Transparency reduces predator detection in mimetic clearwing butterflies

1. Predation is an important selective pressure, and some prey have evolved con‐ spicuous warning signals that advertise unpalatability (i.e., aposematism) as an an‐ tipredator defence. Conspicuous colour patterns have been shown effective as warning signals, by promoting predator learning and memory. Unexpectedly, some butterfly species from the unpalatable tribe Ithomiini possess transparent wings, a feature rare on land but common in water, known to reduce predator detection. 2. We tested whether transparency of butterfly wings was associated with de‐ creased detectability by predators, by comparing four butterfly species exhibiting different degrees of transparency, ranging from fully opaque to largely transpar‐ ent. We tested our prediction using both wild birds and humans in behavioural experiments. Vision modelling predicted butterfly detectability to be similar for these two predator types. 3. In concordance with predictions, the most transparent species were almost never first and were detected less often than the opaque species by both birds and humans, suggesting that transparency enhances crypsis. However, humans were learn to better detect the more transparent species over time. 4. Our study demonstrates the first time that transparency on land likely de‐ creases detectability by visual predators.

Since transparency is often associated with crypsis, for example in aquatic organisms (Johnsen, 2014), transparency in these butterflies may decrease detectability by predators.
To determine whether transparency in clearwing butterflies decreases detectability by visual predators, we compared predator detection of four Ithomiini species that differ in the amount of transparency of their wings ( Figure 1): Hypothyris ninonia (largely opaque and brightly coloured), Ceratinia tutia (brightly coloured and translucent), Ithomia salapia (transparent with a pale yellow tint and an opaque contour) and Brevioleria seba (transparent without coloration other than a white band in the forewing and an opaque contour). Given the proportion of light that is transmitted through the butterfly wing of the different species (Supporting Information Figure S1), we predicted that the opaque species H. ninonia should be the easiest to detect, followed by the translucent species C. tutia. Finally, the more transparent butterfly species I. salapia and B. seba should be the least detectable. However, it is also possible that the coloured opaque elements of the transparent species, such as the white band in B. seba and the opaque contour found in most of these species, enhance detection. We tested our predictions using two complementary behavioural experiments involving birds and humans and further supported by a vision modelling approach.
Detectability of butterflies was first tested using wild great tits (Parus major) as model bird predators. Great tits are sensitive to UV wavelengths (UVS vision in Ödeen, Håstad, & Alström, 2011). Their vision is similar to that of naturally occurring Ithomiini predators such as the houtouc motmot (Momotus momota, Pinheiro, Medri, & Salcedo, 2008), the fawn-breasted tanager (Pipraeidea melanonota, Brown & Neto, 1976), or the rufous-tailed tanager (Ramphocelus carbo, Brower, Brower, & Collins, 1963). However, unlike Neotropical insectivorous birds, great tits are naïve to ithomiine butterflies and have not learned to associate their colour patterns to toxicity. As a bird's propensity to attack prey is the result of both prey detection and motivation to attack the prey, we also performed behavioural experiments using human participants, which can be useful in disentangling these two factors.
F I G U R E 1 Dorsal (top row) and ventral (bottom row) view of butterfly species used in the study (photographed against a black and a white background to show the location and degree of transparency in the wings). Wing transparency (transmission and area occupied by transparent patches) increases from left (most opaque) to right (most transparent): Hypothyris ninonia (largely opaque), Ceratinia tutia (translucent but brightly coloured), Ithomia salapia (transparent with a pale yellow tint and black wing contour), Brevioleria seba (transparent without coloration other than a white band in the forewing and a black wing contour). © Céline Houssin Differences in colour perception between great tits and humans include the presence of a fourth single cone type receptor (instead of three cones in humans) that extend the great tits' sensitivity into the UV light spectrum (Hart, 2001) and oil droplets that refine colour discrimination in birds (Vorobyev, 2003). However, neither humans nor birds are able to detect linear polarization, which excludes the use of polarization cues to detect and discriminate between butterfly species (Foster et al., 2018;Greenwood, Smith, Church, & Partridge, 2003;Melgar, Lind, & Muheim, 2015;Montgomery & Heinemann, 1952). Moreover, humans have been found to be good predictors of insect prey survival in the wild (Penney, Hassall, Skevington, Abbott, & Sherratt, 2012). Finally, models of predator vision (both for birds and humans) were used to complement behavioural experiments and infer the relative detectability of each butterfly species based on their contrast against the background.

