Geographic mosaic of selection by avian predators on hindwing warning colour in a polymorphic aposematic moth

Warning signals are predicted to develop signal monomorphism via positive frequency-dependent selection (+FDS) albeit many aposematic systems exhibit signal polymorphism. To understand this mismatch, we conducted a large-scale predation experiment in four locations, among which the frequencies of hindwing warning coloration of aposematic Arctia plantaginis differ. Here we show that selection by avian predators on warning colour is predicted by local morph frequency and predator community composition. We found +FDS to be strongest in monomorphic Scotland, and in contrast, lowest in polymorphic Finland, where different predators favour different male morphs. +FDS was also found in Georgia, where the predator community was the least diverse, whereas in the most diverse avian community in Estonia, hardly any models were attacked. Our results support the idea that spatial variation in predator and prey communities alters the strength or direction of selection on warning signals, thus facilitating a geographic mosaic of selection.


Predation experiment 116
To estimate the attack risk of white, yellow and red hindwing colour morphs by local 117 predators in the wild we used artificial moth models, resembling real A. plantaginis 118 morphs. Models with plasticine (Caran D'Ache Modela 0259.009 Black) bodies 119 attached to printed waterproof (Rite in the Rain ©, JL Darling Corporation, Tacoma, 120 WA, USA) paper wings were prepared following methods described in Nokelainen et 121 al. (2014). Models were constructed using pictures of one white moth hindwing and 122 two forewings, one with a typical European pattern and another with a typical 123 Caucasian (Georgian) pattern, which were copied and assembled in GIMP 2.8.16 124 SOFTWARE (GNU Image manipulation program) to create six models representing 125 the white, yellow and red morphs in Europe and Georgia ( Figure 1B). A locally 126 common forewing type was used to reduce potential novelty effect caused by the 127 forewing pattern (Hegna & Mappes 2014). Resemblance of the artificial models to the 128 real moths was verified by taking measurements of reflectance from the black and 129 coloured areas of real moth wings and printed wings with a Maya2000 Pro spectrometer 130 (Ocean Optics) using a PX-2 Pulsed Xenon Light Source (Ocean Optics) for 131 illumination and adjusting the model colours with Gimp (2.8.16) to match the natural 132 wing colour as closely as possible with a calibrated (HP Colour LaserJet CP2025) 133 printer (see spectral curves of hindwing colour in Rönkä et al. 2018), where identical 134 models were used). As our study focused on the hindwing coloration, all other variables 135 such as wing size and pattern were kept constant. 136 We set up 60 predation transects across the four study populations (15 in each 137 country) in open, semi-open and closed natural habitats where the wood tiger moth and 138 its potential avian predators were known or presumed to occur. The predation transects 139 were set at least 500 m apart to avoid birds having overlapping territories between the transects. Along each 900 m transect 20 white, 20 yellow and 20 red artificial moth 141 models were set individually every 15 meters using a randomized block design, so that 142 two models of the same colour would never be next to each other. Models were pinned 143 directly on natural vegetation, either to green leaves large enough to support their 144 weight, or to tree trunks, as visibly as possible. All models were left in the field for a 145 maximum of 6 days (2-6 days, 4 days on average), during the A. plantaginis flight 146 season in 2014 (May 31st -July 6th in Estonia, May 26th -July 6th in Finland, June 147 15th -July 30th in Scotland and July 12th -August 3rd in Georgia). Predation events 148 were recorded every 24 hours except for days of heavy rain (as birds were likely not 149 active). For practical reasons (i.e. accessibility of mountain roads and weather 150 conditions) the protocol was modified in Georgia. The 20 white, 20 yellow and 20 red 151 models were set every 10 m totalling up to 600 m, left in the field for 3 consecutive 152 days (72 h), and checked only once. 153 Attacks were recorded based on imprints on the plasticine body and fractures in 154 the wings (see Supplemental Experimental Procedures). Only clear avian attacks were 155 included in the analyses (Supplemental Table 1). Missing and attacked models were 156 replaced with a new model of the same colour to ensure constant morph frequency 157 during the experiment. Excluding or keeping consecutive attacks on the replaced 158 models in the analyses did not markedly change the outcome, reported here (Table 1) 159 for the dataset including replaced models (4004 observations) and for the dataset 160 including original models only (3600 observations; in Supplemental Table 2). 161 Therefore, we kept the replaced models in for all of the analyses, as it increased the 162 sample size. 163

Measures of predator community 166
To estimate the abundances of different insect-feeding birds, which are the most likely 167 predators of wood tiger moths, we counted birds belonging to the orders Passeriformes 168 and Piciformes (Supplemental Table 3). These counts were done once, either before or 169 during the predation experiment, along the predation lines using a modified transect

