Field Sampling.

The 2 focal populations were Matoury, French Guiana (4.897°, −52.350°), where frogs have white stripes, and the Kaw Mountains (4.542°, −52.158°), French Guiana, where frogs have yellow stripes, hereafter referred to as white and yellow, respectively. These sites were chosen because frogs have the same pattern (stripes on a black body with blue legs) but differ in the color of their stripes (white and yellow) (Fig. 1A). We collected frogs from white (n = 10) and yellow (n = 8) populations in 2013 and 2014. Notably, we initially collected 2 frogs from the Kaw Mountains in 2013, which were used for molecular analysis, but to increase sample size for alkaloid variability, we collected 6 additional frogs in 2014. Frogs were killed by deep-freezing to avoid possible contamination from killing agents (58). Livers were preserved in 95% ethanol, and whole skins were preserved in 100% methanol for subsequent alkaloid analysis. Skins were stored in 4-mL glass vials with PTFE caps.

Gene Flow between Populations.

To assess gene flow between populations, we examined thousands of variable sites (SNPs) across the D. tinctorius genome that could then be used to estimate population demography. To do this, we employed the 3RAD method of refs. 59 and 60. Samples were digested in a reaction that contained 100 ng of DNA, 20 units of each of 3 restriction enzymes (XbaI, EcoRI-HF, and NheI-HF; New England Biolabs), 1× NEB Cutsmart Buffer, 1 µL of both a forward and reverse double-stranded adapter at 5 µM and water to 15 µL total volume. Enzymes were chosen based on an in silico digest of the genome of Xenopus laevis, which suggested these enzymes would produce ∼3,500 unique loci within the desired size range of 300 to 350 bp. Each adapter contained a 6- to 9-bp barcode and each sample was assigned a unique forward and reverse barcode combination. Digestions were carried out at 37 °C for 1 h and followed immediately by the addition of ligation mix. As adapters were already present in digestion reactions, the ligation mix consisted of 1.5 µL of 10 mM rATP (New England Biolabs), 0.5 µL 10× ligase buffer (New England Biolabs), 100 units T4 DNA Ligase (New England Biolabs), and water to 5 µL. Ligation reactions were incubated for 2 cycles of 22 °C for 20 min and 37 °C for 10 min followed by 20 min at 80 °C. Ligated samples were then cleaned by mixing with 1.2 volumes of Sera-mag Speedbeads (Fisher Scientific; prepared as in ref. 61), 2 washes of 70% ethanol, and eluted from beads with 20 µL of IDTE pH 8 (Integrated DNA Technologies). Samples were then amplified in 20-µL reactions with 10 µL of postbead template, 1× HIFI Buffer (Kapa Biosystems), 0.75 µL dNTPs (Kapa Biosystems), 0.5 µL HIFI DNA polymerase (Kapa Biosystems), 0.5 µM iTru5 and iTru7 primers containing sample-specific 8-bp barcodes, and H₂O to 25 µL total volume. Library amplification began with 3 min at 95 °C followed by 16 cycles of 98 °C for 20 s, 55 °C for 15 s, and 72 °C for 30 s, followed by a 5-min extension of 72 °C.

Success of individual PCRs was determined by gel electrophoresis. When digestion, ligation, and amplification were successful, 10 µL from each PCR were pooled into sets of 24. Each pool was concentrated down to 30 µL using a Qiagen PCR purification column and then run in a single lane of a 1.5% Pippin Prep gel cassette (Sage Science) selecting for fragments sized 400 to 450 bp. Selected sized fragments were removed from Pippin elution wells and used directly in qPCR library quantification (Kapa Biosystems) and Illumina Sequencing on a NextSEq. 500 to obtain single-ended, dual-indexed reads of 75 bp. Libraries were sequenced in the National Center for Natural Products Research, University of Mississippi.

Raw reads were downloaded from Basespace using basespacerundownloader (Illumina) and demultiplexed using BCL2FASTQ (Illumina) allowing up to 2-bp errors in barcodes as our barcodes had 3-bp degeneracy built in. After demultiplexing, each sample was trimmed from the beginning of each read to remove remaining internal (adapter) barcode and restriction site overhang. All samples were then trimmed to 58 bp in length to prevent downstream difficulties in assignment of homology. Loci were identified within individuals and aligned among individuals using pyRAD (62).

Population Genetic Analysis.

