Relative size matters: eyespots on large insect prey deter small arthropod predators

Abstract

Circular chromatic patterns that appear to resemble vertebrate eyes (‘eyespots’) are commonplace in the animal kingdom and are widely believed to have evolved as an anti-predator defence. For example, experiments have shown that eyespots on caterpillar-like pastry baits can deter predation by birds. However, little is known about the extent to which eyespots deter (or promote) attack by arthropod predators. Here, we describe two separate experiments in which salticid spiders (Salticus scenicus) and Chinese mantids (Tenodera sinensis) were presented with a choice of mealworms (Tenibrio molitor) with or without eyespots. In a complementary experiment, we observed the time taken for adult Chinese mantids to attack hawkmoth (Manduca quinquemaculata) larvae of two different sizes, with and without eyespots. All three experiments indicate that eyespots on insect larvae can deter predation, so long as the larvae are sufficiently large compared with the size of the arthropod predator. However, when larvae are small relative to the arthropod predator, eyespots cease to be protective and may even promote attacks. Our results suggest that small arthropods can show an aversion to large prey with eyespots and help explain the presence of eyespots in medium-sized caterpillars, because these traits are unlikely to deter avian predators.

1. Introduction

Many animal species have evolved conspicuous circular chromatic patterns that appear to resemble vertebrate eyes [1–6]. There are two broad explanations as to why natural selection has favoured the evolution of these ‘eyespots’, and both assume that it is mediated by predators. In some taxa, the position of the false eyes may cause a predator to attack a non-vital body part, giving it a chance to escape [7–9]. Alternatively, or in addition, eyespots may deter predation through some form of intimidation [10–12] in which the eyespots enhance the perceived risk of harm to the predator. Although eyespots have been postulated to deter predation simply through their saliency [13–15], behavioural experiments suggest that their resemblance to eyes at least partly accounts for their anti-predation benefit [16–19].

While eyespots are commonly associated with insect wings, they also occur in the immature stages of insects and have been particularly well documented in the caterpillars of Lepidoptera (butterflies and moths) [20]. The eyespots observed on these caterpillars have long been believed to prevent predation via intimidation, rather than by deflecting attacks to non-vital body parts [5], not least because they frequently occur on or near the head (a vital body part) [21]. Field data using artificial caterpillar baits have provided empirical evidence that caterpillars with eyespots do indeed experience lower predation rates compared with similar baits without them [22,23]. Caterpillars with eyespots are often palatable to predators, so it appears unlikely that predators would have learned to avoid attacking eyespotted caterpillars through previous bad experiences with these prey [24,25]. Rather, predators may have an innate aversion to objects with eye-like markings. However, if eyespots provide protection against predators, one might wonder why have they been selected for some species but not others? A recent phylogenetic study of Sphingid caterpillars found that there was an association between the final instar body size of a species and the presence of eyespots [26]. This cross-species association is also consistent with intra-specific comparisons, which suggest that prominent eyespots are sometimes not evident until later instars when the caterpillar has achieved relatively large body size [27,28]. More generally, insect species with hidden contrasting hindwings that are exposed when fleeing and/or on approach by predators are also more likely to be larger [29,30].

Two non-mutually exclusive hypotheses have been proposed to explain the observed relationship between eyespots and body size in immature Lepidoptera. The first hypothesis is that large body size results in greater selective pressure for defence, because such prey are not only more conspicuous [31] but they also offer a greater caloric reward to would-be predators. The second hypothesis is that eyespots (and startle signals in general) are more effective on larger insects because large organisms may be perceived by predators as more threatening. Behavioural experiments using avian predators, both in the lab and in the field, support the second size-based hypothesis [26,29]; indeed, there is some evidence that eyespots on small prey may actually attract avian predators [26]. The size-dependent deterrent effect of eyespots is likely not mediated by the absolute size of the prey, but rather by the prey’s size relative to the predator. As such, eyespots on small prey may be an effective deterrent, but only to relatively small predators.

