Why has transparency evolved in aposematic butterflies? Insights from the largest radiation of aposematic butterflies, the Ithomiini

Abstract

Defended species are often conspicuous and this is thought to be an honest signal of defences, i.e. more toxic prey are more conspicuous. Neotropical butterflies of the large Ithomiini tribe numerically dominate communities of chemically defended butterflies and may thus drive the evolution of mimetic warning patterns. Although many species are brightly coloured, most are transparent to some degree. The evolution of transparency from a warning-coloured ancestor is puzzling as it is generally assumed to be involved in concealment. Here, we show that transparent Ithomiini species are indeed less detectable by avian predators (i.e. concealment). Surprisingly, transparent species are not any less unpalatable, and may in fact be more unpalatable than opaque species, the latter spanning a larger range of unpalatability. We put forth various hypotheses to explain the evolution of weak aposematic signals in these butterflies and other cryptic defended prey. Our study is an important step in determining the selective pressures and constraints that regulate the interaction between conspicuousness and unpalatability.

1. Introduction

Organisms use a range of defensive strategies that are effective in reducing the probability of being attacked by predators [1]. These include avoiding detection, such as camouflage or crypsis, or deterring predators from attacking when encountered, such as aposematism, whereby defended prey advertise their unpalatability to predators using conspicuous warning displays (see [2]). Aposematism relies on the fact that predators learn to associate certain features that act as warning signals (such as coloration, sounds, odours) with distastefulness upon attacking defended prey. There is evidence that predators learn to avoid unpalatable prey more quickly and remember the association between the signal and unpalatability for longer and make fewer recognition errors if prey are more conspicuous [3–5]. The benefits of being conspicuous are also thought to include producing stronger neophobic avoidance by naive predators, a greater reluctance to attack and a lower probability of killing an aposematic prey in an attack as predators may engage in cautious sampling and/or taste rejection of unpalatable prey [6,7].

Hence, defended species often have conspicuous signals that warn potential predators of these defences. Because predators that attack aposematic individuals learn to avoid prey harbouring a specific signal, aposematic signals are under positive frequency-dependent selection [8,9], whereby the efficiency of a signal increases with abundance. Unpalatable species exposed to the same suite of predators therefore tend to converge on the same warning signals because these individuals effectively share the cost of educating predators, and this is known as Müllerian mimicry [10].

Aposematism and Müllerian mimicry occur in a range of organisms, but it has been especially well studied in butterflies [11]. The Neotropical butterfly tribe Ithomiini (Nymphalidae: Danainae) comprises ca 393 species and 53 genera, and is the largest known radiation of mimetic butterflies. Ithomiini numerically dominate forest butterfly communities and are thought to drive convergence in wing colour pattern in multiple mimicry rings [12,13], whereby co-occurring aposematic species, of both Ithomiini and other Lepidoptera lineages, converge in their aposematic signal. Although several species are brightly coloured, over 80% of Ithomiini have more subdued colour patterns as a result of transparency [13,14], and as such, the tribe is commonly known as clearwing butterflies. Transparency is likely a derived state, since many basal Ithomiini lineages and most species of the sister clade Danaini are conspicuously coloured and not transparent.

Nevertheless, all species of Ithomiini, even those that are transparent, exhibit distinct and recognizable colour patterns and, like other unpalatable butterflies, are similarly coloured on both sides of their wings and are characterized by slow flight (often described as part of a chemical defence syndrome). Some species have been shown to be highly unpalatable to both vertebrate and invertebrate predators [15–17] as a result of the pyrrolizidine alkaloids (PAs) they contain. Like Danaini butterflies (sister-group to the Ithomiini) and arctiid moths, some basal Ithomiini species sequester PAs from larval food plants (Apocynaceae), whereas other Ithomiini that have shifted to Solanaceae or Gesneriaceae as host plants obtain PA compounds as adults by feeding on flowers of Asteraceae and Boraginaceae [18–20].

As such, the evolution of transparency in Ithominii butterflies is very puzzling. Transparency is generally assumed to be involved in concealment, but this has mostly been studied in aquatic systems where it is abundant [21,22]. If transparency has indeed evolved as a result of selection for concealment, these butterflies may benefit from both crypsis and aposematism [23–25]. However, because conspicuousness of unprofitable prey is thought to act as an honest indicator of toxicity (see [26–29] but see also [30,31]) and predators have been shown to preferentially avoid brighter prey [3], unprofitable species that are inconspicuous are generally expected to possess weaker defences.

