2.1 Sampling of wild populations

Sampling was performed in 2012 and 2013 along Pipeline Road, Gamboa (elevation 60 m), which transects open to closed forest, and the nearby Soberanía National Park. We focused on seven common Heliconius species which are locally abundant and for which ecological data on habitat and pollen preference (Estrada & Jiggins, 2002) and host plant preference (Merrill et al., 2013) are known: Red/Yellow co-mimics H. melpomene and H. erato, which generally occupy open canopy forest edge habitats, Black/White co-mimics H. sapho and H. cydno, which generally prefer closed forest habitats, tiger stripe co-mimics H. hecale and H. ismenius, and H. sara which lacks a co-mimic locally, all of which are generally found in mixed or open habitat (see Figure 1 for phylogenetic and ecological summaries). Samples were collected concurrently under permits SEX/A-3-12, SE/A-7-13 and SE/AP-14-18. All species lack seasonal reproductive cycles and generations are overlapping. Sampling all species with a given time period therefore likely samples randomly with respect to age across all species. We sampled 9–10 wild individuals of each species, including 3–5 females/species. For insectary groups we sampled 5 males and 5 females of each species, with the exception of H. ismenius, where only five individuals were available within the sampling period. Full details of all samples are in Table S1.

2.2 Insectary-reared individuals

To determine whether any variation we observed was due to environmentally induced plasticity, we performed common garden experiments for five species in 2013, for which a preferred host plant was available. Insectary-reared individuals were obtained from wild-caught females which were kept under standard conditions in outbred stock cages (c. 2 × 2 × 2 m) of mixed sex and equal densities at the Smithsonian Tropical Research Institute's Gamboa insectaries. These cages are maintained on the edge of the butterfly's native habitat, and light conditions do not substantially deviate from the forest edge environment. Species were reared on their preferred local host plant; H. melpomene on Passiflora menispermifolia, H. sara on P. auriculata, H. hecale on P. vitifolia, H. ismenius on P. triloba and H. erato on P. biflora. Individual larvae were then raised on new growth shoots in outdoor larval cages. After eclosion, adults were aged for 2–3 weeks for the neuroanatomical samples, with individual age recorded (Table S1).

2.3 Experimental manipulation of habitat complexity

To further test whether the spatial complexity of the adult environment impacts post-eclosion brain development, we performed an additional experiment using two larger cages (4 × 4 × 2.5 m), focusing on H. hecale as an experimental subject. These two cages were constructed of the same semi-transparent mesh as the main stock cages. In each cage, food (artificial feeders and flowering Cephaelis tomentosa) and host plant (P. vitifolia) resources were placed in identical positions. One of the cages was left open, while the other (‘divided’ cage) was subdivided by a series of internal walls to create ~1 m wide corridors through which the butterflies had to travel to move between resources. Freshly eclosed, individually marked H. hecale were added to each cage (assigned randomly) over the course of approximately 1 month, after which all surviving butterflies were collected, following a 2-week period without new introductions to ensure a minimum age. These samples were subsequently compared to wild caught and stock-cage reared individuals. Individuals used in this experiment were sourced from the same stock populations, at the same time, as other insectary-reared groups. The experiment was repeated twice, with the position of the open/divided cages swapped to control for effects of the position of the sun.

In addition, behavioural activity was examined by randomly selecting an individual and performing a 5-min observation, recording basic measurements of time in flight, resting, feeding, or inspecting host plants, and the number of social interactions (Table S2). This was done to ensure the butterflies were behaving in equivalent ways in both cages to rule out non-spatial effects. In addition, in the 4 days preceding the final sampling, individuals were collected after dusk and placed in a pop-up cage (~1 m × 0.8 m × 0.8 m) where they slept overnight. Approximately 2 h after sunrise, a period of time in which the butterflies were generally inactive with the exception of wing fanning behaviour to warm the flight muscles, butterflies were released individually from a common position and the time (total and time in flight) to first feed was recorded. We recorded performance in the cage the individual had previously experienced, over two trials in two consecutive days, before recording performance in the alternative cage, again for two trials in consecutive days. By comparing performance in the divided and open cages, we aimed to assess whether prior experience of the divided cage increases foraging efficiency, with the expectation of no effect in the open cage as food resources were visible with no obstacle in the direct flight path from the release site. At the end of the trails, individuals were terminally sampled, with brains preserved for neuroanatomical comparisons. Individual age was recorded (Table S2) and included in statistical models where appropriate (Tables S7 and S8), but this was not possible where models also include wild caught data where age is not known. In total, behavioural and anatomical data were collected for 11 individuals matured in the large ‘open’ cage (6 female, 5 male) and 10 individuals matured in the large ‘divided cage’ (4 female, 6 male). These were compared to 10 recently emerged individuals, and 10 individuals aged in the smaller stock cages, and 10 wild caught individuals (all 5 female, 5 male) (Table S2).