Statistical analyses 177
To investigate how local predator community affects the direction and strength of 178 selection on wood tiger moth morphs, we constructed generalised linear mixed models. 179 Because the artificial moths were presented to predators over a different number of days 180 in each transect, the attack risk (attacked or not) within a day exposed was used as the 181 response variable for all analyses, modelled with a binomial distribution and a logit link 182 function. First, we tested whether predators select for wood tiger moth warning colours 183 in a frequency-dependent manner across populations ( Figure 2). For this, we used local 184 morph frequency calculated from field monitoring data and its interaction with colour 185 morph as the explanatory variables in Model 1 (Table 1). Transect ID, nested within 186 country, was set as the random factor to account for the nested spatial structure of the 187 study design. 188 To test for predator community composition effects, the dimensions of the bird 189 count data, consisting of 12 genera, was first reduced with a principal component analysis using the R function 'princomp'. To avoid overparameterization, the main 191 effects of the first three resulting components (explaining 44.7 %, 33.7 % and 8.5 % of 192 the variation in predator community), and their three-way interactions with morph 193 colour and country, were included one by one as explanatory variables in three separate 194 GLMMs (Table 2). Country was included as an explanatory variable to test for local 195 differences in selection and thus transect ID alone was set as a random effect to each Altogether, we observed a total of 718 bird attacks on the 4004 artificial moths. The 207 relative attack risk of each colour morph was lower when the natural frequencies of the 208 respective morph were higher in relation to the others (Table 1, Figure 2). Also, the 209 morphs with intermediate local frequencies show corresponding levels of attack risk 210 ( Figure 2). This effect did not depend on colour morph itself (Table 1), as expected if 211 the local predator avoidance depends more on local morph frequency than on morph 212 colour. 213

Predator community 215
The attacks were not evenly distributed across countries or transects ( Figure 2C). 216 Predation pressure varied between and within countries, being highest in Scotland and 217 lowest in Estonia ( Figure 2C). Georgia had the lowest amount of insect feeding birds 218 observed (2.1 per 100 meters) compared to Finland (2.6), Scotland (4.0) and Estonia 219 (4.4), respectively. Georgia also had the least diverse predator community measured 220 with Shannon-Wiener diversity index, whereas Estonia was most diverse, followed by 221 Scotland (Figure 3). Across countries, the three most commonly observed potential 222 predators included the common chaffinch, the willow warbler (replaced by green 223 warbler in Georgia) and the great tit (Supplemental Table 3), the latter of which was 224 observed to attack the artificial moths. The first three principal components (PC1, PC2 225 and PC3) that explained 44.7 %, 33.7 % and 8.5 %, respectively, captured 87.0 % of 226 variance in the predator community data. PC1 was dominated by Sylvidae (warblers), 227 Fringillidae (finches) and Muscicapidae (flycatchers), which loaded in the negative end, 228 whereas the positive end of the axis was loaded with Paridae (tits) (Figure 3). PC2 was 229 dominated by Paridae and PC3 with Fringillidae, Muscicapidae and Troglotydidae (the 230 Eurasian wren) (see Supplemental Table 4 for factor loadings). 231 232

Significant association between predator community structure and selection 233
A consecutive analysis, where the effect of predator community on the attack risk of 234 each moth colour morph was addressed, revealed a significant three-way-interaction 235 between moth colour, country and PC1 (Model 2, Table 2a, Figure 3). This significant 236 interaction means that the variation in predator community structure captured by PC1 237 is associated with predation pressure on different colour morphs, but the direction of 238 the association is different between countries (i.e. between local communities). PC2 239 and PC3 were not significantly associated with predation pressure (Table 2C and 2D). All morphs were attacked at equally low levels in Estonia, which implies spatial 288 variation in the strength of selection or even locally relaxed natural selection on the 289 warning signal. The low predation pressure is not explained by a low number of 290 predators, as there were more insectivorous birds in Estonia than in any other study site 291 (Supplemental Table 3 experimental evidence of +FDS in the wild, however, comes from tropical systems 311 the prey and predator community composition is temporally less variable (Mittelbach 313 et al. 2007). In such communities, strong +FDS can lead to very accurate mimicry 314 between warning coloured prey, whereas in more variable conditions, higher levels of 315 variation and polymorphism can be maintained. 316 The paradoxical maintenance of local polymorphism despite +FDS could thus 317 be explained by spatial and temporal variation in morph survival combined with lack of genetic and environmental differentiation between distinct phenotypes. 375 J. Zoolog. Syst. Evol. Res.,54,[127][128][129][130][131][132][133][134][135][136] Bocek