution of conspicuous coloration. Because animal coloration is important in many contexts, conspicuous coloration is expected to generate many trade-offs. An important difference from the factors related to initial evolution is that such ecological costs continue to have an effect even after avoidance learning is established. Therefore, those costs are expected to affect not only the initial evolution but also the persistence of warning coloration. Indeed, our results indicate that higher FC costs make reversion to crypsis more likely. Ecological costs presumably play more important roles in the evolution and persistence of conspicuous anti-predator coloration than previously thought.

Evolutionary Patterns and Selection Regime.

Previous evidence and our findings indicate that snakes largely differ from other animals in the evolution of conspicuous coloration. In butterflies, which feed on plants, reversion to crypsis is unlikely once Batesian mimicry has evolved (7–9). This is not surprising given that mimicry is usually more successful than crypsis (5, 34) and the accuracy of the deceptive signals improves as time goes on (5). Nonetheless, loss of mimetic coloration is frequent and widespread in snakes regardless of whether it is aposematic or mimetic, presenting an evolutionary paradox (10). Also, snakes’ bold coloration is positively correlated with fossorial habitat and nocturnal activity, whereas the opposite trend is seen for some lepidopteran insects and mammals (31–33). Our results indicate that higher FC costs promote losses of warning coloration in snakes. Also, the presence of foraging costs predicts the correlation between warning coloration/mimicry and secretive ecology. Thus, observed inconsistencies are much less confusing when additional selection factors are considered.

An important indication from previous studies is that warning coloration may result in niche expansion and accelerated phenotypic/ species diversification by freeing prey animals from constraints of crypsis (1). Empirical evidence is known in some animal lineages, including amphibians (45, 46) and mammals (47). On the contrary, our results suggest that warning coloration of snakes may result in niche specialization, emphasizing the need for caution when extrapolating findings from one system to another. Various animal lineages should fundamentally differ in the combination of relevant selection factors according to their ecology. Our findings indicate that considering selection regime is essential in understanding the evolution and consequences of warning coloration.

Form and Diversity.

The trade-offs between predator avoidance and foraging may also influence the form of warning coloration. Although this field remains largely unexplored, one of the few studies indicates that warning coloration of venomous black widows looks more conspicuous to their predators (birds) than to their prey (insects) (48). Our results showed that typical bold warning coloration is associated with specific foraging strategies, but it is apparent that other snakes lacking bold coloration have evolved various forms of warning signals.

Cobras (Naja) are typically terrestrial, active both by day and night, dietary generalists, and usually have inconspicuous dorsal coloration. However, when threatened, they raise their head and spread their hood, which makes the snakes absolutely distinctive and conspicuous. Some poisonous or mimetic anuran-eating snakes (e.g., some Rhabdophis, Thamnophis, and Erythrolamprus) have bright colors on the skin between the scales. When these snakes are attacked, they often inflate or flatten their bodies, and the bright colors become (more) visible. These switchable signals are an elegant solution to the conflicting demands of their ecologies (49), although “always-on” signals are expected to be most effective in warning predators (1). Venomous snakes of the family Viperidae face a unique trade-off. Vipers have a defensive quality that could be signaled to their potential predators but are specialized sit-and-wait predators that must hide from their prey while ambushing. There is empirical evidence that the cryptic but distinctive patterns of some vipers serve as warning coloration (50, 51). These examples illustrate that snakes have evolved diverse warning signals, which is unexpected solely from the view of efficacy in warning predators. Thus, foraging costs affect the form and diversity of warning signals, rather than simply constraining their evolution.

Conclusion

The vast majority of studies on warning coloration have focused on predators as the selective agent. Relatively recently, studies focusing on additional selection factors, such as sexual selection and thermoregulation, have provided important findings. Here, we present macroevolutionary evidence that foraging (interactions with their prey) is an important factor in understanding the diversity of the evolution of warning coloration and mimicry in snakes. Our results highlight the significance of trade-offs between predation avoidance and other ecological demands in understanding the evolution, ecology, and diversity of warning coloration. A comprehensive examination of the selection pressures acting on coloration will provide useful insights into when and how warning coloration has evolved and diversified in various animal lineages.

Materials and Methods

Taxon Sampling.