Here, we test the hypotheses that transparent Ithomiini species are both (i) less detectable by predators and (ii) less unpalatable as a result. We estimate detectability by natural predators (i.e. birds) of 33 species of Ithomiini butterflies spanning a range of transparency, from fully opaque to largely transparent and unpigmented, by measuring transmittance and reflectance of the wings, and using models of avian vision. We then assess the unpalatability of 10 species by measuring the PA content and by recording taste rejection of avian predators (domestic chicks).

2. Methods

(a). Specimens used

Butterflies were collected in northeastern Peru (San Martín department) in 2015–2017. Collection sites included forest habitats surrounding Tarapoto (Rio Shilcayo basin: 6°27′30″ S 76°21′00″ W alt 460 m and the Tunel ridge: 6°27′11″ S 76°17′11″ W alt 1090 m), Moyobamba (6°04′34″ S 76°57′27″ W alt 1130 m) and on Pongo-Baranquita road (6°17′53″ S 76°14′38″ W alt 200 m). Specimens (see electronic supplementary material, table S1 for a list of species and number of individuals used) collected for optical (N = 33 species; see electronic supplementary material, figure S1 for photographs of the taxa used) and chemical analysis (N = 10 species) were preserved dry in envelopes until use. Specimens to be used to measure unpalatability using chicks (N = 10 species) were kept in ambient conditions in 2 × 4 × 2 m outdoor insectaries until use. Butterflies were used no more than a few days after capture so as to ensure that they were fresh and reflected naturally occurring toxin content.

(b). Optical measurements, transparency and detectability

To assess the relationship between detectability and transparency, optical measurements were performed on 33 Ithomiini species. Although none of the species harbour obvious sexual dimorphism in wing pattern, all individuals used were female when possible so as to avoid any sex-related differences (see electronic supplementary material, table S1 for details). Transparent species typically possess large transparent areas that can harbour different colours (e.g. yellow, orange, white or none), in addition to opaque areas of varying sizes and colours (electronic supplementary material, figure S1).

Using spectrometry (see electronic supplementary material, Methods for details), the transmittance spectra of the transparent areas and the transmittance and reflectance spectra of the opaque areas were measured. The average of the measurements for transmittance (five on the forewing and five on the hindwing) taken in different transparent areas of the wing was used to assess the transmittance of transparent patches and account for the heterogeneous properties across transparent areas. Differences between individuals of the same species were found to be minimal and less than that found between species (see electronic supplementary material, Methods for details). Therefore, one individual per species was used in subsequent analyses, except for those species for which two individuals were measured and for which we used the mean of both individuals. For each species, a transparency index was computed as the average of the transmittance of transparent and opaque areas in the fore- and hindwing weighted by the proportion of surface occupied by those areas. This index, which can vary between 0 (fully opaque) and 100 (fully transparent), therefore takes into account both the proportion of transparent patches in the wing and the degree of transparency of those patches. Butterfly size was not taken into account in this index.

The brightness and colour contrast of butterflies against a background of vegetation as perceived by birds was quantified using visual modelling and Vorobyev & Osorio's discriminability model [32]. Given that spectral sensitivities of the avian predator community in this study system are unknown, calculations were based on the two main vision systems found in birds: ultraviolet-sensitive (UVS) and violet-sensitive (VS) vision (see details of model parameters in electronic supplementary material). Light conditions used for the vision models simulated ambient light conditions in forest shade, common in rainforest understory and the reflectance spectrum of an average tropical leaf as a visual background [33]. For transparent areas of the wing, the ambient light was assumed to be transmitted by the wing, reflected by the leaf and transmitted again by the wing. For opaque areas, the ambient light was assumed to only be reflected by the wing before reaching the observer. For the semi-transparent species Ceratinia tutia and C. neso, measures of reflectance rather than transmittance were used because these species had very low transmittance in any given transparent areas. The colour and brightness contrast of the wings were obtained by averaging the contrast measured for each colourful or transparent area weighted by the surface area. All computations were done using the Pavo R package [34].