2.4 Immunohistochemistry, imaging and volumetric measurements

Brains were fixed in situ using a ZnCl₂-formaldehyde solution for 16–20 h, following Ott (2008). Further methodological details and anatomical descriptions of the Heliconius brain are available in Montgomery et al. (2016). Briefly, brain structure was revealed using immunofluorescence staining against a vesicle-associated protein at presynaptic sites, synapsin (anti-SYNORF1; obtained from the Developmental Studies Hybridoma Bank, University of Iowa, Department of Biological Sciences; RRID: AB_2315424) and Cy2-conjugated affinity-purified polyclonal goat anti-mouse IgG (H + L) antibody (Jackson ImmunoResearch Laboratories), obtained from Stratech Scientific Ltd., (Jackson ImmunoResearch Cat No. 115-225-146, RRID: AB_2307343). Imaging was performed using confocal laser-scanning microscopy (Leica tcs SP5 or SP8, Leica Microsystem) with a 10× dry objective with a numerical aperture of 0.4 (Leica Material No. 11506511), a mechanical z-step of 2 μm and an x-y resolution of 512 × 512 pixels. The z-dimension was scaled by 1.52 to correct the artefactual shortening (Montgomery et al., 2016). We assigned image regions to brain components, or neuropils, using the Amira 5.5 (Thermo Fisher Scientific) labelfield module and defining outlines based on synapsin immunofluorescence. We reconstructed total central brain volume (CBR), six paired neuropils in the optic lobes (OL), six paired and one unpaired neuropils in the central brain (CBR) in all wild individuals, using the measure statistics module to estimate component volumes. In insectary samples the POTu, a small posteriorly located neuropil, was inconsistently stained and was not measured. The total volume of segmented structures in the CBR was subtracted from total CBR volume to obtain a measure of the remaining, unsegmented CBR (rCBR), which is used as an allometric control throughout. Due to the lack of volumetric asymmetry in Heliconius neuropils (Montgomery et al., 2016) we measured the volume of paired neuropils from one hemisphere, chosen at random unless one hemisphere was damaged, and multiplied the measured volume by two. Data for H. erato and H. hecale were previously published in Montgomery et al. (2016), and data for H. cydno and H. melpomene are available in Montgomery et al. (2021). All volumes were log₁₀-transformed before data analysis.

2.5 Statistics

We identified non-allometric differences between brain component sizes using nested linear models, analysed in the lme4 R package (Bates et al., 2015). Linear models included each brain component as the dependent variable, rCBR and taxonomic/experimental grouping as an independent variable, with sex and individual (where relevant, in the cage experiments only) included as random factors. The likelihoods of nested models were compared using the ANOVA function against a chi-squared distribution. Correction for multiple testing was performed using a sequential Bonferroni procedure (Benjamini & Hochberg, 1995), and unless otherwise stated all results referred to are corrected for multiple tests (indictaed as p_adj). Diagnostics for these models were assessed using the package DHARMa (Hartig, 2022). All post hoc comparisons were made by obtaining the estimated marginal means using the R package lsmeans v 1.7.0 and were corrected for selected multiple comparisons using the Tukey test (Lenth, 2016). As the low species number prohibits formal correction for phylogenetic effects, we followed Estrada and Jiggins (2002) approach and test for effects of clade (Erato [H. erato, sara, sapho] vs. Melpomene [H. melpomene, cydno, hecale, ismenius]) as a way of exploring for phylogenetic effects at the deepest split in the data, in addition to exploring species effects. Effects of plasticity were analysed by constructing mixed models with and without the grouping variable (wild vs. insectary) as a fixed effect (with Species and Sex as random effects) to test for significant neuroanatomical differences between insectary-reared and wild individuals.

