Light environment influences mating behaviours during the early stages of divergence in tropical butterflies

Abstract

Speciation is facilitated when traits under divergent selection also act as mating cues. Fluctuations in sensory conditions can alter signal perception independently of adaptation to the broader sensory environment, but how this fine-scale variation may constrain or promote behavioural isolation has received little attention. The warning patterns of Heliconius butterflies are under selection for aposematism and act as mating cues. Using computer vision, we extracted behavioural data from 1481 h of video footage, for 387 individuals. We show that the putative hybrid species H. heurippa and its close relative H. timareta linaresi differ in their response to divergent warning patterns, but that these differences are strengthened with increased local illuminance. Trials with live individuals reveal low-level assortative mating that is sufficiently explained by differences in visual attraction. Finally, results from hybrid butterflies are consistent with linkage between a major warning pattern gene and the corresponding behaviour, though the differences in behaviour we observe are unlikely to cause rapid reproductive isolation as predicted under a model of hybrid trait speciation. Overall, our results reveal that the contribution of ecological mating cues to reproductive isolation may depend on the immediate sensory conditions during which they are displayed to conspecifics.

1. Background

During ecological speciation, barriers to gene flow evolve as a result of ecologically based divergent selection [1]. In particular, the evolution of premating isolation is facilitated when traits under ecological selection act as mating cues (sometimes known as ‘magic traits’ [2,3]), as this allows divergent natural selection to be transferred to mating behaviours. However, the importance of these ecological mating cues may be affected by rapidly fluctuating sensory conditions, so that mating preferences contribute to reproductive isolation under some conditions, but not others, within the same broad environment. By increasing variation within and between individuals, this may influence population divergence, but this has received relatively little empirical attention. One reason is that robustly detecting these effects during the early stages of divergence, when they are perhaps most relevant, likely requires large, fine-scale behavioural and environmental datasets, presenting a substantial empirical challenge.

It is well established that the sensory environment can alter signal detection and perception [4,5]. This may affect the evolution of reproductive isolation in two main ways. First, adaptations to meet the ecological needs of distinct sensory environments can influence female preferences, subsequently driving divergence in male signals, and leading to reproductive isolation through sensory drive [6]. For example, in Pundamilia cichlids, adaptation of the visual system to the local environment is associated with divergent female mate preference for male coloration [7]. Second, fluctuating environmental conditions may alter the efficacy of signals used in mate choice [4], so that individual mating preferences may act as an important reproductive barrier under some sensory conditions, but not under others (‘context-sensitive preferences' [8]). However, compared to adaptation to broad-scale shifts in the sensory environment, the influence of rapidly fluctuating sensory conditions on mating behaviours and the evolution of reproductive isolation, has received little attention.

If the efficacy of signals depends on the sensory environment, in turn affecting the strength of the interaction between mating cues and preferences, then this may influence if and how premating isolation evolves. For example, in tropical rainforests, variation in cloud coverage and vegetation can rapidly change illumination, available light spectrum and background, across very small temporal and spatial scales, all of which may profoundly affect the perception of colour signals [9]. The efficacy of signalling and detection mechanisms might, therefore, vary considerably from moment to moment, potentially constraining the reliability of visually based mate choice and its efficacy as a reproductive barrier (unless counteracted by choice of display locations maximizing signalling efficacy).

The sensory conditions during signalling may also influence the strength of genetic associations (i.e. linkage disequilibrium, LD) between mating and ecological traits, which are typically required for speciation with gene flow [10]. When ecological traits act as mating cues, LD (between ecological and preference loci) will arise as a natural consequence of mating preferences, but this will be proportional to the strength of the preference [11], which is likely affected by the sensory environment. Regardless, unless preferences are very strong, more robust coupling between mating and ecological components of reproductive isolation likely requires genetic architectures that impede recombination (or ‘one-allele mechanisms’, see [12]). These include physical linkage, or—–in the extreme—pleiotropy [13]. To date, physical linkage between behavioural and ecological components of reproductive isolation has been reported for a few animals, including pea aphids [14], fish [15] and butterflies [16].