| Behavioural experiments using wild birds
Behavioural experiments took place in August and September 2017 at the Konnevesi Research Station (Finland). Thirty wild-caught great tits (P. major) were used. Birds were caught using spring-up-traps and mistnets, individually marked with a leg band and used only once. Each bird was housed individually in an indoor cage (65 × 65 × 80 cm) and was fed with seeds and water ad libitum, except during training and experiments. During training, birds were given mealworms attached to butterfly wings (see Training section). Birds were deprived of food for up to 2 hrs before the experiment to increase their motivation to hunt.

| Training
In indoor cages, birds were taught that all four species of butterflies were similarly palatable by offering them laminated wings of four butterflies (one of each species) with a mealworm attached to the copper wire. Wings were laminated during training only, using transparent thin plastic so as to minimize damage and enabling us to reuse the wings between trials. Butterflies were presented to the birds in the absence of vegetation during training so as to enhance the association between butterfly colour patterns and fully edible prey.
When birds had eaten all four prey items (one of each species), a new set was presented. Training ended when birds had eaten three sets of butterflies. No time constraint was imposed for training, and most birds completed it in < 4 hrs.
In order to familiarize birds with the experimental set-up, which was novel to them, they were released in the experimental cage by groups of two to four birds for approximately 1 hr the day before the experiment. Oat flakes, seeds and mealworms were dispersed over leaves and vegetation so as to encourage searching for edible items in locations similar to where butterflies would be placed during the experiment.

| Experiments
The experimental set-up consisted of a 10 m × 10 m cage that had tarpaulin walls and a ceiling of whitish dense net that let in natural sunlight. Butterflies were disposed in a 5 × 5 grid, delimited by poles all around the borders and a rope defining rows and columns (see Supporting Information Figure S3). Five specimens of each species (20 specimens in total) were placed in the grid, one per cell. Before each trial, butterflies were photographed over graph paper, used as a scale to measure butterfly size on ImageJ (Rueden et al., 2017). Butterflies were pinned on top of meadowsweet leaves (Filipendula ulmaria) that had naturally grown in the outdoor cages. Butterflies were always put in similar places within the cell and could be easily seen from a nearby pole. Butterfly position was randomized, but care was taken in (a) leaving the five cells closest to the observer empty as birds tended to avoid this area, (b) avoiding having more than two specimens of the same species in the same row or column and (c) having two specimens of the same species in neighbouring cells. This ensured that all species were evenly represented along the grid. This random configuration was reshuffled between trials.
For each trial, an observer, hidden to the birds, watched from outside the cage through a small window and took notes of which butterfly species were attacked and in which order. A GoPro camera also recorded the experiments. A butterfly was considered detected only if a bird directly approached to attack it, including when the attack failed. No bird was seen hesitating during an attack once it had initiated it. Experiments took place between 9 a.m. and 5 p.m.
Before each trial, the radiance of ambient light (coming from the sun and sky) was taken by spectrophotometry in the same location. We computed the total radiance (TR) over the bird's spectral sensitivity, which ranges from 300 to 700 nm, to account for the intensity of ambient light associated with each experimental trial in the statistical analyses. Further information on weather conditions (cloudy, sunny, etc) was also recorded. Experiments ended when a bird had eaten half of the available butterflies (i.e., 10 butterflies) or after 2 hrs, whichever happened first. Wings were occasionally reused if they had not been damaged.
To control for any positional effect on overall species detection, we computed the probability of a bird being present in a given grid area. To do so, a 10-min interval of each recorded trial was selected and revised to calculate the proportion of time birds spent on the different poles. The time intervals were possible for all trials as they all lasted at least 10 min and were selected either as a result of the birds actively attacking prey or actively exploring the cage during that time, based on notes taken by the observer. These probabilities were later used to divide the grid into four main areas according to bird occupancy: furthest and closest corner to the observer, grid border and grid centre (Supporting Information Figure S4a). Most birds fed willingly on all butterflies located on the borders of the grid. Given that butterfly species distribution was random and reshuffled between trials, the four species were similarly represented throughout the grid (Supporting Information Figure S4b), so no bias was expected. For more details about permits, husbandry conditions, training and experiments, see Supporting Information.