While there is both a clear trend of eyespots being associated with large and late instar caterpillars, and evidence that vertebrate predators are wary of approaching large prey with eyespots, eyespots are nevertheless observed in some small Lepidoptera species and early instars (see electronic supplementary materials, Figures S1, S2 and table S1, for seven Lepidoptera species that have eyespots at early instars). For these small-sized caterpillars, arthropod predators may serve as important agents of selection on eyespots. Arthropod predators—ranging from spiders to asilid flies—are widespread, diverse, voracious insectivores and they often outnumber birds in a given community by an order of magnitude [31,32]. The extent to which arthropod predators can select for anti-predation signals in other arthropods has been the subject of growing research [33–44]. Most of this work has focused on how arthropod predators might select for aposematic and mimetic signals. Considerably less work has focused on how arthropod predators may react to potential intimidatory signals, such as eyespots in caterpillars [9,45]. Janzen et al. [20] has argued that avian predators are likely to have an unlearned response to these eye-like stimuli and that this may go some way to promoting their effectiveness. As the authors put it, ‘The bird that must learn to avoid what appears to be a predator’s eye is not long for this world’. There is clear evidence that some vertebrates do indeed have an unlearned response to eye-like stimuli [26]. However, while this logic should extend to visual arthropod predators, to our knowledge, this has not been directly tested.

2. Methods

Here, we performed a complementary series of experiments to gauge the reaction of arthropods to insect larvae that had been manipulated to have or not have eyespots. To evaluate the generality of our insights, two visual predator species from two separate orders of arthropod were chosen (a mantid and a spider). The primary aim of these experiments was to quantify the extent to which eyespot markings on larvae deterred or promoted attack. Our first two experiments varied the size of the predator relative to a fixed size of the prey, while the third experiment varied the size of the prey relative to a fixed size of predator.

Our electronic supplementary material details how the experimental insect prey, namely, mealworms (Tenibrio molitor) and five-spotted hawkmoth (Manduca quinquemaculata) caterpillars, were acquired and reared, as well as how they were painted. Experiment 1 investigated the preferences of zebra jumping spiders (Salticus scenicus) of different sizes, for mealworms of a fixed size that were with and without eyespots. Experiment 2 followed a similar approach by investigating the preferences of Chinese mantids (Tenodera sinensis) of different instars for mealworms of a fixed size that were with and without eyespots. In experiment 3, we continued our investigation of Chinese mantid behaviour, by determining the latency of adult mantids to attack smaller or larger caterpillars with and without eyespots.

(a). Experiment 1

Zebra jumping spiders (S. scenicus) were collected from the brick walls of local buildings in Ottawa, Ontario, Canada. The total body length (cephalothorax and abdomen) of each spider at the time of collection was based on a digital photograph of the specimen (see electronic supplementary material for more details). All experimental trials involved presenting two mealworm prey (~10 mm in length) to a jumping spider in a one-off choice test. One of the mealworms was modified to have eyespots of contrasting colours and one did not (electronic supplementary material). Presentations took place in a small plastic tapered cylindrical container (0.12 l, base diameter: 50 mm, lid diameter: 76 mm). Our response variable was the prey type that was attacked first within the observation period. Each spider and mealworm prey were used once.

During the test phase, the cups containing the animal subjects were placed on brown cardboard, and cardboard barriers were placed between the cups to prevent the S. scenicus from seeing one another. The room’s temperature was set at 22.5°C during the day and 18.5°C at night; ambient light entered the room from two windows. The initiation of the trial involved the simultaneous presentation of both mealworms to a spider. The mealworms were placed in the cup and were not tethered, allowing them to roam freely within the cup. Once the trial was initiated, the jumping spiders were checked three times per day (at 7 a.m., 12 noon and 7 p.m. local time) for attacks on the mealworms. The treatment group of the first-attacked mealworm was recorded, and the trial was terminated at this time. Given the relatively large size of the mealworms compared with the spiders, there was no instance of both mealworms being attacked during the inter-observation period.

(b). Experiment 2

This experiment followed a similar format in that mealworms with and without painted eyespots were presented to Chinese mantid (T. sinensis) nymphs. In this experiment, the mantids were classed discretely by instar (L3–L6 instars inclusive, size range from 25 mm for L3 mantids to 60 mm for L6 mantids) rather than continuously by size (although owing to growth later instars are larger). To reflect the larger size of mantids compared with the spiders, the experimental mealworms were somewhat larger (~20 mm in length) and consequently had slightly larger eyespots (figure 1). During presentations, the mealworms were tethered 7 cm apart at the bottom of the presentation arena (0.95 l cup) to ensure equal accessibility (see electronic supplementary material for details of how the mealworms were tethered).