For neuropils showing a significant clade/species/group effect, we subsequently explored the scaling parameters responsible for group differences using SMATR v.3.4-3 (Warton et al., 2012). Using the standard allometric scaling relationship: log y = β log x + α, where y is the brain component of interest and x is rCBR, we first performed tests for significant shifts in the allometric slope (β) between taxa. If this test indicated conserved slopes between species/groups, we performed two further tests which assume a common slope: (1) for differences in α (or the intercept) that suggest discrete ‘grade-shifts’ in the relationship between two variables, (2) for major axis-shifts along a common regression line. Deviation from a shared scaling relationship, by slope or elevation, can indicate an adaptive change in the functional relationship between two brain structures (Montgomery, 2013).

To analyse patterns of covariance between functionally linked neuropils in the optic lobes (lamina, medulla, accessory medulla, lobula plate, lobula, ventral lobula), multiple linear models were constructed using the lm function in the lmerTest package (Kuznetsova et al., 2017) where each optic neuropil of interest was regressed against the other five optic neuropils, while also controlling for rCBR and Species. This allowed us to test whether interspecific volumetric differences were a result of co-ordinated evolution or whether selection acts on each neuropil independently, leading to non-allometric species differences in the relationships between functionally linked neuropil in the optic lobe (Barton & Harvey, 2000; Montgomery & Merrill, 2017). We also created another similar covariance matrix, this time including the anterior optic tubercle (AOTU), the primary optic neuropil in the central brain. The AOTU is thought to be involved in the parallel processing of a range of visual stimuli (Mota et al., 2011, 2013; Pfeiffer et al., 2005) the importance of which might vary between the ecological niches Heliconius species occupy. Separate covariance matrices were built for the wild and insectary-reared individuals.

Finally, we performed tests of whether the species differences we detected could be simply explained by habitat preference (open/mixed/closed; Estrada & Jiggins, 2002), pollen or host-plant use. Two strategies were followed to explore correlates of interspecific variation between mushroom body morphology and pollen use. First, species were sorted into large- and small-pollen-grain specialists as per Boggs et al. (1981), and Estrada and Jiggins (2002). Second, the proportion of pollen loads collected from different plant species was summarized with a principle component analysis (PCA), which largely separates Heliconius taxa along one axis of variation, which we refer to as pollen species preference. A similar two-step approach was followed to test for effects of host plant ecology. First, butterfly species were classified as specialist or generalist based on egg-laying data (Jiggins, 2017; Merrill et al., 2013). Second, host plant species preference was summarized with a PCA, with the main axis of variation separating H. hecale from the other five species. Note, some ecological traits are completely predicted by clade, meaning possible phylogenetic effects could not be ruled out.

3.1 Brain structure varies extensively across a community of closely related species

There is significant interspecific variation in brain component volumes in wild-caught individuals of the seven species sampled when scaled against the rest of the central brain (rCBR) to control for allometric effects (Figure 2a; Table S3). Results from linear mixed models detected significant species effects in the scaling relationship with rCBR for 9 of the 15 neuropils measured, suggesting that distinct differences in brain structure exist between species (Table S3A). No significant clade effects (Erato clade vs. Melpomene clade) were observed for any neuropils (Table S3B). This suggests that interspecific variation is not explained by the phylogenetic split between the two major Heliconius clades. SMATR comparisons reveal that all interspecific differences among wild-caught individuals were a result of significant elevation shifts (α), with the slopes of each neuropil's scaling relationship being conserved between species (Table S4A). In general, we find little evidence of extensive sexual dimorphism across species, with only the AOTU (larger in males) and mushroom body lobes + peduncle (larger in females) showing statistical support for dimorphism after correction for multiple tests (Table S3D).