| Experimental set-up
As with the behavioural bird experiments, artificial butterflies (N = 10 of each of the four species, for a total of 40 butterflies) consisted of one real forewing and one real hindwing assembled with copper wire and placed on leaves, but without the mealworm used in the bird experiments. These butterflies were set up along two corridors in a forest-like understorey habitat of similar vegetation and light conditions. Butterfly order followed a block randomization, with five blocks each consisting of eight butterflies (i.e., two of each species; see Supporting Information Figure S5). This ensured that observers were similarly exposed to the four species all throughout the experimental transect. Whether a butterfly was placed on the left or right side of the corridor was also randomized and both order and corridor side were changed daily. Participants could start the path from either end of the set-up and were given unlimited time to complete the trial. However, they could only move forward on the path. Only one participant was allowed in the path at any given time, and they were accompanied by an observer who recorded which butterflies were found. Trials ended when the participant had completed both corridors.

| Statistical analyses
Experiments using birds and humans were analysed independently.
Differences in the total number of butterflies of each species that were attacked by predators (for the sake of simplicity, we use "attacked" hereafter for both birds and humans) were compared by fitting generalized linear mixed-effect models (GLMM), with bird/human identity as a random factor. A binomial distribution was used for the response variable (attacked or not). For the experiments using birds, butterfly species, butterfly size, trial duration, age and sex of the bird, time to first attack, first butterfly species attacked, butterfly position on the grid (corner-furthest or closest to the observer-, grid border, grid centre), weather (as a qualitative variable), and TR, as well as their interactions, were all included as explanatory variables. For human trials, butterfly species, first species attacked, butterfly position, corridor, left or right side of the path, time of day, gender and age of the participant, duration of the experiment and their interactions were all used as explanatory variables. In each case, the best fitting model was selected based on minimization of Akaike's information criteria (AIC), assuming that models differing by two units or less were statistically indistinguishable (Anderson, Burnham, & White, 1998). Coefficients and standard errors were computed using a restricted maximum-likelihood approach and a Wald z test was used to test for factor significance.
In addition to the total number of butterflies attacked per species, an "inconspicuousness" rank was calculated for each butterfly species, as done in a previous study (Ihalainen, Rowland, Speed, Ruxton, & Mappes, 2012). This ranking takes into consideration both the specimens that were attacked and those that were not for each species. Lower values are assigned to those specimens that were attacked (from 1 to 10, according to the sequence of overall prey discovery), and higher values are given to those specimens that were not attacked (all unnoticed specimens are given a value of 11: the maximum number of butterflies that could be attacked before the experiment ended + 1). For example, if a bird captures two H. ninonia second and fifth in the sequence of captured prey, leaving three specimens unnoticed (out of a total of five placed in the cage), this species gets a rank value of 2 + 5 + (3 × 11) = 40 for that trial.
This inconspicuousness rank distinguishes species attacked first and in higher numbers (lower values of inconspicuousness) from those attacked last and in lower numbers (higher values of inconspicuousness). We fitted a linear mixed-effect model to test for differences in rank for each species, assuming a normal distribution, with rank as the response variable. We fitted independent models for birds and human experiments. For bird experiments, bird individual was considered a random factor, and butterfly species, age and sex of the bird, date, time until first attack, first butterfly species attacked, weather as a qualitative variable and TR were explanatory variables.
For humans, participant identity was a random factor, and butterfly species, first species attacked, time of day, gender and age of the participant, duration of the experiment and their interactions were all explanatory variables. Again, the best fitting model was selected using AIC minimization. GLMMs were fitted using nlme (Pinheiro, Bates, DebRoy, Sarkar, & R Core team, 2009) and lme4 (Bates, Maechler, Bolker, & Walker, 2014, p. 4) packages for R. Moreover, whether specific species were more frequently detected first by either birds or humans was tested using a chi-square test.
Additionally for birds, we tested whether butterfly location in the grid could explain differences in the overall species' detection, that is whether species more likely to be attacked were more often placed on areas more likely to be visited. To do so, the frequency per species on the four different grid zones was compared using a chi-square test.
Finally, we tested whether birds and humans created a "search image" (i.e., improved ability in finding butterflies of a given species after encountering a similar one) by counting the number of butterflies of each species attacked consecutively. Results were compared among butterfly species using a chi-square test. Additionally, whether finding some species improved a bird's or a human's ability to find others was also tested. For each combination of two species, we calculated how many times a butterfly of species 1 was found after a butterfly of species 2. Differences between combinations of butterfly species found by birds were tested using a chi-square test.
For humans, observed results and the frequency at which each possible pair of species was placed consecutively in the original experimental set-up were compared using a chi-square test. All analyses were performed in R (R Foundation for Statistical Computing, 2014).