We used the SNP data with the Bayesian coalescent program G-PhoCS (Generalized Phylogenetic Coalescent Sampler) (26) to estimate rates of migration between the 2 populations. Six independent, simultaneous runs were conducted, each with a 500,000 Markov chain Monte Carlo iterations, which were then compiled for a total of 3,000,000 iterations. We used a 10% burn-in independently for each run, which resulted in 2,700,000 iterations from which summary statistics were calculated. We defined a constant mutation rate in order to conduct these analyses as a constant mutation rate would be a null assumption when lacking data, suggesting fluctuating mutation rates over evolutionary history (26). Bidirectional migration rates between the 2 populations were then calculated in G-PhoCS. We examined population structure between the 2 populations using snmf function in the LEA R package (27). We tested from 1 up to 10 populations and ran 1,000,000 iterations.

Avoidance of Novel Phenotypes.

Plasticine clay model experiments are commonly used to assess predator responses to novel aposematic signals (e.g., refs. 29 and 63⇓–65). These allow observation of predation attempts and provide insight into selection pressures on various phenotypes. We constructed 45-mm-long (snout to vent) replica frogs using Van Aken polymer modeling clay (29, 66), which does not harden and retains evidence of predation attempts. We poured melted clay into silicone molds made from plastic model replicas (67). In order to assess how predators responded to novel colors and patterns, we created models as follows. In the white population, we deployed models with the local color (white), that could have either the local pattern (stripes) or a new pattern (solid white), and also individuals with the local pattern (stripes), but a different color (yellow). Conversely, in the yellow population, we had a novel color (white) with the local pattern (stripes), but also the local color (yellow) both with the local pattern (stripes) and a novel pattern (solid yellow). All models had blue legs (as do both populations sampled), and the stripes were placed on a black body. Both dendrobatid frogs and the polymer clay have been shown to lack UV reflectance (28). Both stripes and eyes were created with clay and affixed to models (63).

Models were placed along transects on which 3 model types were randomized with 1 model every 5 m (29, 68). The 3 model types for each location were local color and pattern, local color and novel pattern, local pattern and novel color, ensuring that at least 1 component of the aposematic signal was familiar to predators. Transects were separated by at least 100 m. We placed 7 transects from 495 m long to 1.5 km long in the white population and 13 transects from 495 m long to 1.5 km long in yellow population. Models (1,378 and 1,136 in the yellow and white populations, respectively) were left for 72 h to allow for predation attempts (29, 63). Following this period, we collected models and determined which ones had been attacked by avian predators. Avian attacks were recognized by a characteristic U or V shape, as well as stab marks. We focused on attacks by avian predators because they are likely the most important in driving phenotypic diversity in conspicuous signals given their ability to discern different colored phenotypes (69, 70). The nonavian organisms that attack plasticine models, namely arthropods and mammals, are also capable of seeing at least differences in brightness, but mostly forage following chemical cues, making the biological interpretation of their attacks difficult (71, 72). Similarly, snakes, which are some of the few known predators of poison frogs (73, 74), can also see color, but they are highly motivated by movement and chemical cues, making it highly unlikely that they would attack clay models.

Clay Model Data Analysis.

If models had multiple bite marks, they were scored as a single predation attempt as we were not able to determine whether 1 predator attacked multiple times or multiple predators attacked once (29). Missing models were excluded from the analysis (43 of 1,136 and 57 of 1,378 for white and yellow, respectively). We used a general linear mixed model (GLMM) with binomial distribution to compare predation attempts among phenotypes (novel/local pattern, novel/local color) at both sites to determine whether aposematic signal is a predictor of predation risk, using transect ID as a random factor. Thus, we used color novelty (y, n), pattern novelty (y, n), and population (w, y) as independent variables, and tested for their effect on whether each model was attacked (1) or not (0). Therefore, local or novel traits to the avian community are what are notable in this analysis rather than individual clay model phenotypes.

Avoidance Learning and Generalization.