Heliconius butterflies are known for their diversity of bright warning patterns, which also act as mating cues (e.g. [17], and are considered one of the best examples of ‘magic traits’ in nature [3]). Variation in warning pattern is largely controlled by just a few major effect genes [18]; and recent work implicates candidate genes associated with the corresponding preference behaviours in tight linkage to one of these colour pattern genes, specifically the transcription factor optix [16,19]. There is substantial evidence that colour pattern alleles have introgressed across otherwise separately evolving Heliconius lineages (e.g. [20]), and the red–yellow banded H. heurippa (figure 1a) acquired its red colour pattern element from local populations of H. melpomene via introgression of optix alleles [21]. By contrast, its close relative H. timareta linaresi, which is assumed to represent the ancestral wing colour pattern of the heurippa/timareta group [22], only displays a yellow band (figure 1a). The red pattern element of H. heurippa is presumably maintained by strong frequency-dependent selection imposed by predators. However, its role as a premating reproductive barrier may depend on how it is perceived by con- and heterospecifics under varying sensory conditions, as the light environment is predicted to alter how colour patterns are perceived by butterfly visual systems [23] (and possibly also by that of their avian predators [23,24]). In particular, red signals are expected to be stronger (and perceived with greater contrast) with increased intensities of long wavelength light [23]. Whether prevailing light conditions affect mating behaviours has not been investigated in Heliconius, but such an effect would suggest that the efficacy of wing patterns as premating reproductive barriers may depend on fluctuations in sensory conditions as individuals move through the environment.

Figure 1.

(a) Locations in Colombia where H. heurippa (red, North) and H. t. linaresi (blue, South) are known to occur. A contact zone has not yet been described, but is likely to exist between the species' distributions shown on the map. (b) Video recording set-up. Mounted females were presented simultaneously with wings spread open. Each was filmed and an associated light logger recorded illuminance every second. The ‘tripod’ was not casting shade on logger or mounted females at any time. (c) Post hoc motion-detection pipeline. A difference image is formed between each frame and its predecessor (identifying pixels with a change in value). Thresholding, blurring and again thresholding pick out significant local changes (signals). ‘Signal’ frames (and surrounding frames) were extracted. (Online version in colour.)

To test whether warning patterns contribute to premating isolation between H. heurippa and H. t. linaresi, and how the strength of divergent mating preferences depends on the signalling environment, we combined real-time measurements of local illuminance with a novel computer vision pipeline extracting behavioural data from corresponding video footage. We also determined levels of assortative mating in choice trails. Our data, including approximately 17 000 behavioural interactions for 387 individuals, allowed us to test (i) whether the two species show differences in visual attraction towards con- and heterospecific patterns, (ii) whether these differences segregate with the colour patterns, consistent with the physical linkage between behavioural and colour patterning genes and (iii) whether the strength of behavioural differences is dependent on the local, but rapidly fluctuating, light conditions.

2. Material and methods

(a). Study species, collection, maintenance and crossing design

H. t. linaresi and H. heurippa occur in similar forest habitats on the eastern slopes of the Andes, are geographically adjacent (figure 1a) and presumably share a contact zone. Despite their nominal species status, H. t. linaresi and H. heurippa likely represent an early stage of divergence; hybrids between the two are viable and fertile [25] and any post-mating isolation is probably limited to selection acting against immigrant warning patterns. Asides from warning patterns, there are no known differences in ecology or habitat use.

We maintained stocks of H. t. linaresi and H. heurippa, established from wild individuals (electronic supplementary material, table S1), between January 2016 and September 2019 at the Universidad del Rosario insectaries in La Vega, Colombia (4°59′ N, 74°20′ W, elevation 1257 m). We generated F₁ and backcross hybrids (BL, backcross to H. t. linaresi and BH, backcross to H. heurippa) (electronic supplementary material, table S2). All butterflies were supplied with approximately 10% sugar solution and Lantana, Psiguria and/or Gurania spp. Females were kept individually, with Passiflora for oviposition, whereas males were kept in groups. Eggs were collected every other day and the caterpillars were raised in plastic cups, while fed on fresh Passiflora leaves.