We based our taxon sampling on ref. 52. We added one species, Xenodon rabdocephalus, to our dataset to cover all species in that genus. Four dipsadid species, Philodryas chamissonis, Philodryas trilineata, Tachymenis chilensis, and Conophis vittatus, were excluded for presumable misidentification (20) or being rogue taxa in phylogenetic inference. An outgroup was included in phylogenetic comparison for elapids (Buhoma depressiceps), but not for dipsadids because its sister clade is uncertain. Consequently, the subjects of our phylogenetic comparison were 194 species of elapids+Buhoma and 238 species of dipsadids. For phylogenetic inference we used their respective outgroups, Atractaspis bibronii and Sibynophis chinensis; these were pruned from the resultant trees before further analyses.

Data Compiling and Categorization.

We compiled phenotypic data (coloration, foraging mode, activity time, habitat, and diets) for the target species through extensive surveys of published articles and books. We defined “bold coloration” as a distinct pattern with the combination of red/yellow/white and black or uniform red. We considered species to have bold coloration if they possess these patterns throughout most of their geographical range and over most life stages. For the purpose of this study, species with bright colors only on the ventral side or the skin between the scales (i.e., “switchable” aposematism, 49) were not regarded as having bold coloration.

Foraging mode was classified into active foraging, ambushing, and both strategies. Activity time was categorized into nocturnal, diurnal, and both (all-time). Habitats were categorized into terrestrial/arboreal, fossorial/litter-dwelling, and aquatic. For the purpose of this study, we attempted to search for foraging habitats rather than sheltering habitats. For diets, we constructed a food composition dataset (Dataset S4) based on published data on the number of prey items from stomach content surveys, direct observation, and book descriptions. Diet composition was clustered into eight major groups by divisive hierarchical cluster analysis in the package cluster (53) in R (54). We were able to obtain data on foraging mode, activity time, habitat, and diets for 91%, 77%, 100%, and 97% of elapid and 78%, 89%, 99%, and 92% of dipsadid species, respectively. We categorized the species into one of the five foraging types. Those species that feed predominantly on animals with no/poor vision, specifically, invertebrates lacking image-forming eyes, such as annelids, gastropods, and onychophorans, vertebrates with highly reduced eyes, such as scolecophidians, amphisbaenians, caecilians, and Gymnotiformes, and eggs, were coded as Type A (low FC cost regardless of when/where/how they forage). The remaining snake species, which feed predominantly on animals with more considerable visual and escaping abilities, were classified based on habitat, behavior, and/or activity time: Type B, foraging in concealed habitat (fossorial, leaf litter, and murky water) or specialized for preying upon hiding animals, such as most seasnakes (their prey is unlikely to see their coloration); Type C, actively forage in exposed habitats during the night (relaxed FC cost due to dim-light conditions); Type D, actively forage in exposed habitats during the day (high FC cost); Type E, sit-and-wait foraging in exposed habitats (high FC cost).

Phylogenetic Tree Reconstruction.

We reconstructed phylogenetic trees separately for Elapidae and Dipsadidae. DNA sequences of five mitochondrial and five nuclear regions were obtained from GenBank (Dataset S1) and aligned with MAFFT 7.487 (55). The best-fit nucleotide substitution model was selected based on BIC for each sequence region with ModelTest-NG 0.1.7 (56). We performed partitioned maximum likelihood analysis and 2,000 bootstrap sampling using RaxML-NG 1.1 (57). The resultant trees (SI Appendix, Fig. S2) were time-calibrated with the mcmctree program in PAML 4.9j (58). Fossil calibrations were based on the refs. 20 and 59 and specified as a skew-normal distribution (ω = 0.15, α = 20). The minimum node age was constrained as follows: Elapidae stem, 24.9 Mya; Naja stem, 17.0 Mya; Bungarus stem, 10.215 Mya; Oxyuranine stem, 10.0 Mya; Dipsadidae stem, 12.5 Mya; the divergence of Heterodon and Farancia, 12.08 Mya. These values were used as the location parameter (ε) of the distribution. The maximum age for the stem of Elapidae and Dipsadidae was constrained to 54 Mya. We carried out Bayesian divergence time estimation by running Markov chain Monte Carlo (MCMC) samplings for 10 million generations. Parameters were sampled every 1,000 generations and the first 20% was discarded as a burn-in. We also constructed polytomic trees from the bootstrap tree sets (see above) by collapsing low-support (bootstrap value < 50) nodes using the “consense” option implemented in RaxML-NG.