Figure 1.

Top: Mealworms (T. molitor) manipulated to have eyespots or no eyespots, along with three zebra jumping spiders, S. scenicus, demonstrating the wide size variation of predators used in experiment 1. Scale bar = 10 mm. Middle: Mealworm T. molitor manipulated to have eyespots or no eyespots, along with Chinese mantids T. sinensis instars 3−6 used in experiment 2. Scale bar = 20 mm. Bottom: Manipulated live five-spotted hawkmoth M. quinquemaculata larvae used in experiment 3 (from left to right: small no eyespots, large no eyespots, small with eyespots, large with eyespots). Scale bar = 20 mm.

The test phase took place indoors in a room with the temperature set to 22.5°C, with standardized ambient artificial lighting. The cups containing the tethered mealworms were held in cardboard cells to prevent mantids from seeing one another. The mantids were then transferred to the cups containing a tethered mealworm. This was accomplished by first encouraging them to perch on the screen of the lid of the original cup (where they had been housed prior to the experiment) by gently shaking the cup. Once the mantids were perched on the screen of the original cup’s lid, the lid was removed and transferred to the cup with the tethered mealworm. Transfer of the mantids was alternated such that the first mantid received a mealworm with eyespots on the right, while the second received a mealworm with eyespots on the left, and so on. Trials began at 10 a.m. and the mantids were checked every 20 min until 6 p.m., when the trials were concluded. Each mantid (20 of each instar) and the each of the two types of painted mealworms were used only once.

(c). Experiment 3

To explore outcomes when size differences are caused by prey, we used adult Chinese mantids with body lengths of 90–100 mm as model predators and sequentially presented each of them with a single five-spotted hawkmoth (M. quinquemaculata) larva from one of two different size treatments. These larvae had been manipulated either to have or not have visible eyespots (electronic supplementary material). The trial lasted a single day. The order in which T. sinensis were presented with each type of caterpillar differed between blocks (electronic supplementary material, table S2,S3) and each block consisted of three individuals (1 male and 2 females). Thus, the attack times of 12 adult mantids were repeatedly evaluated. The response variable was recorded as the time taken for the mantid to attack the presented prey item.

In order to standardize their hunger prior to the trial, the mantids were fed darkling beetle (superworm) Zophobas morio larvae to satiation, 2 days before the trial. After this, they were not fed for a day before the trial. Experimental trials lasted 7 days (a total of 4 days of testing, alternated with a total of 3 days without food). On each test day, the adult mantids were first presented with a final instar larva of T. molitor that was not manipulated in any way (‘treatment 0’). The purpose of treatment 0 was to provide a measure of the mantises’ motivation to eat on that specific day, so that if latency changed over time then this potentially important source of variation could be adjusted for in the analysis. Shortly after the mantid had finished consuming this mealworm, a five-spotted hawkmoth M. quinquemaculata caterpillar was presented. These hawkmoth caterpillars were one of two length classes (small: 33−47 mm, large: 61−74 mm) and either had two visible eyespots (white circles containing a smaller black circle) or did not (circles of identical dimensions, this time painted with clear polish; figure 1). Given the repeated use of the same individual mantids, upon seizing a test hawkmoth, the prey item was taken from the T. sinensis to prevent the ingestion of polish. After the treatment prey was taken away (or the experiment ended without attack) the T. sinensis was fed Z. morio larvae to the point of satiation. In between each test day, to increase motivation, the T. sinensis mantids were not fed for a day.

Since the goal was to observe the mantids' response to the eyespotted and non-eyespotted prey without distraction by the researcher, the trial took place in a ‘blind’ constructed from grey corrugated plastic. The blind consisted of four sides, allowing light to enter from the top while preventing the T. sinensis from seeing the researcher. One researcher (K.L.H.) observed the trial and recorded the results in real-time from a webcam (Logitech HD 1080 p) that was placed inside the trial blind. The larva of a given size, with and without eyespots (or mealworm control, see above) was placed in a 0.95 l transparent plastic cup lined with filter paper. When the T. sinensis was resting on the screen perch in the original enclosure, the top of the original enclosure was placed over the cup containing the treatment. The time at which the top cup containing the mantid was placed over the bottom cup containing the treatment marked the beginning of a given test. The duration of each test was 15 min (900 s). This provided ample time for the mantids to exhibit a predatory response to the prey. The total time taken for T. sinensis to attack the prey offered was recorded and marked the end of the trial. If no prey were attacked within 15 min then the test was terminated. We also noted the time at which the T. sinensis’s head crossed a line that was 6 cm from its original position on the perch and in the direction of the prey (time to ‘first base’).