(b). Trials with mounted females

To test for differences in visual mating behaviours, we assayed males in choice trials with dead mounted H. heurippa and H. t. linaresi females presented simultaneously in an exposed 4 × 4 × 2 m insectary, in which light conditions varied due to changes in cloud coverage and shade cast by vegetation. Virgin females were frozen with their wings spread on the day of eclosion and kept at −20°C for more than 168 h. They were then dried and washed in hexane to remove residual cuticular hydrocarbons and other pheromones. Throughout the experiment, we used 21 H. heurippa and 23 H. t. linaresi females, randomly choosing a pair each day. ‘Females’ were mounted to a horizontal wire (approx. 70 cm above ground) at one of six locations within the experimental cage. Every hour (hereafter ‘trial’), a new location was chosen randomly for each ‘female’ (but female types were never in the same location twice during the same day). A GoPro Hero 5 Black (GoPro, San Matteo, CA) camera (settings and equipment in electronic supplementary material, table S3) was installed at each position, approximately 50 cm above the mounted female (figure 1b). At both mounted females, we attached a HOBO UA-002-64 logger (sensor facing up) to measure illuminance [lux] every second at the same vertical position and approximately 30 cm away from the mounted female (figure 1b); these loggers are sensitive to wavelengths across the full range of Heliconius vision [26]. Cameras and illuminance loggers were time-synced using GoPro Quik and HOBOware software, respectively.

Most of the naive virgin males matured in a separate communal cage before being introduced into the experimental cage, 5 or more days after eclosion, where they were tested in mixed groups (median group size = 22). Males were numbered and received a unique combination of dots on the dorsal side of the wings, allowing identification from videos. Each male was tested over multiple days (median = 12 d). We recorded an average of 3.01 h of material on each of the 246 recording days, with a median starting time at 9.35 (earliest = 7.48, latest = 14.23) and a median end time at 12.43 (earliest = 9.34, latest = 16.23). Conducting behavioural trials across different seasons and at different daytimes allowed us to capture a variety of light conditions.

(c). Computer vision and video analyses

We used a custom motion-detection pipeline, which post hoc discarded video footage with no activity. The detection of frames with motion (signals) was based on difference imaging and thresholding–blurring–thresholding, as implemented in the OpenCV library available for C++ (figure 1c, electronic supplementary material, methods). Not all of the frames of male motion were recognized as ‘signal’ frames, so we also determined frames 1 s before and after a ‘signal’ in R [27]. Reduced videos were created with the OpenCV library in C++ (all code accessible at: github.com/SpeciationBehaviour/visual_preference_heurippa_linaresi). Video material was processed on an HP Desktop computer (i7-7700 CPU, 4 cores), at approximately 1.25× real-time speed. All videos were processed under the same threshold (145) and blur (30) settings, which resulted in low false positive and false negative rates across lighting conditions. The remaining footage was scored manually at 66.6% speed using the MPCHC player. We recorded three male behaviours: ‘approach’ (change of direction towards mounted female), ‘courtship’ (sustained hovering above mounted female) and ‘sitting’ (sitting on mounted female), which were combined for subsequent analyses. For ‘pure males’, we also scored the prevailing light habitat class at the H. heurippa mounted female by eye (electronic supplementary material, methods).

(d). Tetrad experiments with live females

We performed ‘tetrad’ trials with virgin males and females to test for assortative mating. For each trial, sexually mature unmarked H. heurippa and H. t. linaresi males (one of each) were allowed to acclimatize for 15 min in a 2 × 4 × 2 m insectary, at which point H. heurippa and H. t. linaresi virgin females (one of each) were introduced. Cages were regularly monitored by human observers, and once the first mating occurred, the experiment was stopped.

(e). Statistical analyses

Illuminance, measured in lux, is the intensity of light falling onto a surface. Long wavelength-intensity determines the brightness/contrast of the red band of H. heurippa. Overall illuminance can be used as a proxy for long wavelength-light intensity, and this proxy is strongest when the available spectrum of light is similar across the dataset. However, in contrast to other broad light habitats in our cage, ‘woodland shade’ conditions (electronic supplementary material, methods) likely caused a blue-biased spectrum [9]. To correct for this, we also calculated a ‘corrected’ log-illuminance at the H. heurippa ‘female’ for data from pure males by multiplying lux from ‘woodland shade’ conditions (approx. 20% of the data) by 0.85, a factor based on fig. 8 from [9] and the spectral response curve of the illuminance logger (electronic supplementary material, figure S2b, electronic supplementary material, methods). Finally, we log10-transformed ‘corrected’ and ‘uncorrected’ lux measurements (log-illuminance), accounting for the logarithmic response of animal eyes to light intensity [28].