| Colour measures and vision modelling
Finally, models of predator vision (both for birds and humans) were used to complement behavioural experiments and infer the relative detectability of each butterfly species based on their contrast against the background. First, we measured colour (i.e., reflectance) and transmission properties (i.e., transmittance of transparent wing areas) using spectrophotometry. Vorobyev and Osorio's discriminability model (1998) was then used to calculate the contrast between butterfly and background for birds and humans. Detailed methods for measurements and vision modelling can be found in the Supporting Information (additional materials and methods).

| Behavioural experiments using wild birds
The model that best explained whether butterflies were attacked or not included only the time required before the first attack and the cage area in which the butterfly was located (Supporting Information   Table S1). Butterflies were most likely to be attacked when located in the furthest corners and in the borders than in the rest of the cage (z = 9.13, p < 0.001). By contrast, the inconspicuousness rank of a butterfly species was best explained by a model including butterfly species as an explanatory variable (Supporting Information Table S2).
Which species was attacked first closely matched wing transmission properties: H. ninonia, the fully opaque species, followed by the translucent C. tutia, the transparent and yellow-tinted I. salapia and the most transparent species in our study, B. seba (Χ 2 = 11.07, df = 3, p = 0.011; Supporting Information Table S3). Hypothyris ninonia, which was the most colourful species, was usually the first species attacked (t = −3.15, p = 0.002, Figure 2a; Supporting Information Tables S2 and   S3). Species distribution along the four different grid zones was similar (Χ 2 = 6.19, df = 9, p = 0.72; Supporting Information Figure S4b).
Generally, birds did not attack several butterflies of the same species consecutively (Supporting Information Figure S6a). In the rare instances when they did, no differences between species were found (Χ 2 = 0.6, df = 3, p = 0.90), suggesting that birds did not form a "search image" for any of the butterfly species. No combination of species attacked consecutively at high frequencies were found either (Χ 2 = 10.88, df = 11, p = 0.45).