This experimental design not only allowed for a binary outcome (attacked or not attacked) to be recorded, but also the mantids' attack and approach latency. In contrast to experiments 1 and 2, prey size, rather than predator size, was modulated. When predator size is modulated, relative attack latency is not expected to be a very insightful metric of their behavioural reaction to eyespots. This is because predators of different sizes differ in locomotory capability, metabolism and relative motivation, and are therefore likely to differ in the amount of time taken to attack prey for reasons unrelated to the presence of eyespots. However, when the predators are of the same ontogenetic stage and the prey are of different sizes, attack latency can provide significant insight into the predator’s behavioural reaction to eyespots.

3. Statistical analysis

(a). Experiment 1

To visualize the responses of the jumping spiders to the two types of mealworm, we plotted how the probability density of spider responses (mealworm with eyespot attacked, mealworm with no eyespot attacked, no mealworm attacked) varied with spider size, using the geom_density function of ggplot2 [46], setting the running bin width as 0.5 mm. We then fitted a multinomial logistic model to the spider responses using the multinom function of the nnet [47] package, with attacking no mealworm as the reference outcome and spider size as a covariate. Relative risk (the ratio of the probability of one outcome happening compared with the probability of the reference) was quantified by exponentiating the coefficients of the logit model. The significance of coefficients (with the null hypothesis that they were zero) was evaluated based on evaluating the degree to which the (untransformed) scaled coefficients departed from a standardized normal distribution.

To further explore how the spiders’ decisions were related to their size, we fitted a binary logistic model to our data in which outcomes were classified as no mealworm attacked (1) or a mealworm (of any type) attacked (0). We then reduced our data to consider only those instances in which a mealworm was attacked (33 of 50 trials) and fitted a binary logistic model to elucidate whether the log odds of attacking a mealworm with eyespots varied with spider size. The fits of these models to the data were displayed using the logi.hist.plot function of the popbio package [48].

(b). Experiment 2

To test whether there was an overall association between the outcome of the choice test (neither mealworm attacked, mealworm with eyespot attacked, mealworm without eyespot attacked) and the instar of the mantid, we conducted a Pearson chi-squared test for independence. Given that the row totals for each instar were relatively low (so that the Pearson test statistic only approximates a chi-squared distribution if the null hypothesis of independence were true), we estimated statistical significance using the simulated p-value argument of the chisq.test function (although the results were similar either way). With simulated p-values the null distribution for the Pearson was determined by repeatedly generating datasets using the same totals, while assuming independence.

To investigate the distribution of the outcomes in more depth, we fitted a multinomial logistic regression model as above. Here, the instar of the mantid was treated as a covariate rather than a factor because of the lower Akaike Information Criterion (AIC) arising from the fit of the former model. Not attacking either mealworm type was again treated as the reference in this multinomial model. We then fitted a binary logistic model to determine whether the propensity to attack a mealworm of any type varied with the instar of the mantid, treating instar as a covariate owing to the lower AIC of this fitted model. Finally, we reduced the data to include only those choices where a mealworm was attacked (n = 64 cases from 80) and fitted a logistic model to test whether the log odds of attacking a given type of mealworm varied with instar. We used this fit to estimate the probability (with 95% confidence intervals (CI) based on the original logit scale, back-transformed to the response) that a mantid of any given instar would choose to attack a mealworm with eyespots.