Statistical analyses were conducted in R [27] (scripts: electronic supplementary material, R Markdown and github.com/SpeciationBehaviour/visual_preference_heurippa_linaresi), using the brms package [29], an interface to the Bayesian software Stan [30]. Posteriors are described with a 95% equal-tailed credible interval and the mean as point estimate. We analysed data from trials with mounted females using logistic regression. Male behaviours directed towards the H. t. linaresi or H. heurippa mounted females were fitted as binary Bernoulli-distributed response variable (i.e. 0 and 1, respectively); estimates from the model can be understood as a proportion of interactions with the mounted H. heurippa female, with higher values indicating stronger relative attraction to H. heurippa and lower values indicating stronger relative attraction to H. t. linaresi. Models initially included all possible nested variations of the fully saturated model explaining effects of (i) male type, (ii) log-illuminance at the H. heurippa ‘female’, (iii) log-illuminance at the H. t. linaresi ‘female’, and their interactions. Individual ID and trial were fitted as random effects.

To test for ‘species’ differences, we initially fitted models for the ‘pure’ males (male type = ‘H. heurippa’ or ‘H. t. linaresi’). Segregation of the red bar in BL hybrids (controlled by alleles at optix [18]) allowed us to test for linkage between colour and preference loci [16] (here male type = ‘red’ (Bb genotype) or ‘no red’ (bb genotype) and brood was additionally fitted as a random effect). To determine which terms to retain [31], we calculated the widely applicable information criterion (WAIC) and the leave-one-out information criterion (LOOIC) for each set of models, using the loo package [32], and WAIC weights using the brms package [29]. Weakly informative priors (logit-scaled normal distribution with mean = 0 (no preference) and s.d. = 3) were set for coefficients corresponding to the different male types, giving small prior probabilities for extreme values very close to 0 or 1. For all other coefficients, default (non-informative) priors were applied. Diving illuminance data into categorical data may assure illuminance to be a more stable proxy for long-wavelength intensity. Therefore, we also repeated the model selection process for ‘categorical illuminance’ models, where values less than or equal to median were ‘poorly lit’, and values greater than median were ‘brightly lit’ (electronic supplementary material, figure S1). Finally, we also fitted a model to data from pure males, which included male type, ‘corrected’ log-illuminance at the H. heurippa ‘female’ and their interaction. Posteriors of the estimated marginal means (EMMs) were calculated using the emmeans package [33]. From this, we retrieved posteriors of contrasts and calculated the posterior probability (PP) that EMMs differ.

For the tetrad data, we fitted observed counts of each mating outcome (type of male and female involved) as Poisson-distributed response variable. We included the specific male–female combination of the mating outcome as a fixed factor. We transformed the resulting posteriors of count estimates for each of the male–female combinations into proportions, which effectively results in the same outcome as an intercept-only multinomial model [34]. Non-informative default priors were applied throughout. PPs were calculated with the brms package [29]. We compared the observed frequency of each mating combination to the predicted frequencies based on our measurements of vision-based male preferences from the mounted female experiment. Predictions were derived by multiplying the frequency that a male type was involved in any mating combination by its respective EMMs from the models fitted to the mounted female data. Predictions were based either on the overall EMM for each type, or the interaction term EMMs from the categorical model. Posterior distributions for each prediction were calculated using the binom package under the default prior [35].

3. Results

(a). Species comparisons: (i) divergent visual attraction behaviours in H. t. linaresi and H. heurippa males are dependent on the light environment

Over 1.5 years, we collected 1481 h of footage. Our computer vision pipeline reduced this to 66 h requiring human evaluation (i.e. 4.5% of the total footage recorded), including 16 995 behavioural ‘interactions’ from 387 males (approx. 43.9 per male). We measured illuminance at both mounted females for 83% of recorded behaviours. These data allowed us to determine the effects of male type and illuminance on relative visual attraction to the H. heurippa mounted female (hereafter ‘preference’). The best-fitting model for the pure H. heurippa and H. t. linaresi males retained male type, log-illuminance at the H. heurippa ‘female’ and their interaction (electronic supplementary material, table S4: model #1). This model structure also had the best fit among categorical illuminance models (electronic supplementary material, table S5: model #1). Overall, our results suggest that the local light environment influences the strength of visual attraction.