As avian predators are the likely drivers of aposematic signal evolution in D. tinctorius (29, 30, 63), we used a model avian predator, the domestic chicken (Gallus gallus domesticus), to assess how naïve predators learn and extend experience (generalize) with aposematic signals. Chickens are well-known for their capabilities of recognizing and learning different colors quickly (75), and thus widely used in this type of experiment (33, 37, 76⇓⇓–79). We trained naïve 1-wk-old chickens to eat mealworms (Tenebrio molitor) from a Petri dish with an image of a frog beneath the dish. Frogs were printed illustrations that were brown (control) or had white or yellow stripes on a black body and blue legs, similar to the clay models (SI Appendix, Fig. S3). White and yellow colors were taken from photographs of frogs in the white and yellow populations, respectively, using the color dropper tool in Photoshop. Unmodified mealworms were served in association with the control, brown frog, whereas mealworms associated with white- or yellow-striped frogs were tainted with a distasteful chloroquine solution. Using this design, we measured predator-learning rates for white-striped and yellow-striped frogs and then explored their response to similar, but novel aposematic signals. To make mealworms distasteful, we soaked them in a chloroquine diphosphate (98%, Arcos Organics) solution, which is an alkaloid and known to be distasteful, but not harmful, to birds (80, 81). Over successive trials, we examined whether chicks would learn to avoid distasteful mealworms associated with an aposematic signal, with learning defined as 3 consecutive refusals of a distasteful mealworm (82). Once a chick learned to avoid a particular aposematic signal, it was presented with the other signal (e.g., trained to avoid yellow stripes, presented with white stripes). We further examined the effect of distastefulness level on learning by varying levels of distastefulness by training chicks with mealworms that had been soaked in either a 5% or 10% chloroquine solution for 1 to 3 h. Chicks were equally divided (n = 15 per treatment; 60 chicks total) into 4 different treatments: 5% chloroquine with a yellow signal, 5% chloroquine with a white signal, 10% chloroquine with a yellow signal, and 10% chloroquine with a white signal. Thus, we were able to explore avoidance in the context of associated color or distastefulness (chloroquine concentration). The trials were done in 30-cm × 60-cm wooden compartments under full spectrum lights. Chicks were placed individually into a compartment and allowed to habituate for 2 h. Chicks were food-deprived during this acclimation period to ensure motivation to feed during the trials. Once in the compartment and after the 2-h acclimation period, training consisted of teaching the chicks to eat a dried mealworm from a Petri dish on top of an illustration of a brown frog on a tan background. The training phase was completed once the chick had eaten 3 consecutive times, after which we allowed the birds to rest for a period of 5 min.

The avoidance-learning trials consisted of the consecutive presentation of mealworms on a Petri dish on top of an illustration of D. tinctorius on a tan background. (i.e., a frog with either white or yellow dorsal stripes) (SI Appendix, Fig. S3). We recorded the latency (i.e., the time until the chick approached and picked up the mealworm) and noted any behavioral reaction to the distastefulness of the mealworm after tasted or eaten. Such behaviors most often involved beak wiping and head shaking. Each trial ran for 5 min, followed by 5-min rest, after which a new Petri dish with an unpalatable mealworm was offered. This procedure was repeated until the chick refused to eat the mealworm over 3 consecutive trials, at which point the chick had learned the signal. The test ended either when chicks “learned” the signal or proceeded through 10 trials, whichever came first. If chicks proceeded through 10 trials without 3 consecutive refusals, they were considered to not have learned the signal. When a chick did not eat the chloroquine-soaked mealworm on the D. tinctorius pattern, a palatable mealworm was offered on the neutral (brown frog) background, which the chick did not associate with an unpleasant experience. In that way, we made sure that the chicks had refrained from eating the treated mealworm because of an association with the warning signal, rather than satiation. In 5 cases, chicks stopped eating and subsequently refused the palatable mealworms, and as a result were excluded from analysis. In these cases, we tested additional chicks to ensure each treatment had 15 replicates. We registered the number of trials that it took for each chick to learn to avoid an aposematic (white or yellow) signal.

The next step was to run a generalization trial, which aimed to test whether, once a signal was learned, the aversion would be extended to other signals. Therefore, after the 3 trials in which the chick would refrain from eating the presented mealworm, chicks were given a 5-min rest, and then a new unpalatable mealworm was presented in association with the other (white/yellow) aposematic signal. The unpalatable mealworm had the same chloroquine concentration on which the chick was originally trained. Chicks that had learned to avoid yellow-striped frog images were presented with a mealworm atop a white-striped frog image, and vice versa. We recorded the chicks’ response both as a binary variable (whether the mealworm was eaten or not) and the hesitation time. Generalization trials were only conducted on chicks that learned to avoid a signal in the first set of trials. Chicks were considered to have generalized avoidance to the novel signal if they avoided the first presentation of the novel signal. This situation best simulates choices wild predators that have learned to avoid a locally common aposematic signal would make when encountering a novel signal. All trials, both avoidance learning and generalization, were filmed in order to extract details on the chick’s behavior afterward. See SI Appendix, Fig. S3 for a schematic on how the learning and generalization experiments were conducted.

Learning Experiment Data Analysis.