Because species are not independent due to shared ancestry, the relationship between transparency and detectability was assessed by performing phylogenetic regressions of our detectability estimates against the transparency index using the R package caper [35]. We extracted the phylogeny of our 33 species from the Ithomiini phylogeny [36], which contains 340 species, including all those used in this study. For each estimate of detectability, the phylogenetic signal of the residuals of the regression (Pagel's λ) was estimated and the phylogeny was rescaled accordingly, such that only the amount of phylogenetic signal in the actual data was accounted for [37].

(c). Experiments measuring unpalatability using domestic chicks

Domestic chicks (Gallus domesticus) of ca 10 days old were used as model predators (see electronic supplementary methods for further details of the methodology described below; see also [38–40]) to test the unpalatability of 10 different transparent or conspicuously coloured Ithomiini butterfly species (figure 1).

Figure 1.

The 10 taxa of Ithomiini butterflies that were used to quantify toxicity and unpalatability, ranked by decreasing transparency from top left to bottom right. Species on the top row, from left to right, are Ithomia agnosia, Ithomia salapia aquinia, Oleria onega, Godyris zavaleta and Methona confusa. Species on the bottom row, from left to right, are Napeogenes inachia, Ceratinia tutia, Mechanitis polymnia, Tithorea harmonia and Hypothyris ninonia. The transparent group (transparency index greater than 20%) consisted of the six first species (the entire top row + Napeogenes inachia) and the opaque group consisted of the four last species on the bottom right. The left-hand side of each photo shows the dorsal side of the wings against a dark background so as to highlight transparency, and the right-hand side shows the ventral side of the wings against a white background so as to highlight colour patterns. (Online version in colour.)

Prey consisted of orange or green coloured pellets of chick feed, either with or without freshly killed butterflies mixed into the paste (experimental versus control pellets; see electronic supplementary material, Methods for details). A total of six chicks were tested for each butterfly species. Birds had no prior experience of unpalatable prey and were initially trained to eat the dyed chick feed. Chicks were tested with all palatable pellets of both colours before the start of the experiment, and only those that showed no bias were used. Chicks were then presented with a plastic tray containing 20 pellets of each colour, presented alternatingly, for a total of 40 pellets (i.e. the chicks were given a choice-test with 40 pellets, half of which were of one colour and palatable, and the other half another colour and unpalatable). Chicks were allowed to attack (peck or eat) 20 pellets before the food tray was removed. These chicks were then tested for a total of 12 successive trials, always with the same colour combination. By reversing the colour association for half the birds that were tested (i.e. three chicks were tested with green experimental pellets and three chicks were tested with orange experimental pellets, for a total of six chicks tested for each butterfly species), we ensured that there was no colour bias in the pellets that were preferentially attacked. In each session, the number of palatable and unpalatable pellets attacked were recorded. Intra-individual variation in the response of chicks to a given butterfly species was found to be minimal (see electronic supplementary material, Methods).

To evaluate unpalatability of the different butterfly species, the following exponential decrease model was fitted to the data pooled for all six chicks (see electronic supplementary material, figure S2) using the nls function in R

$$E(t)=A+(10-A)\hspace{0.17em}{\mathrm{e}}^{(-(t-0.5)/T)},$$

where t is the trial number, E(t) is the number of experimental pellets eaten at trial t, and A and T are constants. This is because the shape of avoidance learning curves, whereby predators reduce their attack rates on defended prey over a series of test sessions, has been shown to occur in two phases [41]. The first is an acquisition phase, which shows the speed with which a predator learns to associate the warning signal with the defence, and is depicted by the half-life of the exponential decrease model. This is illustrated in the above equation by the variable T, and the higher the value, the more trials are necessary for the chicks to learn to avoid the unpalatable pellets. The second is an asymptotic phase, illustrated by the variable A, where predators appear to have learned the association and do not change their attack rate. A and T therefore decrease with increasing unpalatability. Fully palatable pellets should lead to infinite values of T, which cannot be analysed subsequently, and we therefore arbitrarily set the upper limit to 100.