Across the entire dataset, the illuminance at the mounted H. heurippa female increased the difference in preference between the male types, as evidenced by the interaction between male type and log-illuminance at the H. heurippa ‘female’ (figure 2a and electronic supplementary material, figure S3, PP simple slope H. heurippa > H. t. linaresi = 0.996). This was largely driven by an increase in log-illuminance at the H. heurippa ‘female’ leading to a stronger conspecific preference in H. heurippa males (PP simple slope > 0 = 0.994); there was only limited support for an effect on H. t. linaresi males (PP simple slope < 0 = 0.755). When considering the same model structure, but with ‘corrected’ log-illuminance at the H. heurippa ‘female’, the difference between the slopes becomes only marginally stronger (PP difference between slopes > mean difference from ‘uncorrected’ model = 0.515). Overall, H. heurippa males showed a higher proportion of interactions with the H. heurippa pattern than H. t. linaresi males. Although supported with high credibility (PP relative visual attraction H. heurippa > H. t. linaresi = 0.968), this difference in preference was relatively small (0.07, CrI: 0.00–0.14) and characterized by considerable within population variation (figure 2b). Nevertheless, this effect nearly doubled when the H. heurippa ‘female’ was in brighter light (0.13, CrI: 0.05–0.22; figure 2d) and was absent when the H. heurippa ‘female’ was poorly lit (0.01, CrI: −0.07–0.09; figure 2c). We see the same pattern when scoring lighting conditions by eye from video material. H. heurippa males show a higher proportion of interactions with the H. heurippa pattern when the H. heurippa ‘female’ is in bright sunlight (0.16, CrI: 0.07–0.25; electronic supplementary material, figure S2c), but this difference is absent under cloudy (0.03, CrI: −0.06–0.12; electronic supplementary material, figure S2c) or ‘woodland shade’ conditions (−0.03, CrI: −0.15–0.09; electronic supplementary material, figure S2c).

Figure 2.

The effect of illuminance on relative attraction towards H. heurippa ‘females’ (i.e. proportion of interactions with the H. heurippa ‘female’ as opposed to the H. t. linaresi ‘female’). (a) Relative attraction towards H. heurippa ‘females’ of H. t. linaresi (blue, falling line) and H. heurippa males (red, rising line) under changing illuminance at the H. heurippa ‘female’. Illuminance on the x-axis is log-scaled. Coloured area around each regression line represents the 95% credible interval (CrI). Dashed vertical line represents the median log-illuminance used as a cut-off to define the poorly and brightly lit conditions (below). Relative attraction towards H. heurippa ‘females’ for H. t. linaresi (blue) and H. heurippa males (red): (b) across light environments; (c) for poorly lit H. heurippa ‘female’; (d) for brightly lit H. heurippa ‘female’. Gardner–Altman plots in (b–d) show the difference between the two male types: horizontal lines project from the means of the posteriors for each male type (means and CrIs in electronic supplementary material, table S6). The Δ-axis has the same scaling as the regular y-axis, but with the 0-value shifted to the mean of the H. t. linaresi posterior (hence showing the difference H. heurippa–H. t. linaresi). The mean and the 95% CrI for the posterior of the difference between the male types are shown on the right. Each point represents a single individual and its size is scaled to the number of observations. Custom swarmplot was used to distribute dots horizontally. (Online version in colour.)

(b). Species comparisons: (ii) H. heurippa males mate more often with conspecific females in choice trials

During 89 ‘tetrad’ trials we observed 50 con- and 39 heterospecific matings (PP for positive assortative mating = 0.88). This trend was driven by a higher proportion of conspecific matings involving H. heurippa males (0.405, CrI: 0.305–0.508) than heterospecific matings involving H. heurippa males (0.270, CrI: 0.183–0.366) (PP = 0.94). H. t. linaresi males did not participate more frequently in con- rather than heterospecific matings (PP = 0.43) (figure 3). In general, H. heurippa males mated more often than H. t. linaresi males (60 versus 29 times). Our results closely match predictions derived from the mounted females experiment with brightly lit H. heurippa ‘female’ (horizontal purple bars in figure 3), and, to some extent, those derived from all of the illuminance conditions combined (electronic supplementary material, table S7).

Figure 3.

H. heurippa males show a preference for live, conspecific females in the tetrad experiments. Dashed vertical line indicates expected proportion under random mating. Posterior distributions for each mating pair combination are displayed as histograms (shaded in red = 95% CrI, vertical orange line = mean). Predictions based on the visual attraction data from the mounted females experiment with brightly lit H. heurippa ‘female’ are displayed with their 95% CrI as horizontal purple lines. (Online version in colour.)