3.2 Interspecific differences in sensory neuropils, but not other brain structures, are maintained in captive-reared butterflies

In our analysis of wild individuals, four of the six functionally linked neuropils in the optic lobes (medulla, ${\chi }_6^2$ = 33.903, p_adj < 0.001; lobula plate, ${\chi }_6^2$ = 33.489, p_adj < 0.001; lobula, ${\chi }_6^2$ = 37.86, p_adj < 0.001; ventral lobula, ${\chi }_6^2$ = 44.707, p_adj < 0.001) and a downstream central visual neuropil, the AOTU (${\chi }_6^2$ = 59.950, p_adj < 0.001), showed significant species differences (Figure 2a). To confirm that variation in visual neuropils is not explained by plasticity alone, we repeated tests of interspecific variation using insectary-reared individuals for 5 of the 7 species. With this common garden dataset, all components of the lobula system (lobula, lobula plate and ventral lobula) no longer show robust interspecific variation after correcting for multiple tests, but the medulla (${\chi }_4^2$ = 17.385, p_adj = 0.023, see Table S3Aiii) and AOTU (${\chi }_4^2$ = 32.863, p_adj < 0.0001) remain significant, while the lamina instead becomes significant (${\chi }_4^2$ = 22.442, p_adj = 0.002) having not been so in the wild dataset (Table S3Ai). This difference in result in wild-caught and insectary-reared specimens is not explained by sub-setting the data from seven to five species (Table S3Aii). However, when the volumes of visual neuropils in wild and insectary individuals were directly compared, the only significant difference is for the accessory medulla (${\chi }_1^2$ = 26.982, p_adj < 0.001; Table S3C). This suggests that the loss of significance for some neuropils may primarily reflect a reduction in effect size, with only the accessory medulla showing statistical evidence of a degree of plasticity. The volume of the antennal lobe, the primary olfactory neuropil, does not vary across species after correcting for clade effects in wild-caught (${\chi }_6^2$ = 18.048, p_adj = 0.086) or insectary-reared individuals (${\chi }_4^2$ = 9.756, p_adj = 0.627).

To further explore which neuropils are most directly impacted by species-specific selection regimes, we analysed co-variation among the visual neuropils in the wild data. Neuropils are physically connected through projection neurons, and as a result it is possible that an expansion in one neuropil would lead to knock-on volumetric effects in other neuropils (Kinoshita et al., 2015; Montgomery et al., 2013). We therefore examined how these neuropils co-vary across species. Patterns of covariance from linear multiple regressions revealed that the four largest neuropils in the optic lobes (lamina, medulla, lobula and lobula plate) form a covarying network (Figure 2c; Table S5A1). The lamina (${\chi }_1^2$ = 21.047, p < 0.001), lobula plate (${\chi }_1^2$ = 13.627, p < 0.001) and lobula (${\chi }_1^2$ = 10.052, p = 0.002) all covary significantly with the medulla after controlling for species as a random effect. After accounting for this covariance and including species as a fixed effect, significant species differences are then only observed for the medulla (F₆ = 2.831, p = 0.018), lobula plate (F₆ = 4.475, p < 0.001) and ventral lobula (F₆ = 3.602, p = 0.004). This suggests that volumetric changes in these neuropils may be partly independent and respond to species-specific differences in ecology, but differences in other neuropils may be the result of indirect selection. Models including the AOTU, a major visual neuropil in the central brain, produced similar results (Figure 2d; Table S5A2). The AOTU covaries significantly with the medulla (${\chi }_1^2$ = 9.081, p = 0.003) and accessory medulla (${\chi }_1^2$ = 4.442, p = 0.035), independently of rCBR, consistent with the AOTU receiving projections from these two neuropils (Homberg et al., 2003; Mota et al., 2011). However, significant species differences were still observed for the AOTU (F₆ = 3.198, p = 0.009) when medulla size was included in the model, suggesting that the AOTU varies across species independently of the medulla. In contrast, in the same model a species effect is no longer observed for the medulla (F₆ = 0.852, p = 0.536]).