In the above equation, discretization of the time variable t, which is the number of trials, was accomplished by adding an offset of −0.5. This was done because avoidance learning does not take place at the end of trials, but during the trials. Consequently, the number of pellets recorded as attacked by the chicks during a given trial t actually occurs over time between t − 1 and t. The start of each trial can therefore be set to t − 0.5 if the attack rate is assumed to be linear between these two points. Changes in attack rates are especially fast during the first few trials and the model parameters are therefore very sensitive to the offset. Finally, the rate of attack by chicks was set to 10 before the start of the first trial so as to indicate that chicks were feeding on both colours without bias, as was tested at the start of each experiment. In the equation, this is done by setting the y-intercept (A + (10 − A)) to 10 before the start of the first trial (i.e. for E(0.5)).

To test the relationship between unpalatability (estimated by both the above variables A and T) and estimates of detectability, a phylogenetic regression of detectability on unpalatability was performed as detailed in the previous section (correcting for the exact amount of phylogenetic signal), using the R package caper [35]. In addition, we tested whether transparent species as a group (transparency index greater than 20%; figure 1) differed from opaque species as a group in their unpalatability by performing a phylogenetic ANOVA using the R package phytools [42], and implementing 1000 simulations.

(d). Chemical quantification of pyrrolizidine alkaloids

PA content (free bases and N-oxides) of the 10 species tested with the chick behavioural assays above was quantitatively analysed by colorimetric assay [43]. Individual butterflies (body and wings) were weighed and pulverized using liquid nitrogen, and placed into vials containing 2 ml MeOH. These vials were then stirred continuously for 3 h at room temperature. After filtration using glass pipettes and cotton, 100 µl of the liquid extract was mixed with 0.5 ml of oxidation reagent (5 mg of Na₄P₂O₇ in 1 ml of H₂O₂ 30%, diluted 1 : 200 with MeOH) and then heated in open test tubes at 100°C for 25 min, until complete evaporation of the solution. The dried samples were left to cool to room temperature, and placed overnight in a desiccator so as to remove any traces of residual water. Anhydrous isoamyl acetate (1.0 ml) and acetic anhydride (0.1 ml) were then added to the tubes, covered so as to avoid any water contamination, and placed at 100°C for 2 min. Samples were again cooled to room temperature and treated with 1.0 ml of freshly prepared modified Ehrlich reagent (8 ml of 20% methanolic boron trifluoride (BF₃), diluted in 72 ml of absolute EtOH and 1.4 g dimethylaminobenzaldehyde). Tubes were sealed and heated at 60°C for 5 min and cooled to room temperature for 23 min. Samples were assayed photometrically at 561 nm against a blank solution (a mixture of isoamyl acetate, acetic anhydride and modified Ehrlich reagent). An initial solution of 5 mg of crotaline and 5 mg of retrorsine dissolved in 10 ml of MeOH (initial concentration: 1 mg ml⁻¹) was used to produce six different dilutions in MeOH, which were then used to produce a reference curve.

The average PA content (per unit of weight for each butterfly) was compared between species and between sexes using a non-parametric Kruskal–Wallis test done using SPSS (SPSS Inc., Chicago, IL, USA). A phylogenetic regression (as above) was done to test the correlation between PA content and unpalatability, using both the variables A and T calculated from the above avoidance curves. Finally, the relationship between PA content and (i) detectability and (ii) transparency were assessed by performing phylogenetic regressions and a phylogenetic ANOVA, respectively, as described in the previous section.

Table 1 shows average transparency, detectability, unpalatability and toxin content for all 10 Ithomiini species tested.

Table 1.

Average transparency, detectability by avian predators (chromatic contrast in just noticeable difference units (JNDs) for UVS vision in large gap light conditions), unpalatability, shown as the speed of aversion learning (T) and number of attacks by educated predators (i.e. residual attacks A), both of which increase with increasing palatability, and PA concentration for the 10 taxa of Ithomiini butterflies that were used. The transparent group consisted of species with transparency index greater than 20% (i.e. the six last species). The colourful but semi-transparent species C. tutia was classified as opaque because measures of transmittance were low in any given transparent areas (i.e. although the overall transparency index is relatively high as a result of the surface area of the wing that is transparent, the individual wing areas that are transparent are effectively almost opaque; refer to figure 1 above).