(c). Hybrid comparisons: patterns of behaviour in backcross hybrids are consistent with linkage between colour and preference loci

Estimates of preference indicate that F₁ males behave like H. heurippa males (electronic supplementary material, figure S5d), perhaps suggesting that H. heurippa alleles for attraction to red are dominant; however, challenging this, BH males seem to display a preference similar to H. t. linaresi males (electronic supplementary material, figure S5e). For BL males, the model including all possible interactions was the best fit (electronic supplementary material, table S8, model #1). As for the ‘pure’ males, the six best-fitting models included the interaction between type (i.e. red (Bb) and non-red (bb) wing pattern) and log-illuminance at the H. heurippa ‘female’ (95% of cumulative WAIC weight for an effect of this interaction term). Among categorical illuminance models, however, the best-fitting model only included type, illuminance at the H. heurippa ‘female’ and their interaction, i.e. the same model structure supported for data from pure males (electronic supplementary material, table S9: model #1).

As for males from the parental populations, the difference between the two genotypes was higher when the H. heurippa ‘female’ was in bright light (0.08, CrI: 0.02–0.13; figure 4b), but absent when it was poorly lit (0.00, CrI: −0.06–0.06; figure 4a). Illuminance at the mounted H. heurippa female increased the differences in preference between the two BL genotypes (electronic supplementary material, figure S3b, PP simple slope Bb > bb = 0.999). While the slope for Bb males is only slightly positive (PP = 0.867), the slope for bb males is strongly negative (PP = 0.999). By contrast to the pure males, we also observed an effect of illuminance at the H. t. linaresi ‘female’ when illuminance was treated as a continuous variable. Specifically, bb males showed a stronger preference for the H. heurippa ‘female’ when the H. t. linaresi ‘female’ was bright (electronic supplementary material, figures S3d, PP = 0.999); however, the model with three-way interaction was not well supported when illuminance was treated as a categorical variable (electronic supplementary material, table S9: model #15), perhaps suggesting that extreme lux measurements at the H. t. linaresi ‘female’ might be driving this pattern in the continuous data (electronic supplementary material, figure S4). Overall, Bb males were more likely to interact with the H. heurippa ‘female’ than bb males (electronic supplementary material, figure S6). This difference was supported with moderately high probability (PP relative visual attraction Bb > bb = 0.938). Although these represent small effects in absolute terms (0.04, CrI: −0.01–0.08), this difference accounts for approximately 50% of the difference in means of the parental populations, consistent with linkage between colour and preference loci.

Figure 4.

The effect of illuminance on relative attraction towards H. heurippa ‘females’ (i.e. proportion of interactions with the H. heurippa ‘female’ as opposed to the H. t. linaresi ‘female’) for BL hybrid males (no red/bb = blue dots; red/Bb = red dots): (a) for poorly lit H. heurippa ‘female’; (b) for brightly lit H. heurippa ‘female’ (means and CrIs in electronic supplementary material, table S6). Gardner–Altman plots are as in figure 2b,d. (Online version in colour.)

4. Discussion

By collecting approximately 1500 h of mate choice data, we have shown that the light environment can influence visual mating behaviours in Heliconius butterflies. Although our data are characterized by considerable individual variation, we observed significant differences in behaviours of H. heurippa and H. t. linaresi males, and these differences are stronger when the female patterns are more brightly lit. In other words, there is an interaction between innate population-level preferences and the immediate sensory environment in which the female cue is encountered. Experiments with live males and females revealed a degree of assortative mating, and the differences in visual attraction behaviours that we observe are sufficient to explain this. We have also shown that differences in visual attraction are associated with the presence of the red forewing band in interspecific hybrids under bright light conditions, perhaps suggesting physical linkage between an ecologically relevant colour pattern gene and those for the corresponding behaviour.

Studies of speciation often focus on already diverged groups, which are maintained by multiple reproductive barriers [36,37], making it difficult to draw conclusions about the role of individual barriers and their ecological context. Early acting barriers may be of small effect and, especially in the case of behavioural phenotypes, require very large datasets to draw robust conclusions, which may not always be feasible. Behavioural researchers increasingly use computer vision, and software for automated individual tracking and pose estimation is available [38]. However, applying these to footage with heterogeneous backgrounds, variable light environments and arenas larger than the camera's field of view remains a challenge. To overcome this, we combined computer vision with subsequent human evaluation, permitting a low-cost solution to efficiently analyse footage recorded across heterogeneous lighting conditions.