(c). Experiment 3

The primary endpoint interest was the time taken for a mantid to attack the hawkmoth caterpillars of a given size class, with or without eyespots. Since attacks did not always occur within the 900 s observation period following release, some of the data were effectively (right) censored, so a survival (more generally, ‘time to event’) analysis was appropriate. Our primary predictors were caterpillar body size (large and small) and the presence/absence of eyespots (yes and no). We were interested in testing for an interaction between body size and the presence/absence of eyespots on the attack time. Given that individual mantids were presented with all four possible treatments in one of four possible orders, we also fitted a block term to represent treatment order, with individual mantid nested within blocks (since each mantid only experienced one treatment order). Block and individual mantid were considered as random effects. To fit the Cox proportional hazards model with fixed effects as well as nested random effects, we used the mixed-effects survival analysis package coxme [49]. We compared the fit of the full model to the time to event response (Surv~Size*Eyespot +(1|Block/Individual)) with a series of simpler models lacking the product interaction and main effects, maintaining the nested random effects and therefore consistent with the underlying design. Model comparison was via log likelihood ratio (LR) tests (to compare the goodness of fit of hierarchical models) and AIC. Hazard ratios (based on exponents of estimated model parameters) for having an eyespot were determined against its appropriate reference treatment (the same-sized caterpillar without eyespots).

We were interested in whether the mantids show any change in the time taken to attack prey over consecutive presentations—for example, as a consequence of them learning that the caterpillars are safe to attack or because their motivation changes as they age. Although we controlled for a block effect above, the most direct way to assess any change in hazard over day was to extract data on the attack times of mealworms that were presented before each experimental treatment. Here, the fitted model for the attack times of mealworms was simply Surv ~ Day +(1|Block/Individual) and the importance of Day (first treated as a factor, then as a covariate) was evaluated by comparing similar models with and without Day.

Finally, although our primary endpoint was time to attack, to obtain a more complete understanding of the nature of the attack sequence and whether eyespots attract the attention of mantids from a distance, we fitted the same mixed-effect survival models as above, with the times taken to reach ‘first base’ in much the same way as we analysed the time to attack.

4. Results

(a). Experiment 1

Figure 2 shows the density plot of the responses of spiders of given sizes towards the two types of mealworm. The sample data indicate that as the size of a spider increases, it is less likely to attack nothing during the experimental period and more likely to attack a mealworm with eyespots.

Figure 2.

Density plots of the responses (three possible outcomes) of zebra spiders of a given size when presented with a choice between two types of mealworms (with and without eyespots). The figure simply shows a non-parametric smoothed density estimate, with probability estimates based on a rolling bin width of 0.5 mm.

The fit of the overarching multinomial logistic model indicated that with every 1 mm increase in spider size the relative risk (probability ratio, compared with not attacking a mealworm) of attacking a mealworm with eyespots increased by a factor of 21.38 (p < 0.001), while the relative risk of attacking a mealworm without eyespots increased by a factor of 4.38 (p = 0.032).

The log odds of not attacking a mealworm (compared with attacking one) declined significantly with increasing spider size (slope −1.902, z = −2.91, p = 0.004) (electronic supplementary material, figure S3). On fitting the logistic model to the subset of instances in which a mealworm (of whatever type) was attacked, we found that the log odds of the spider attacking a mealworm with eyespots increased with increasing body size (slope = 2.123, z = 2.29, p = 0.022; figure 3). The smallest spider (3.1 mm) had an estimated 0.085 probability of choosing the mealworm with an eyespot rather than the mealworm without an eyespot (95% CI 0.014–0.375). In contrast, the largest spider (5.1 mm) had an estimated 0.867 probability of choosing the mealworm with an eyespot rather than the mealworm without an eyespot (95% CI 0.445−0.981).

Figure 3.

Fit of the logistic model to the 33/50 cases in which a mealworm was attacked by a spider. This model indicates that the probability of a spider choosing to attack a mealworm with eyespots (as opposed to a mealworm without eyespots) increases as the spider's size increases. Observations are shown as dots (jittered to avoid overlap). The fitted logistic model is shown in blue and the borders of the 95% CI for the fitted logistic model are depicted in red.

(b). Experiment 2

The chi-squared test revealed that there was an overall association between the outcome of the choice test and mantid instar (X₆² = 16.01, p = 0.011). Figure 4 shows that, in general, there was a greater tendency of earlier instar mantids to not attack any of the mealworms that were presented (45% of third instars did not attack a mealworm, while only 5% of sixth instars did not attack a mealworm). Likewise, there was a greater tendency of later instar mantids to attack mealworms with eyespots (15% of third instars attacked mealworms with eyespots, while 60% of sixth instar mantids did).

Figure 4.

The frequency of responses of each instar of mantid (n = 20 for each instar) when presented with the mealworms with and without eyespots.