species	transparency index (%)	average chromatic contrast (JNDs)	speed of aversion learning (T)	residual attacks (A)	PA concentration (μg mg−1)
Hypothyris ninonia	0.29	7.03	2.47	3.14	23.55
Tithorea harmonia	0.37	6.53	1.70	0.00	2.83
Mechanitis polymnia	1.30	6.65	1.61	4.77	23.98
Ceratinia tutia	17.57	6.37	100.00	10.71	20.83
Napeogenes inachia	21.10	4.70	0.88	1.11	25.13
Methona confusa	22.22	3.49	1.03	0.25	0.32
Godyris zavaleta	24.06	5.07	2.72	0.10	17.76
Oleria onega	29.35	4.27	0.97	0.36	27.23
Ithomia salapia aquinia	32.28	4.54	7.54	0.00	23.22
Ithomia agnosia	38.19	3.90	0.68	0.99	21.69

3. Results

(a). Transparency and detectability

The relationship between transparency (as measured by a transparency index that takes into account the proportion of the different colour patches and their respective transparencies) and detectability by predators was assessed by measuring the transmittance (the amount of light that passes through the wings) and the reflectance (the amount of light that is reflected) spectra of wings using spectrometry and models of avian vision. The 33 species used in this study span all major clades of Ithomiini (electronic supplementary material, figure S3) and a wide range of transparency, ranging from completely opaque, with a transparency index of zero (e.g. Melinaea mothone, Mechanitis messenoides, Hypothyris mansuetus), to highly transparent (e.g. Heterosais nephele, Ithomia agnosia, Pteronymia gertschi) (electronic supplementary material, figure S1). Detectability (estimated by the chromatic and achromatic contrasts of wing patterns against a green leaf) decreased with increasing butterfly transmittance, regardless of vision type (UVS or VS), type of contrast (chromatic or achromatic) or simulated light conditions (table 2 and figure 2).

Table 2.

Butterfly average transparency in relation to average detectability by bird predators for both UVS and VS vision, in both forest shade and large gaps, and for chromatic and achromatic contrasts. Relationships were assessed by performing phylogenetic regressions of detectability on transparency index, accounting for the phylogenetic signal of the residuals (λ).

bird vision	light environment	chromatic/achromatic	λ	model p-value	adjusted R2	F-stat	d.f.	intercept estimate	coefficient estimate
UVS	forest shade	chromatic	0.43	<0.001	0.53	36.17	30	2.23	−0.03
UVS	forest shade	achromatic	<0.001	<0.001	0.80	123.46	30	6.86	−0.08
VS	forest shade	chromatic	0.44	<0.001	0.53	36.23	30	2.068	−0.03
VS	forest shade	achromatic	<0.001	<0.001	0.80	122.49	30	6.83	−0.08
UVS	large gap	chromatic	0.44	<0.001	0.54	37.36	30	7.58	−0.11
UVS	large gap	achromatic	<0.001	<0.001	0.79	117.90	30	11.65	−0.14
VS	large gap	chromatic	0.45	<0.001	0.54	37.36	30	7.50	−0.11
VS	large gap	achromatic	<0.001	<0.001	0.79	116.64	30	11.58	−0.14

Figure 2.

Relationship between detectability of Ithomiini butterflies by birds (chromatic contrast in just noticeable difference units (JNDs) for UVS vision), in forest shade (white circles) and large gaps (black circles), and butterfly transparency as measured by the transparency index. Lines were fitted by the method of least-squares linear regression for forest shade (dotted line: y = −0.028x + 2.170; R² = 0.607) and large gap (dash line: y = −0.097x + 7.378; R² = 0.616). Results of the phylogenetic regressions are shown in table 2.

(b). Detectability and unpalatability

The unpalatability of 10 of the previous 33 Ithomiini species (figure 1) that differ in their degree of transparency was tested using domestic chicks. Two facets of unpalatability were assessed in our analyses (electronic supplementary material, figure S2): the speed of avoidance learning, depicted by variable T and the number of residual attacks sustained by educated predators, depicted by variable A. This last variable can also be thought of as the strength and durability of avoidance learning. Both A and T decrease with unpalatability (i.e. the more unpalatable a species, the more quickly predators learn to avoid them (T) and the stronger the avoidance over time (A)).