These data revealed that H. heurippa males were no more likely than H. t. linaresi males to interact with the H. heurippa female when it was poorly lit (lower 50% quantile of illuminance values; figure 2c), but were 1.3 times (i.e. 0.57/0.44, electronic supplementary material, table S6) more likely to interact with the H. heurippa female in brighter light (upper 50% quantile of illuminance values; figure 2d). Similar effects of the light environment have been observed in aquatic organisms. For example, the attractiveness of red colour during courtship in fish is dependent on the spectrum of available light, influenced by water depth and turbidity, and recognition of conspecifics may fail entirely under certain light environments [7,39,40]. However, far fewer studies have directly tested how changes in the abiotic environment influence mating preference behaviours that contribute to population divergence in terrestrial organisms (but see [41–43]).

The mechanism underlying the differences we observe under contrasting lighting conditions is not immediately clear. Insects, including butterflies [44], have frequently evolved colour constancy across light environments [45,46], and we would expect individuals to distinguish H. heurippa and H. t. linaresi patterns under different lighting conditions. Male attraction to female colour patterns might to some degree be ‘wavelength-specific’, where triggering of behaviours depends greatly on an object's emission of specific wavelengths and their intensity [46]. Under this scenario, a colour cue might only trigger a behaviour when presented at intense irradiance from the relevant part of the spectrum [46–48]. Modelling of the Heliconius visual system predicts their red pattern elements to be more conspicuous to conspecifics in bright sunlight [23], and, in general, we expect radiance and contrast of H. heurippa's red band to increase with the intensity of light at long wavelengths. We were unable to identify feasible methods that would allow us to capture spectral data, with second-precise resolution, across 100 s of hours of observation in our outdoor insectaries. If suitable methods can be identified, future work could combine behavioural data with spectral irradiance measurements. Nevertheless, because most of our data were generated under the same broad light habitat class (i.e. ‘open/cloudy’ sensu [9], see electronic supplementary material, methods and results), our illuminance measurements can be used as a proxy for irradiance levels at long wavelengths (and correcting for deviations from this class did not qualitatively change our results).

Regardless of the proximate mechanism, our data indicate that prevailing light conditions influence visual mating behaviours in Heliconius. Unlike other examples [39,40,49], the effects we see here act across very short time and spatial scales, reflecting the fluctuating sensory conditions experienced by our study species in their rainforest environments. By contrast to more divergent Heliconius species pairs [50,51], H. heurippa and H. t. linaresi are not known to differ in ecology or habitat type, and there is little reason to suspect that differences in their sensory environment drive divergence in mating preferences, or sensory behaviours more broadly. Instead, differences in female wing pattern (driven by local selection for aposematism) likely impose divergent sexual selection on male preferences to improve their ability to find receptive females [16]. However, by increasing both within and between individual variation, rapidly changing light conditions may dampen the contribution these divergent preferences make to reproductive isolation, though this may be mitigated if females choose to display in microhabitats more suitable for signalling. As such, the extent to which divergent preferences contribute to speciation may depend on individuals' ‘choice’ of signalling location (for example between dimly and brightly lit patches).

Heliconius butterflies bask in the sun with their wings open, particularly during the morning hours [52]. This may primarily be for thermoregulation, though other species are known to display in environments where they are most conspicuous. Anolis lizards, for example, occupy microhabitats in which their dewlap colour is most conspicuous [53]. Visual modelling indicates Heliconius red patterns are more conspicuous to avian predators when presented in bright sunlight, suggesting sun-basking could enhance aposematism [23]. It is also conceivable that virgin females choose basking sites to reduce the chance of costly heterospecific matings [54]. Notably, such behaviours could evolve by substituting the same alleles in otherwise diverging populations (i.e. via Felsenstein's [12] ‘one-allele mechanism’), which by evading the homogenizing effects of recombination is especially conducive to speciation with gene flow [12]. Whatever the ultimate reason for aggregating in sun-exposed patches, our data suggest that these behaviours could enhance the strength of divergent mating preferences and hence might contribute to population divergence.