The fit of the multinomial logistic model suggested that for each step up in instar, the relative risk of the mantid attacking a mealworm with eyespots increased by a factor of 3.4 (p < 0.001) compared with not attacking a mealworm. Similarly, the relative risk of a mantid attacking a mealworm without eyespots increased by a factor of 2.1 (p = 0.03) compared with not attacking a mealworm. Separate fits of the generalized linear (logistic) model indicated that the log odds of not attacking a mealworm (compared with attacking one) declined with increasing mantid instar (slope = −0.960, z = −2.983, p = 0.003). When only mealworms that were attacked were analysed, there was a significantly higher propensity of mantids to attack mealworms with eyespots compared with mealworms without eyespots the later the instar of the mantid (27.3% of the mealworms attacked by third instar mantids had eyespots, compared with 63.2% of the mealworms attacked by sixth instar mantids (slope = 0.499, z = 2.013, p = 0.044; see figure 5)). The fit of this model revealed that third instar mantids had an estimated 0.31 probability of attacking a mealworm with eyespots (95% CI 0.15−0.55), whereas a sixth instar mantid had an estimated 0.67 probability of attacking a mealworm with eyespots (95% CI 0.48−0.82).

Figure 5.

Fit of the logistic model to the 64/80 cases in which a mealworm was attacked by a mantid. The fitted model indicates that the probability of a mantid choosing to attack a mealworm with eyespots increases as its instar increases. Observations are shown as dots (jittered to avoid overlap), while the fitted logistic model is shown in blue and the borders of the 95% CI for the logistic model are depicted in red.

(c). Experiment 3

On fitting the proportional hazard model, there was no evidence that the survival distribution of mealworms changed over time whether days were treated as a factor (LR tests, factor: X²₃ = 2.963, p = 0.397) or a linear covariate (X²₁ = 0.429, p = 0.513).

Our primary questions were whether there was any evidence that the distribution of the time to attack differed between caterpillars of different sizes, and between caterpillars with and without eyespots. Comparing the full model containing interactions (AIC = 219.8) with the main effects model (AIC = 239.6), eyespots only model (AIC = 238.2), size only model (AIC = 237.7) and the intercept only model (AIC = 236.2), each containing the block and individual as random factors, it was clear that the full model with interactions most parsimoniously explained the variation in attack times (∆AIC from next-best fitting model >16). Deconstructing the full model using sequential log-LR tests (analysis of deviance) indicated that body size alone does not explain significant variability in attack times over and above the null model (X²₁ = 0.542, p = 0.469) and neither does the presence/absence of eyespots alone (X²₁ = 0.010, p = 0.920). However, the full model was a significantly better fit than both the null model (X²₃ = 21.985, p < 0.001) and the main effects model (X²₁ = 21.298, p < 0.001) and all models in between. Figure 6 shows the fitted survival plots. Having an eyespot reduces hazard in large caterpillars by a factor of 0.17 (95% CI 0.051−0.567), yet in contrast, having an eyespot increases hazard in small caterpillars by a factor of 6.59 (95% CI 2.46−17.70).

Figure 6.

Survival plot of time taken for adult mantids to attack the presented prey. Small caterpillars with eyespots were most readily attacked while large caterpillars with eyespots were least readily attacked.

Intriguingly, when we fitted a survival model based on the time distribution to ‘first base’, then a rather different pattern emerges from that of time to attack (electronic supplementary material, figure S4). Here, the main effects model had the lowest AIC compared with all possible component models. Having eyespots in large caterpillars increases the hazard by a factor of 10.79 (95% CI 3.77−30.84), while having eyespots in small caterpillars increases the hazard by a factor of 26.86 (95% CI 7.48−96.39).

5. Discussion

Eyespots have attracted attention from naturalists and evolutionary biologists for over 150 years [4,21,50]. Here, we have conducted some of the first experimental tests to evaluate the extent to which eyespots in insect larvae deter arthropod predators. One key difference between this study and the previous predation experiments that have tested the protective value of caterpillar eyespots is that we have utilized live insect larvae manipulated to have or not have eyespots, rather than artificial targets. While studies that utilize artificial targets have provided valuable insights regarding the nature of selection for/against eyespots, it is possible that movement significantly affects how predators respond to prey with eyespots, such that an entirely sedentary eyespotted artificial prey may elicit a very different response from a moving insect larva with eyespots.