| Behavioural experiments using human participants
Younger participants found more butterflies than older ones (number of butterflies: z = −2.34, p = 0.019; Supporting Information Figure   S7a). Additionally, participants found more butterflies earlier than F I G U R E 2 Sum of the inconspicuousness rank for each butterfly species calculated from the behavioural experiments using (a) great tits and (b) humans. Species for which butterflies were detected first and most often by birds or humans have lower values of "inconspicuousness rank." Butterfly transparency increases from left to right: Hypothyris ninonia (H), Ceratinia tutia (C), Ithomia salapia (I) and Brevioleria seba (B). Letters above the bars mean significant differences below 0.05 later in the afternoon (number of butterflies: z = −2.80, p = 0.005; Supporting Information Figure S7a). Generally, the more time participants spent on the experiment, the more butterflies they found (number of butterflies: z = 5.21, p < 0.001), although this was most significant for women (number of butterflies: z = −2.96, p = 0.003, Supporting Information Figure S7b). Participants found more butterflies on the corridor that had slightly larger vegetation cover (number of butterflies: z = 3.14, p = 0.002). Participants also found more butterflies at the end rather than at the start of the experiment (number  Figure   S7c). Hypothyris ninonia was also the species most frequently found first, followed by C. tutia, B. seba and I. salapia (Χ 2 = 19.5, df = 3, p < 0.001, Supporting Information Table S3). More butterflies of each species were found when C. tutia was found first (t = −3.96, p < 0.001).
There were also differences in the consecutive order in which butterflies were found. Participants were more likely to find two consecutive butterflies of the same species when they were colourful (H. ninonia-50 times-and C. tutia-58 times) than when they were transparent (B. seba-32 times-or I. salapia-18 times; Χ 2 = 29.14, df = 3, p < 0.001). Brevioleria seba and H. ninonia were found consecutively up to four times in a single trial. Some species were also more likely to be found consecutively after another species. The two most opaque butterflies H. ninonia and C. tutia (found 278 times consecutively), and the two transparent species B. seba and I. salapia (found 186 times consecutively) were found consecutively more frequently than any of the other possible combinations after correcting for the number of butterflies found for each species (Χ 2 = 170.95, df = 5, p < 0.001). These observed frequencies differed significantly from expected as a result of their physical position along the path (Χ 2 = 79.12, df = 11, p < 0.001, Supporting Information Figure S6b).

| Models of bird and human vision
The achromatic-weighted contrast between butterfly colour patches and green-leaf background was similar for both birds and

| Transparency reduces detectability
As initially predicted based on wing transmittance, and as demonstrated by our behavioural experiments and visual modelling results, transparency decreases butterfly detectability.
Interestingly, detection by human participants was similar to that of naïve birds, as shown in other studies (Beatty, Bain, & Sherratt, 2005;Sherratt, Whissell, Webster, & Kikuchi, 2015), providing further support for using human participants to measure predator detection. Surprisingly, experimental results from the bird experiments differed slightly from predictions based on the measures of transmittance of transparent patches and results obtained from the vision models. For instance, according to the transmittance and the chromatic contrast measured between butterflies and their background, birds should have detected C. tutia more easily than the two more transparent species. Indeed, semi-transparent objects should be more easily detected than fully transparent objects at short distances and when more light is available (Johnsen & Widder, 1998), such as conditions present during our experiments. Yet this transparent but brightly coloured species was detected at rates similar to those of the most transparent species, perhaps because transparent butterflies were more easily detected and attacked by birds than we predicted (e.g., if an opaque contour enhances detectability of otherwise transparent prey). Alternatively, the semi-transparent C. tutia could have been less detectable by birds, because it shows less strongly delimited contours than those of the most opaque species H. ninonia.
Perhaps this hampered its detection as occurs in disruptively coloured prey (Honma, Mappes, & Valkonen, 2015;Stevens & Cuthill, 2006). These contradicting results highlight the importance of combining both modelling and behavioural experiments to better understand the evolution of transparency and other prey defences.