ASR.

Terminal taxa in the phylogenetic trees were coded with discrete trait states according to the phenotypic data. Bayesian MCMC sampling of ancestral states was carried out with BayesTraits 3.0.5. We obtained the probability of character states for each node by averaging the estimates of 7,000 samplings resulting from 10 million generations of reversible-jump MCMC with the first three million discarded as burn-in. We used a hyper-prior approach, seeding the mean of the exponential prior from a uniform distribution on the interval 0 to 100. When the state on a node was uncertain (the probability < 0.6), the state was inferred following the parsimony or majority rule (when equally parsimonious).

Bayes Factor Test for Correlated Evolution.

We tested for correlated evolution between coloration and ecological traits (diet, habitat, activity time, foraging mode, and foraging type) by comparing the marginal likelihood of two models; the independent model, in which two binary traits evolve independently of each other, and the dependent model, in which two binary traits evolve correlated with each other along given phylogenetic trees (17). For this analysis, we transformed the categorical data into binary data by combining categories. The independent and dependent models implemented in BayesTraits 3.0.5 were fitted to the biological data with reversible-jump MCMC sampling. The number of sampling/burn-in generations and prior distribution were the same as the ASR. The marginal likelihood was estimated by stepping stone sampling with 100 stones, each of them running for 10,000 generations. We calculated the log Bayes factors from the marginal likelihood of the two models. We interpreted the support for correlated evolution based on the log Bayes factor: <2, weak evidence; >2, positive evidence; 5 to 10, strong evidence; >10, very strong evidence (60). The direction of correlations was determined based on the phenotypic data (SI Appendix, Fig. S7). We also compared the patterns of correlated evolution among elapid bold coloration, dipsadid bold coloration, and dipsadid pitviper mimicry based on the probability of transitions. We fitted the data on coloration and foraging type to the dependent model in BayesTraits 3.0.5 with 10 million generations of MCMC. Number of burn-in generations, prior distribution, and sampling methods were the same as above.

CCT.

We tested whether changes in coloration were concentrated on branches of clades with a particular state in ecological traits more or less than expected by chance using CCT implemented in MacClade 4.0. We ran 10,000 simulations to generate randomly scattered changes on the phylogenetic trees and obtain the distribution of the number of changes situated on the focal branches as predicted by the null hypothesis. We examined the probability of obtaining the observed numbers of changes on the branches with particular ecological conditions under the null hypothesis. Mapping of ancestral states and counting of the number of changes were based on the results of Bayesian ASR.

Regression Analysis.

We constructed three Bayesian GLMMs, M0–2, which explain coloration by phylogenetic factor and foraging type. M0 is the null model, which has no explanatory variables except for a random variable which adjusts phylogenetic “closeness” between species (61). M1 and M2 both have the five-level foraging type as an explanatory variable in addition to the random variable. M1 uses foraging type as a continuous variable, which is a typical approach to deal with an ordinal explanatory variable (62, 63). M2 is a penalized regression model (64), where foraging type was used as an ordinal explanatory variable. Each category of the foraging type was treated as a dummy variable, but the regression coefficients are assumed to be similar between adjunct levels (64). As the probability distribution of the objective variable, we used Bernoulli distribution and a categorical distribution with three categories for Elapidae and Dipsadidae, respectively. We used the logit link function. We fitted the models using MCMC methods implemented in an R package rstan version 2.21 (65). The iteration of MCMC was 20,000, where the first 5,000 samples were discarded as a warmup. We ran four independent Markov chains, and random samples from posterior distributions were obtained for each model. The approximated marginal likelihood was calculated using the bridge sampling algorithm of an R package bridgesampling version 1.1 (66). We conducted a sensitivity analysis for these regression analyses using CPO statistic (67, 68), which evaluates, for each sample i = 1, 2, …, N, how the posterior distribution changes if the i-th sample was removed from the dataset.

Supplementary Material

Appendix 01 (PDF)

Dataset S01 (XLSX)

Dataset S02 (XLSX)

Dataset S03 (XLSX)

Dataset S04 (XLSX)

Dataset S05 (XLSX)

Data, Materials, and Software Availability

All data and scripts for analyses are available from Dryad (https://doi.org/10.5061/dryad.fqz612jzn) (69).