The experiments indicate that eyespots may indeed have a selective benefit in deterring predation by arthropods. However, just as has been observed in experiments with avian predators [26], there is an indication that larvae with eyespots may be selected against when the prey item is small relative to the size of the predator, possibly because it draws attention to the prey without posing a threat. Specifically, in experiments 1 and 2 we found that eyespots on mealworms tended to deter smaller spiders and mantids from attack (significant for spiders and of borderline significance for mantids). Similarly, eyespots on mealworms tended to attract predation by larger spiders and mantids (both of borderline significance). Conversely, when prey size was varied in experiment 3, then eyespots on smaller larvae significantly attracted predators while eyespots on larger prey deterred them. Collectively, these three experiments show that the probability of eyespotted larvae being attacked decreases as their size increases relative to that of their predator.

The presence of eyespots in some small and early instar caterpillars (electronic supplementary material, figure S2) had previously been challenging to explain, since these traits may attract avian predators [26]. One possibility is that the eyespots may be the result of a developmental constraint, so that some species are unable to express eyespots at late instars without also having them during early instars. Alternatively, eyespots in small caterpillars may simply be a neutral trait retained from shared ancestry with larger species where they are beneficial. However, our findings suggest that predation pressure by arthropods, which are typically smaller than vertebrate predators, may contribute to their evolution, or at least mitigate some of the costs imposed by avian predators and very large arthropods. Thus, the critical body size in prey above which eyespots are beneficial may be lower if the larvae are also preyed upon by invertebrate predators of small size relative to their prey.

It is worth noting that Brose et al. [51] found that as the body size of insects increases, the probability that they will be attacked by vertebrate predators increases exponentially. Conversely, the probability that an insect will be predated by an arthropod predator increases with decreasing body size (see also [41]). Our experiments are consistent with these observations. Thus, in experiments 1 and 2 we found that the small arthropod predators rarely attacked any larvae, with the L3 mantids being on average nine times less likely to attack any prey than L6 mantids. Experiment 3 also showed that even adult mantids took longer to attack large caterpillars, with a relatively high proportion not being attacked at all.

Our finding that small S. scenicus spiders tend to avoid larvae with eyespots may well be an adaptive response to the perceived risk of injury. Although salticid spiders are capable of preying on insects much larger than themselves, such prey may be more difficult to subdue and have greater potential to harm the would-be predator. The presence of eyespots may promote this perception [5]. Additionally, high-contrast black and white markings may resemble the eyes of potentially predatory larger conspecifics, which notably have black corneal lenses with (white-looking) light reflections. A recent study suggests that salticid spiders are averse to attacking prey with high-contrast black and white markings similar to those of our eyespotted mealworms [36], possibly because the spiders associate conspicuous colouration with distastefulness. Intriguingly, Huang et al. [34] found that the relative body size of chemically defended ants (and their mimics) had no effect on their rates of attack by salticid spiders. While chemically defended prey may be avoided by spiders irrespective of relative body size, here we have found that only relatively large prey with eyespots are avoided, which suggests that, unlike chemical defences, the perceived costs varies with size.

Our data do not directly address whether the response of our arthropod predators is owing to eyespots resembling vertebrate eyes, or simply because of their conspicuousness and novelty. Indeed, even for avian predators, whose response to caterpillar eyespots has been extensively studied, ascertaining the mechanism through which eyespots prevent predation has proven difficult. A recent systematic review and meta-analysis of 33 experiments that presented eyespot-like stimuli to avian predators suggests that novelty and conspicuousness alone may be sufficient to explain the antipredation benefit of eyespots [15] (see also [13]). However, the authors stress that there is still substantial uncertainty regarding how predators perceive and respond to eyespots.

Previous studies have demonstrated that mantids are capable of learning to avoid prey with signals that indicate chemical defense or evasiveness [43,44]. However, our findings in experiments 2 and 3, that eyespots on insect larvae influence the predatory behaviour of mantids reared from eggs, indicate that these mantids have an innate response to these visual signals. As with experiment 1, our finding that this response is dependent on the relative size of predator and prey is consistent with the view that eyespots deter attack through some form of intimidation.