| Transparency in potentially unpalatable butterflies?
Our results demonstrate that transparency can effectively reduce prey detectability in ithomiine butterflies, where several species have been experimentally demonstrated to be chemically protected (Brown, 1985;Trigo et al., 1996). This is surprising as aposematic colour patterns, rather than inconspicuousness, are more common in toxic and unpalatable prey (Mappes et al., 2005;Poulton, 1890;Ruxton, Sherratt, & Speed, 2004). In fact, conspicuousness is positively correlated with toxicity or unpalatability in some species and can thus be an honest indicator of prey defences (Arenas, Walter, & Stevens, 2015;Blount, Speed, Ruxton, & Stephens, 2009;Maan & Cummings, 2012;Prudic, Skemp, & Papaj, 2007;Sherratt & Beatty, 2003). Moreover, predators learn more quickly to avoid unpalatable prey when colours are more conspicuous (Gittleman & Harvey, 1980;Lindstrom, Alatalo, Mappes, Riipi, & Vertainen, 1999). This might suggest that the evolution of transparency in these butterflies is the result of a loss or a reduction in unpalatability. If this is the case, the existence of mimicry rings of transparent clearwing butterflies remains unexplained, as this is usually the result of convergence of warning signals promoted by the positive frequency-dependent selection exerted by predators (Willmott et al., 2017). Alternatively, if defences are costly, prey may invest in either visual or chemical defences (Darst, Cummings, & Cannatella, 2006;Speed & Ruxton, 2007;Wang, 2011), as such options have been shown to afford equivalent avoidance by predators (Darst et al., 2006). In which case, transparency should instead be associated with an increase in unpalatability. This relationship between transparency and chemical defences in clearwing butterflies remains to be explored.
Alternatively, transparency may lower detection and function as a primary defence, with aposematism taking over as a secondary defence if the prey is detected. Indeed, transparent butterflies were not completely cryptic for either birds or humans. In fact, although birds detected the most colourful species first, in total they found a similar number of both colourful and transparent butterflies. Moreover, humans appear to learn to detect and perhaps remember common elements between the more transparent species, which might be the result of a search image. As such, Ithomiini butterflies may be cryptic from afar, but perceived as conspicuous from up close. The combination of crypsis and conspicuousness has also been shown for other defended prey (Järvi, Sillén-Tullberg, & Wiklund, 1981;Sillén-Tullberg, 1985). For example, toxic salamanders of the genus Taricha are generally cryptic, only revealing their warning coloured underbelly when threatened (Johnson & Brodie, 1975). In Ithomiini, conspicuous elements such as opaque areas that delineate the edges and contrast with the background likely increase detection, as has been shown for artificial moths (Stevens & Cuthill, 2006). Furthermore, pigmentary or structurally produced opaque colours, such as the white band in B. seba, may also enhance butterfly detection. This suggests, as do our results and the occurrence of co-mimics in natural habitats, that these butterflies may reduce the cost of conspicuousness using transparency in addition to maintaining the benefits of detectable warning signals. Further behavioural experiments testing the distance at which Ithomiini butterflies are detected are needed to shed further light on the function of aposematism in less conspicuous prey.
Finally, transparency may have evolved as an additional protection against birds such as adult kingbirds (Tyrannus melancholicus, Pinheiro, 1996), which are able to tolerate their chemical defences.
Indeed, both theoretical (Endler & Mappes, 2004) and experimental (Mappes, Kokko, Ojala, & Lindström, 2014;Valkonen et al., 2012) studies have shown that weak warning signals (not overtly conspicuous) can evolve and be maintained in communities where predators vary in their probability of attacking defended prey. Larvae of Dryas iulia butterflies, pine sawfly larvae (e.g., Neodiprion sertifer), and shield bugs (Acanthosomatidae, Heteroptera) are only a few of the examples that exist of unpalatable species that display weak visual warning signals (Endler & Mappes, 2004). As in the polymorphic poison frog Oophaga granulifera, clearwing species may reflect a continuum between aposematism and crypsis, possibly shaped by differences in the strength of predator selection as a result of the frequency of naïve predators and/or the variation in predator sensitivities to chemical compounds (Willink, Brenes-Mora, Bolaños, & Pröhl, 2013). A thorough characterization of unpalatability, microhabitat and predator communities would be useful in better understanding conditions that promote the evolution of transparency.

| CON CLUS IONS
Our study, which combines behavioural experiments with different predators and vision modelling, provides important insights into the complex role transparency may play in predator defences of terrestrial aposematic organisms. We show for the first time that transparency results in the reduction of detectability of terrestrial prey. We also demonstrate that Ithomiini butterflies may in fact be decreasing the costs of conspicuousness, while still retaining visual elements that are recognized as warning signals. Future studies exploring the efficiency of combining transparency and warning signals in decreasing predation risk will further contribute to our understanding of the evolution of cryptic elements in aposematic prey.