3.3 No evidence that micro-habitat convergence drives patterns of investment in visual structures

Given evidence of interspecific variation that is not solely explained by plasticity or phylogenetic effects, we next sought to test whether micro-habitat preference explained variation in sensory investment. This follows evidence of neural divergence between H. melpomene and H. cydno, a pair of sister species which are consistently separated across an open versus closed habitat (and therefore light intensity) gradient (Montgomery et al., 2021). However, our models revealed no significant effect of habitat on any of the sensory neuropils (Figure 2a; Table S6C) suggesting that occupation of the same microhabitat has not generally resulted in convergence in sensory neuroanatomy within this community. We therefore turned to more specific ecological traits, preferences when pollen foraging and host plant use. Both ecological axes are impacted by habitat preference, resulting in biases in the available plants in different forest types (Estrada & Jiggins, 2002). We again found no significant variation in sensory neuropil investment between pollen groups (small vs. large grains), or signs of preference-associated plasticity, after controlling for species (Table S6A1,B1). This result was also found when pollen species preference (i.e. plants visited, data from Estrada & Jiggins, 2002) was summarized using a Principal Component Analysis (PCA). As expected, the main axis of variation was between the two Heliconius clades (t_66.857 = 3.572, p < 0.001) suggesting that evolutionary relatedness predicts the range of pollen sources certain species feed on, with no effect on any of the sensory neuropils (Table S6A2). Given the role of visual cues in host plant detection and preference (Dell'Aglio et al., 2016), we also reasoned there may be an effect of levels of specialism on visual neuropils, but again found no significant associations between host plant strategy or species preference and investment levels (Table S6A2,B2).

3.4 Strong environmental effects on mushroom body size, but no association with foraging or host plant ecology

Beyond the sensory neuropils, the major integration centres of the Heliconius brain also show evidence of interspecific variation in wild individuals (Tables S3 and S4; Figure 3). In contrast to the sensory neuropils, the effect of species becomes non-significant for all of the central brain neuropils when comparing individuals reared in common garden conditions (mushroom body calyx, ${\chi }_4^2$ = 0.059, p_adj = 0.822; mushroom body lobes + peduncle, ${\chi }_4^2$ = 8.797, p_adj = 0.929; total mushroom body, ${\chi }_4^2$ = 9.782, p_adj = 0.620; central body, ${\chi }_4^2$ = 6.038, p_adj = 1.000; Figure 3d). The mushroom body in particular shows significant effects of group (wild vs. insectary; mushroom body calyx, ${\chi }_1^2$ = 17.829, p_adj <0.001; mushroom body lobes + peduncle, ${\chi }_1^2$ = 22.072, p_adj <0.001, total mushroom body, ${\chi }_1^2$ = 21.474, p_adj < 0.001; Table S3Cb). This indicates a consistent, substantial degree of environmentally induced plasticity in these neuropils.

Of the three visual neuropils still showing species differences after controlling for variation in connected neural tissue, two co-varied significantly with the mushroom body calyx in wild individuals (lobula plate, ${\chi }_{1,68}^2$ = 8.617, p = 0.003; ventral lobula, ${\chi }_{1,68}^2$ = 0.182, p < 0.0001). In insectary-reared butterflies, the association with the ventral lobula is lost, but in addition to the lobula plate, new associations are found between the mushroom body neuropils and the lamina, lobula, accessory medulla and anterior optic tubercle, as well as the antennal lobe (Table S5Bii), indicating that morphometric associations between sensory neuropil and the mushroom bodies are potentially highly plastic. When accounting for the effect of the two significant visual neuropils (ventral lobula and lobula plate), the calyx, lobes + peduncle, and mushroom body as a whole are no longer significantly different between species using wild-caught individuals (mushroom body calyx, ${\chi }_{6,68}^2$ = 10.824, p = 0.094; mushroom body lobes + peduncle, ${\chi }_{6,68}^2$ = 8.1857, p = 0.225; total mushroom body, ${\chi }_{6,68}^2$ = 10.537, p = 0.104), which suggests a joint plastic response of mushroom body neuropils and some primary visual neuropils.