While eyespots in large caterpillars slow down the rate at which they are attacked, one might ask whether a delay in attack by a matter of seconds provides any survival advantage. In our experiments, the caterpillars were contained within a cup. However, under more natural conditions a delay in attack may give the larvae an opportunity to withdraw, escape (by dropping to the ground, for instance) or deploy a backup defence. Our finding that eyespots on relatively small Lepidoptera are harmful for survival is most likely because eyespots in these cases increase the conspicuousness of the caterpillars without appearing too threatening. Indeed, experiment 3’s data on the time to 'first base' support this conjecture (see electronic supplementary material, figure S4), in that eyespots in both small and large caterpillars attract the interest and approach of mantids, even if they are ultimately hesitant to attack the latter. These findings are consistent with the observation that eyespots on adult lepidoptera wings may attract attacks from mantids [9].

As with most experiments, there are some limitations. First, experiments 1 and 2 used larvae of Coleoptera rather than Lepidoptera. Therefore, it is possible that differences in behaviour or appearance caused the arthropod predators to react differently to eyespots on these larvae than they would to Lepidoptera larvae with eyespots. While we cannot rule out this possibility, it appears unlikely because the mantids in experiment 3 exhibited an analogous response to Sphingid (Lepidoptera) larvae with and without eyespots. Of course, Coleoptera are not evolutionarily irrelevant since eyespots are also seen in beetles, such as the adult stage of the eyed elater click beetles (Elateridae). Another consideration is the possibility that arthropod predators could see the markings on larvae painted with transparent polish, although spectral analysis and visual modelling suggest that this is unlikely (electronic supplementary material, figures S5,S4). Indeed, the fact that our predators reacted differently to larvae with and without eyespots strongly suggests that the predators discriminated between the two types of markings. While the large and small caterpillars in experiment 3 had proportionately scaled eyespots reflecting many Lepidoptera (electronic supplementary material, figure S1), it remains unclear if larger eyespots rather than larger body size might explain the benefit of eyespots experienced by our larger treatment group in this particular experiment. Therefore, while it is possible that eyespot size and body size may affect predator behaviour independently of one another, such effects are likely indistinguishable in nature. Finally, since both predator and prey size were modulated through the use of different developmental stages of the same species, we cannot rule out the possibility that ontogeny itself introduced some confounders. For instance, different instars of the same species may differ in metabolic rate, locomotion and even cognition. In particular, experiment 1 utilized wild-caught jumping spiders, so it is possible that the larger, older individuals may have been more likely to have had previous experience with eyespotted prey. However, our finding that the predator-to-prey size ratio determined the deterrent or attractant effect of eyespots, regardless of whether predator size (experiment 2) or prey size (experiment 3) was modulated, suggests that body size played the key role.

Given their abundance, diversity and capacity to learn, researchers have long speculated that arthropod predators could be a driver in shaping the anti-predator defences in other invertebrates [39–41,52]. Here, we provide evidence that eyespots in caterpillars may not only deter attacks by avian predators, but also by arthropod predators. This suggests that eyespots have broader protective benefits, and that the likelihood of their evolution will depend not only on the absolute size of the caterpillar but also on the relative size of their would-be predators.

Ethics

This work did not require ethical approval from a human subject or animal welfare committee. However, we sought to ensure appropriate housing and care for our arthropods. After experimentation, the mantids continued to be reared in a laboratory environment for the rest of their lives, while the spiders were released in the area where they were captured (see electronic supplementary material for more details).

Data accessibility

All experimental data are accessible via Dryad [53]. All R code to reproduce the analysis and figures is available on Zenodo [54].

Supplementary material is available online [55].

Declaration of AI use

We have not used AI-assisted technologies in creating this article.

Authors’ contributions

K.L.-H.: conceptualization, data curation, funding acquisition, investigation, methodology, project administration, writing—original draft, writing—review and editing; C.K.: formal analysis, methodology, validation, writing—review and editing; T.N.S.: formal analysis, supervision, visualization, writing—review and editing.

All authors gave final approval for publication and agreed to be held accountable for the work performed therein.

Conflict of interest declaration

We declare we have no competing interests.

Funding

K.L.-H. is funded by a Natural Sciences and Engineering Research Council of Canada (NSERC) postdoctoral fellowship, while T.N.S. is supported by an NSERC Discovery Grant. C.K. is supported by National Research Foundation of Korea (RS-2024-00333709 and RS-2024 00405751) and Creative-Pioneering Researchers Program through Seoul National University.