The relationship between unpalatability and detectability was first ascertained with a phylogenetic regression. In this analysis, neither the speed of avoidance learning T (table 3 and figure 3a) nor the number of attacks sustained by educated predators A (table 4 and figure 3b) were significantly associated with detectability of butterflies, across both types of predators, both types of contrasts and both types of simulated light conditions, despite a trend for a positive relationship between chromatic contrast and variable A. However, when transparency was considered as a categorical variable (i.e. largely transparent or opaque), transparent species were found to be significantly more unpalatable when considering the number of attacks sustained by educated predators, as estimated by variable A (transparent: A = 0.469 ± 0.468; opaque: A = 4.653 ± 4.496; phylogenetic ANOVA F = 5.446, d.f. = 1, p = 0.044; figure 3b), but not the speed of avoidance learning, as estimated by variable T (transparent: T = 2.305 ± 2.670; opaque: T = 26.446 ± 49.038; phylogenetic ANOVA F = 5.446, d.f. = 1, p = 0.245; figure 3a). A single Ithomiini species, the colourful and weakly transparent C. tutia (considered as opaque, see electronic supplementary material, Methods), was found to be fully palatable to chicks.

Table 3.

Speed of aversion learning (T) in relation to butterfly detectability by bird predators for both UVS and VS vision, in both forest shade and large gaps, and for chromatic and achromatic contrast. Relationships were assessed by performing phylogenetic regressions of T on detectability, accounting for the phylogenetic signal of the residuals (λ).

bird vision	light environment	chromatic/achromatic	λ	model p-value	adjusted R2	F-stat	d.f.	intercept estimate	coefficient estimate
UVS	forest shade	chromatic	<0.001	0.38	−0.02	0.86	8	−28.94	26.15
UVS	forest shade	achromatic	<0.001	0.92	−0.12	0.01	8	6.31	1.10
VS	forest shade	chromatic	<0.001	0.46	−0.05	0.60	8	−21.16	23.42
VS	forest shade	achromatic	<0.001	0.92	−0.12	0.009	8	6.49	1.07
UVS	large gap	chromatic	<0.001	0.28	0.04	1.33	8	−48.89	11.34
UVS	large gap	achromatic	<0.001	0.97	−0.12	0.002	8	17.59	−0.62
VS	large gap	chromatic	<0.001	0.43	−0.04	0.68	8	−31.45	8.32
VS	large gap	achromatic	<0.001	0.96	−0.12	0.0039	8	19.05	−0.78

Figure 3.

Relationship between unpalatability of Ithomiini butterflies, shown as (a) the speed of aversion learning (T) and (b) the number of attacks by educated predators (A), both of which increase with increasing palatability, and detectability by birds (chromatic contrast in just noticeable difference units for UVS vision in large gap light conditions). Transparent species are shown as white circles and conspicuous/non-transparent species are shown as dark circles.

Table 4.

Number of attacks by educated predators (A) in relation to detectability for both UVS and VS vision, in both forest shade and large gaps, and for chromatic and achromatic contrast. Relationships were assessed by performing phylogenetic regressions of A on detectability, accounting for the phylogenetic signal of the residuals (λ).

bird vision	light environment	chromatic/achromatic	λ	model p-value	adjusted R2	F-stat	d.f.	intercept estimate	coefficient estimate
UVS	forest shade	chromatic	<0.001	0.10	0.21	3.46	8	−5.76	5.06
UVS	forest shade	achromatic	<0.001	0.41	−0.03	0.75	8	−2.96	0.10
VS	forest shade	chromatic	<0.001	0.15	0.15	2.54	8	−4.62	4.78
VS	forest shade	achromatic	<0.001	0.41	−0.03	0.74	8	−2.94	0.10
UVS	large gap	chromatic	<0.001	0.08	0.25	3.98	8	−8.03	1.90
UVS	large gap	achromatic	<0.001	0.87	−0.12	0.03	8	−0.29	0.27
VS	large gap	chromatic	<0.001	0.15	0.14	2.51	8	−6.16	1.59
VS	large gap	achromatic	<0.001	0.88	−0.12	0.02	8	−0.04	0.24

(c). Chemical quantification of pyrrolizidine alkaloids, and comparison with unpalatability and detectability

PA content (μg mg⁻¹) of the 10 species tested with the chick behavioural assays was quantitatively analysed by colorimetric assay and was found to be significantly different between species (Wald = 62.877; d.f. = 9; p < 0.001; electronic supplementary material, figure S4) but not between males and females (Wald = 2.021; d.f. = 1; p = 0.155; electronic supplementary material, figure S4), although content appears to be more variable within males. Surprisingly, no PAs were detected in any individual of the species Methona confusa. PA content was also low for the species Tithorea harmonia, known to sequester PAs from their host plants as larvae [20].