We analyzed the data from the learning experiments using a GLMM in 2 ways. First, we used a survival analysis (Cox regression) to examine the effect of color and chloroquine content (5% vs. 10%), and the interaction between the 2, on the latency to attack (i.e., time to event) each mealworm. Chick ID was included in the model as a random factor to account for repeated presentations to the same individual. Then, for each chick, we counted the number of trials in which the mealworm was attacked, to a maximum of 10, and whether or not the chicks learned to avoid the signal they were presented (as defined by 3 consecutive refusals). In the first case, we used a Poisson error distribution, whereas in the second we used a binomial distribution. In both cases, our predicting variables were color, chloroquine concentration, and the interaction between the 2. Finally, we used a last GLMM to test whether the chloroquine concentration and the color of the signal learned predicted whether or not the chicks would generalize their learned aversion to the alternative signal. All analyses were done in R (83), with the RStudio interface and using the packages lme4 (84) and coxme (85).

Population Variability of Alkaloids.

Whole frog skins were collected from 18 specimens (8 yellow, 10 white) and stored in 4 mL of methanol. Alkaloids were extracted and analyzed from skins using the procedure outlined in ref. 34, which is described here briefly. One milliliter of each methanol extract was transferred into a 10-mL conical vial and 10 μg of nicotine as an internal standard [(-)-nicotine ≥ 99%, Sigma-Aldrich] and 50 µL of 1N hydrochloric acid (to acidify the solution) were added. Each sample was then mixed and evaporated with nitrogen gas to a volume of 100 µL. Subsequently, each sample was diluted with 200 µL of deionized water. The samples were then extracted with 4 300-µL portions of hexane. The resulting hexane (organic) layer was disposed of and the remaining aqueous layer was basified with saturated sodium bicarbonate. Basicity was tested with pH paper. Once the pH was greater than 7, each sample was extracted with 3 300-µL portions of ethyl acetate. Anhydrous sodium sulfate was added to the newly extracted mixture to remove any excess water. The remaining samples were carefully evaporated to dryness with nitrogen gas. Alkaloid fractions were resuspended in 100 µL of methanol and stored at −20 °C.

Gas chromatography-mass spectrometry (GC-MS) was performed for each individual frog on a Varian Saturn 2100T ion-trap MS instrument, which was coupled to a Varian 3900 GC with a 30-m × 0.25-mm inner diameter Varian Factor Four VF-5-ms fused silica column. GC separation of alkaloids was achieved using a temperature program from 100 to 280 °C at a rate of 10 °C per minute with helium as the carrier gas (1 mL/min). Each alkaloid fraction was analyzed in triplicate with electron impact MS and once with chemical ionization (CI) MS with methanol as the CI reagent.

Individual alkaloids of D. tinctorius were identified based on comparison of MS properties and GC retention times with those of previously reported alkaloids in dendrobatid frogs (86). Alkaloid quantities for each individual frog were calculated by comparing the average observed peak area of individual alkaloids to the average peak area of the nicotine standard from the triplicate electron ionization-MS analyses using a Varian MS Workstation v.6.9 SPI. Only alkaloids that were present in quantities of ≥0.5 μg were included in the analyses (34). We compared average alkaloid quantity between populations with a GLM with Gaussian distribution using the frog’s body size as a covariate. Differences in alkaloid composition among individual frogs from each population were graphically visualized using nonmetric multidimensional scaling (nMDS), and statistical differences were examined with a 1-way analysis of similarity (ANOSIM). nMDS and ANOSIM statistics were based on Bray–Curtis similarity matrices, and were performed in PRIMER-E (v5).

Predator Response to Chemical Defenses.

Prior research into dendrobatid alkaloids has largely quantified toxicity through either LD50 (31) or subcutaneous injections (32, 33) (but see refs. 34 and 35 for assays on alkaloid unpalatability using arthropods as predators), which, while informative, are limited in their evolutionary significance, as predators experience alkaloids through ingestion, not injection into the bloodstream or musculature (38). In other studies (33, 37), naïve chickens have been offered live frogs, which does not allow investigators to account for interindividual variation in toxin profiles and alkaloid amounts. This is because the amount of alkaloids in living frogs cannot be known and, thus, the most aversive response may be assumed to result from the highest amount of alkaloids. Consequently, we sought to examine unpalatability of skin extracts, as this is what drives predator decision making, in a controlled manner. We used an additional 1 mL of the methanol extract from the same skins of the 18 assayed frogs (10 white, 8 yellow) and prepared toxins for 2 unpalatability assays using wild-caught blue tits (Cyanistes caeruleus), which have been shown to be able to respond to subtle variations in animal chemical defenses (87, 88). These methanol extracts were not fractionated and were simply whole extracts of skin contents.