Key to establishing a link to population divergence is testing whether divergent preferences translate to assortative mating. The large number of trials in our experiment with live females allowed us to detect positive assortative mating, albeit at low levels, though this was much stronger for the 60 trials in which H. heurippa males mated (figure 3). Unfortunately, we do not have illuminance data for these experiments, which would be difficult to measure given the movement of females. Considering that the live females may have been basking in the sun, that overall activity of Heliconius is highest when it is sunny [55] and that mating in Heliconius occurs more frequently on sunnier days (M. Linares 1981–present, personal observation), it is likely that a large proportion of the trials with mating outcomes included sun-exposed females. When accounting for the more frequent involvement of H. heurippa males in mating, the mating rates between H. heurippa and H. t. linaresi in our tetrad experiments closely match estimates derived from our experiments with dead mounted females in bright light. Although we cannot entirely rule out a role for other modalities (e.g. olfaction [56]), differences in visual attraction alone are sufficient to explain the levels of assortative mating we observe.

Finally, it is difficult to ascertain patterns of genetic dominance for visual attraction from our data. This is not surprising given that the differences in behaviour between populations are subtle and shaped by considerable variation. However, segregation of optix alleles, which control red pattern elements in Heliconius [18], did allow us to test for linkage between the warning pattern cue and the corresponding behaviour. Once again, we observed illuminance-induced shifts in visual attraction, so that the two types of backcross to H. t. linaresi males (i.e. red/Bb versus non-red/bb) differed in behaviour, but only under higher illuminance at the H. heurippa mounted female (figure 4). These results are consistent with physical linkage between behavioural loci and optix, as has been shown elsewhere—but with much greater effects—between the closely related species H. cydno and H. melpomene [16]. Physical linkage will help to maintain genetic associations (i.e. LD) between loci underlying ecologically relevant traits and those for premating isolation, facilitating speciation with gene flow in general [10,12], and hybrid trait speciation [57] in particular. Although the differences between backcross to H. t. linaresi genotypes account for approximately 50% of that observed in the two parental taxa, in absolute terms the effects are unlikely to permit sufficient power for a formal QTL study, and caution should be exercised when interpreting these results. Our results are consistent with a simple genetic mechanism by which behavioural alleles were acquired alongside red colour pattern alleles through introgression from H. melpomene into the ‘heurippa/timareta/cydno’ lineage [57]. However, it seems unlikely that the behavioural differences we observe here would rapidly lead to strong reproductive isolation, as predicted under a model of hybrid trait speciation [57,58].

In conclusion, our results reveal that the degree to which Heliconius warning patterns contribute to premating isolation may depend on local illuminance, which can change rapidly in both time and space. The behavioural differences we observe for H. heurippa and H. t. linaresi are similar in strength to those reported elsewhere for Heliconius taxa at the early stages of divergence (e.g. [36]), and are increased when female patterns are in bright light. Alone, these may represent only weak barriers to gene flow, but by augmenting divergent ecological selection acting on the warning pattern cue, they may facilitate the accumulation of additional barriers as speciation proceeds. Nevertheless, by increasing both within and between individual variation in mate preference, changes in light environment may constrain the evolution of premating isolation, though this may depend on ‘choice’ of display site. Traits predominantly shaped by ecology frequently act as mating cues, which by coupling divergent natural selection to premating isolation can promote speciation with gene flow [3]; our results highlight that this effect may depend on the sensory conditions during which these ecological mating cues are encountered, and that premating isolation may be strengthened by the evolution of behaviours that enhance signalling integrity.

Data accessibility

Behavioural data, methods, figures and tables and analysis scripts are included as electronic supplementary material. Behavioural data are also available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.pc866t1ng [59].

Authors' contributions

A.E.H. and R.M.M. conceived the research, with input from M.L., C.P.-D. and C.S. A.E.H. and C.-Y.K. wrote the motion-detection software, and designed the experiments with R.M.M. A.E.H., M.F. and N.R.-M. raised butterflies and performed behavioural experiments. A.E.H. analysed the data. M.L., C.P.-D. and C.S. established butterfly stocks from the wild and provided substantial logistical support. M.L., C.P.-D., C.S. and R.M.M. secured funding, contributed resources and provided supervision. A.E.H. and R.M.M. wrote the manuscript with input from all the authors.

Competing interests

We declare we have no competing interests.

Funding

N.R.-M., C.S. and C.P.-D. were funded by Universidad del Rosario (grant no. IV-FGD005). M.L. was partially funded by the Faculty of Natural Sciences at Universidad del Rosario. A.E.H., C.-Y.K., M.F. and R.M.M. were supported by an Emmy Noether fellowship and research grant awarded to R.M.M. by Deutsche Forschungsgemeischaft (DFG) (grant no. GZ: ME 4845/1-1).