Surprisingly, PA content did not predict unpalatability and there was no relationship between PA content and the speed of avoidance learning T (electronic supplementary material, table S2a and figure S5a) nor the number of attacks sustained by educated predators A (electronic supplementary material, table S2b and figure S5b). For example, M. confusa, which contained little to no PAs, were found to be highly unpalatable to chicks, and C. tutia, which possessed average quantities of PAs, were found to be palatable to chicks.

There was no significant relationship between PA content and detectability (electronic supplementary material, table S3). When transparency was considered as a categorical variable (i.e. largely transparent or opaque), there was no significant difference in PA content between groups (free PA content of transparent species = 19.2 ± 9.8 and of opaque species = 17.8 ± 10.1; phylogenetic ANOVA F = 0.0496, d.f. = 1, p = 0.836. N-oxides PA content of transparent species = 18.6 ± 9.5 and of opaque species = 17.5 ± 10.0; phylogenetic ANOVA F = 0.0378, d.f. = 1, p = 0.850).

4. Discussion

Transparency has likely evolved in clearwing butterflies (Ithomiini) from a brightly coloured aposematic ancestor, and this study shows that it has resulted in a significant decrease in detectability by avian predators (i.e. improved concealment). This is consistent with results from four Ithomiini species that showed that transparency decreases detectability for both human observers and blue tits [44]. Surprisingly, we did not find a concurrent decrease in unpalatability with increasing transparency. Despite considerable empirical support that conspicuousness promotes the effectiveness of aposematic signals, and thus that selection should favour conspicuous signals in well-defended individuals, there are many examples of unprofitable species that are rather cryptic [30,45].

Various reasons have been put forward to explain weak aposematic signals. Some studies have shown that in addition to greater conspicuousness, greater toxicity of prey can also increase the speed and efficiency of predator avoidance learning [9,39], and this could potentially reduce selection on brightly coloured visual patterns [30]. Here, we found that the most transparent species were all highly unpalatable. In fact, transparent species were found to be significantly more unpalatable than opaque species, the latter spanning a larger range of unpalatability, including one fully palatable species. Indeed, although the speed of avoidance learning (T) was not significantly different for both transparent and conspicuously coloured species, the number of residual attacks sustained by educated predators (A) was significantly reduced for transparent species in the absence of a visual cue. This suggests that some aspects of taste, rather than conspicuousness, may improve the memorability of unpalatability in transparent species. However, it should be noted that as a result of the relatively small number of species tested, the difference in unpalatability between transparent and opaque species may not be accurately estimated. This does not refute the findings that all transparent species tested are highly unpalatable, but it may be premature to conclude that transparent species are systematically more unpalatable than conspicuous species. Reduced conspicuousness in aposematic species can be advantageous for several reasons. Some studies have shown that high levels of conspicuousness can potentially lead to aposematic prey experiencing higher attack rates than cryptic prey because conspicuousness increases prey visibility to predators, especially when the prey community is exposed to a large number of naive predators [46–48]. In addition to naive predators, educated predators may also, at times, exert significant selection pressures on aposematic prey [49]. This is because chemically defended prey contain not only toxins, but also nutrients, and such prey may be profitable to attack, for example, when alternative prey are scarce [50,51], or when predators do not have a high toxic load [52,53].