We assessed skin contents in 2 ways. First, we took equal proportions of methanol extracts (unpalatability assay A). This represents how predators respond to natural variation among individuals. Our second assay (unpalatability assay B) sought to examine how skin content composition affected predator response by controlling for dry weight of the skin contents. This aids in inferring how different alkaloid profiles coupled with nonalkaloid content impact predator response. We note, however, that predator responses could be skewed if nonalkaloid content varied among individuals (see below). The 2 different assays were identical except for the skin secretion toxin preparation. The purpose of doing 2 assays was to attempt to 1) understand how toxins influence predator response and 2) ensure that nontoxin components of skin secretions do not disproportionately affect results and mask potential aversion to the alkaloids. For each assay, we tested 1 frog sample with 1 bird (white, n = 10, yellow, n = 8), with 6 control birds for the first assay and 7 control birds for the second assay.

In unpalatability assay A, we diluted each methanol extract equally, evaporating 1.0 mL of methanol to dryness under N2, and then reconstituted with 0.5 mL ethanol regardless of dry mass. We added 15 µL of the reconstituted sample to each oat and allowed the oats to dry. These oats would then be presented to birds. While this is a more realistic scenario of how birds would respond to natural variation among these frogs, we sought to control for dry mass in an effort to examine how composition affected response. Thus, we conducted a second assay that controlled for dry mass.

For unpalatability assay B, we evaporated 1 mL of methanol extracts to dryness and weighed to determine the approximate quantity of skin content present for each sample (notably, this included everything in the methanol, such as mucus, cholesterol, fatty acids, carotenoids, and so forth, which, while not necessarily defensive in nature, their presence may impact the efficacy of distasteful alkaloids). Samples were then reconstituted in a volume of ethanol such that the concentration of toxin for each sample was the same based on the mass of the dried sample (approximately 1:1 toxin mass to ethanol). Toxins extracts were then transferred to oats (15 µL per oat) and allowed to dry. This was in an effort to control for the quantity of toxins present in the extracts and not allow individual variation in alkaloids quantity skew results. However, if the mass of the nonalkaloid content varied among frogs, especially in a manner independent of alkaloid mass, this may have unintentionally altered our concentrations among samples, eliciting nonbiologically relevant behaviors among the birds.

Prior to experiments, blue tits were trained to eat untreated oats. After training, birds were given oats to which toxins had been added following the protocol described in Burdfield-Steel et al. (88). Aversive behavior (i.e., beak wiping) and percentage of oat eaten were recorded.

For each assay, each of 2 oats were soaked with 15 µL of extract of 1 frog skin and left for 24 h at room temperature to ensure that all ethanol had evaporated. Two other oats were soaked—and subsequently allowed to dry for 24 h—with 15 μL of pure ethanol and used at the beginning and end of the experiment with each bird. The first ethanol-only oat needed to be consumed entirely by the bird before the experiment could be initiated, to ensure motivation to eat; the second ethanol-only oat was offered in the final trial to ensure that the birds were not refusing to eat the oats coated with toxins due to satiation or lack of motivation to eat in general. Birds in the control treatment received oats soaked with pure ethanol for all trials in order to compare directly the response of birds to oats containing frog toxins versus oats with ethanol only.

Each oat (1 at a time) was presented on a hatch that had a visual barrier, which allowed us to detect the exact moment at which the oat was seen, which set the actual beginning of the trials. We measured the number of times the bird wiped its beak, which is a known aversive behavior (89), and the percentage of the oat eaten. Birds were watched for a 2-min period after they finished eating the oats, or for a maximum of 5 min in those instances in which the oat was not fully eaten, to make sure that any delayed response to the oat taste would not be missed. Data were analyzed using GLMM using the package lme4 (84). We entered frog population as the predicting variable, while number of times the beak was wiped and percentage of oat eaten were entered as the response variables. Because the response of each bird was measured twice, we included bird ID as a random factor. Given that the duration was not the same for all trials, we used the function “offset” in R to account only for the effect of population on our response variable once the effect of duration was controlled for. These and all other statistical analyses were done in R (83) using the RStudio interface (90), unless stated otherwise.

Ethics Statement.

Chick and blue tit experiments were conducted under University of Mississippi Institutional Animal Care and Use Committee protocols 14-025 and 14-026 and the Central Finland Centre for Economic Development, Transport, and Environment and license from the National Animal Experiment Board (ESAVI/9114/04.10.07/2014), and the Central Finland Regional Environment Centre (VARELY/294/2015), respectively.

Data Availability.

Data supporting the findings reported in this study are available from the data repository of the University of Jyväskylä (DOI: 10.17011/jyx/dataset/65263).