To explore the environmental relevance of this plasticity in the mushroom bodies, we examined two potential hypotheses: (i) that plasticity relates to foraging for pollen resources (Montgomery et al., 2016), and (ii) that plasticity relates to the degree of host plant specialization (van Dijk et al., 2017) (Figure 4). Contrary to expectations, among wild individuals mushroom body size did not vary significantly with pollen preference (mushroom body calyx, ${\chi }_1^2$ = 1.490, p = 0.222 p_adj = 1.000; mushroom body lobes + peduncles, ${\chi }_1^2$ = 1.666, p = 0.197, p_adj = 1.000). This was also the case for the insectary-reared individuals (mushroom body calyx, ${\chi }_1^2$ = 2.070, p = 0.150, p_adj = 1.000; mushroom body lobes + peduncles, ${\chi }_1^2$ = 0.135, p = 0.714, p_adj = 1.000). No significant effect of host plant strategy was observed on any of the mushroom body components among the wild-caught individuals (Table S6B1; Figure 4b), or on group effects between wild and insectary-reared specimens (mushroom body calyx, ${\chi }_1^2$ = 0.612, p = 0.434; mushroom body lobes + peduncles, ${\chi }_1^2$ = 1.313, p = 0.252). Similarly, host plant preference did not explain variation in mushroom body size in wild or insectary-reared butterflies, and no group effects were found (Table S6B2).

3.5 Effects of the spatial complexity of rearing environment on mushroom body development

To further explore how environmental experience affects mushroom body growth, we placed freshly eclosed adults of H. hecale in two types of large experimental cage: (i) an open 4 m × 4 × 2 m cage; and (ii) a 4 m × 4 m × 2 m cage subdivided using net walls into a simple maze (Figure 5a). Survival rates in the open and divided cages were similar (${\chi }_1^2$ = 0.038, p = 0.844) and observations of individual butterfly behaviour showed that time in flight, resting and feeding, inspecting host plants, and the number of social interactions were largely consistent between cages (Table S7A). To provide an initial test of whether an individual's experience of its home cage (i.e. open or divided) impacted its ability to navigate through our experimental cages, we tested all surviving individuals in both the open and divided cages, and recorded time required to reach the food source. Consistent with a potential learning effect, there was a significant interaction between home cage and test cage on time in flight before finding food resources (${\chi }_1^2$ = 7.573, p = 0.006; Figure 5b). This indicates that prior experience of the divided cage improved foraging efficiency in that cage, but prior experience of the open cage had no effect.

We compared the size of the mushroom bodies of individuals from these experimental cages with those from stock butterflies reared concurrently in smaller cages and derived from the same stock population. We found that among surviving, aged individuals, absolute mushroom body size varied significantly between samples from the large open and divided cages, and the smaller stock cage (total body calyx: F_2,28 = 5.701, p = 0.008; mushroom body calyx: F_2,28 = 4.672, p = 0.018; mushroom body lobes and peduncles F_2,28 = 5.694, p = 0.008). Post-hoc comparisons identified significant pairwise differences between the large and small experimental cages, with larger mushroom bodies in the larger cages, despite individuals from all cages having larger mushroom bodies than newly eclosed adults (Figure 5c,d; Table S8). Overall age and sex do not significantly improve the model (Table S8). The cage effects are consistent for the mushroom body calyx, lobes + peduncle, and the mushroom body as a whole and, while individuals from the smaller stock cage have significantly smaller mushroom bodies than wild-caught individuals, individuals matured in the two larger cages fall within the range of wild-caught individuals (Figure 5; Table S8b).

However, when accounting for allometric scaling with the rCBR, cage type is still a significant factor, but the only partial effect large enough to be significant is between young insectary-reared individuals and individuals from the large, divided experimental cage (total mushroom body, t = −3.304, p = 0.002). Further dissection of the allometric relationships demonstrate group differences are explained by shifts along a major axis (total mushroom body, Wald₄ = 48.33, p < 0.001), which distinguish wild-caught individuals and individuals matured in either large cage, from young and old individuals raised in smaller cages, but not from each other. This pattern suggests experience-dependent co-expansion of the mushroom body and rCBR (Table S8C; Figure 5c,d).