Furthermore, although clearwing butterflies may not be very conspicuous, they are also not entirely cryptic (see [44]) as they possess bold and colourful markings on their wings. This is unlikely to be disruptive coloration, as the pattern elements appear to enhance the outline of the wings, and pattern edges have been shown to improve predator avoidance learning [54]. Moreover, Ithomiini engage in mimetic interactions with other unpalatable butterfly species. Because Müllerian mimicry is the result of convergent selection on a warning display, this suggests that the colour pattern in this system is both recognized and avoided by predators. The coloration of transparent Ithomiini butterflies, or lack thereof, could therefore have a cryptic function from a distance and prevent the individual from being seen by predators, but also have an aposematic function once the individual has been seen [23–25]. It is also possible that maximum conspicuousness is not important as long as the signal is distinguishable from the undefended and edible prey, which are typically cryptic, and/or that the signal contrasts with the background [28,55–57]. This explanation has been suggested for the zigzag pattern of European vipers (Vipera), which is not very conspicuous, but appears to be a compromise between crypsis and aposematism [58]. As such, a less detectable signal, although potentially less efficient in educating predators, could be beneficial due to its lower detectability, particularly in a situation where predators may kill defended prey.

Chemical defences in most Ithominii butterflies are believed to be the result of PAs acquired solely pharmacophagously as adults [15], and this would certainly explain the large variation in PA content observed in this study. However, our study suggests that PAs may not be the only chemicals responsible for unpalatability in clearwing butterflies. For example, no PAs were found in M. confusa, despite being strongly unpalatable, and some species may therefore also sequester other chemical compounds from their Solanaceae or Gesneriaceae host plants. In support of this, larvae of another Methona species (M. themisto) were found to elicit aversive reactions in chicks, suggesting that they contain chemical defences that are potentially sequestered from their host plants [59]. In those species that sequester other toxic compounds, collecting PAs as adults may provide additional protection. Similarly, C. tutia was the least unpalatable species despite PA concentrations similar to those of other more unpalatable species, suggesting that PA sources may differ in their ability to cause aversion in predators. Whatever the reason, large variations in PA contents such as those observed in our study suggest that individuals within the population may differ in their level of chemical defences and that in some species, younger butterflies may be more palatable as they would not have accumulated sufficient toxins. This may have important repercussions on predator–prey dynamics as the presence of potentially more palatable individuals (i.e. automimics) has been shown to increase sampling by predators [40,60].

In conclusion, we found that transparent Ithomiini butterflies, or clearwing butterflies, are less detectable than their brightly coloured counterparts. Although less conspicuous, transparent clearwing butterflies are not any less unpalatable, and may in fact be more unpalatable than opaque species. It seems likely that the colour pattern of transparent Ithomiini butterflies is sufficiently distinguishable from palatable prey to trigger recognition and avoidance learning by predators. The decrease in conspicuousness of transparent wing patterns thus allows butterflies to benefit from both limited detection by predators (crypsis), but also predator deterrence when detected (aposematism). This may be the result of a shift in microhabitat [61], which may alter signal properties or result in differences in the predators that are present or their propensity to attack defended prey. The fact that the transparent species tested were all highly unpalatable suggests that strong unpalatability may decrease selection for conspicuousness and favour the evolution of some forms of crypsis under certain conditions. Strength of unpalatability and potentially some aspects of taste appear to improve memorability in the advent of reduced conspicuousness. Whether this is also true for the other clades of aposematic butterflies with transparent species that engage in mimicry with Ithomiini species (e.g. Pericopina and Dioptini) is unknown. Further studies focusing on the selective pressures acting on aposematic signalling will improve our understanding of the factors that regulate the interaction between conspicuousness and unpalatability.

Ethics

This study was approved by the ethical committee for Animal Welfare of the National Museum of Natural History (ethical declaration no. 68-064).

Data accessibility

Data are available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.d2h629q [62].

Authors' contributions

M.E. designed the study. M.M. designed the behavioural experiments with the chicks, and M.M., C.C., M.C., L.B.-H. and J.B. performed the experiments in the field and analysed the data. M.M., M.E. and D.G. collected specimens for optical and chemical measurements. C.D., D.G. and C.P. performed optical measurements and analyses, under the supervision of C.A., S.B. and M.E. A.M. performed chemical measurements, under the supervision of B.N., M.E. and V.L. C.H. performed colour pattern imaging and analyses. M.M., D.G. and M.E. wrote the paper with contributions from all authors.

Competing interests

We have no competing interests.

Funding

This study was funded by a grant from the Human Frontier Science Program (RGP0014/2016), an ATM grant from the MNHN (TRANSPATOX) and a grant from the French National Research Agency (ANR CLEARWING ANR-16-CE02-0012). M.M.'s postdoc grant was supplied by ANR SPECREP (ANR-14-CE02-0011).