Correlational selection in the age of genomics

Correlational selection in the age of genomics 1 2 3 4 Authors: Erik I. Svensson1*, Stevan J. Arnold2, Reinhard Bürger3, Katalin Csilléry4, Jeremy 5 Draghi5, Jonathan M. Henshaw6,7, Adam G. Jones6, Stephen De Lisle1,8, David A. Marques9, 6 Katrina McGuigan10, Monique N. Simon2,11 and Anna Runemark1. 7 Affiliations: 8 1 Department of Biology, Lund University, SE-223 62 Lund, SWEDEN 9 2 Department of Integrative Biology, Oregon State University, Corvallis, Oregon, USA 10 3 Faculty of Mathematics, University of Vienna, Vienna, AUSTRIA 11 4 12 SWITZERLAND 13 5 Department of Biological Sciences, Virginia Tech, USA 14 6 Department of Biological Sciences, University of Idaho, Moscow, Idaho, USA 15 7 Institute of Biology I (Zoology), University of Freiburg, GERMANY 16 8 Department of Ecology & Evolutionary Biology, University of Connecticut, Storrs, USA 17 9 Eawag, Seestrasse 79, 6047 Kastanienbaum & University of Bern, Baltzerstasse 6, 3012 Bern, 18 SWITZERLAND 19 10 School of Biological Sciences, The University of Queensland, AUSTRALIA 20 11 Department of Genetics and Evolutionary Biology, University of Sao Paulo, BRAZIL Swiss Federal Research Institute WSL, Zürcherstrasse 111, 8903 Birmensdorf, 21 22 *Correspondence to: [email protected] 23 24 This document is the accepted manuscript version of the following article: 1 Svensson, E. I., Arnold, S. J., Bürger, R., Csilléry, K., Draghi, J., Henshaw, J. M., … Runemark, A. (2021). Correlational selection in the age of genomics. Nature Ecology & Evolution, 5, 562-573. https://doi.org/10.1038/s41559-021-01413-3 25 Abstract: 26 27 Ecologists and evolutionary biologists are well aware that natural and sexual selection do not 28 operate on traits in isolation, but instead act on combinations of traits. This long-recognized and 29 pervasive phenomenon is known as multivariate selection, or – in the particular case where it 30 favours correlations between interacting traits – as correlational selection. Despite broad 31 acknowledgement of correlational selection, the relevant theory has often been overlooked in 32 genomic research. Here, we discuss theory and empirical findings from ecological, quantitative 33 genetic and genomic research, linking key insights from different fields. Correlational selection 34 can operate on both discrete trait combinations and on quantitative characters, with profound 35 implications for genomic architecture, linkage, pleiotropy, evolvability, modularity, phenotypic 36 integration and phenotypic plasticity. We synthesize current knowledge and discuss promising 37 research approaches that will enable us to understand how correlational selection shapes genomic 38 architecture, thereby linking quantitative genetic approaches with emerging genomic methods. We 39 suggest that research on correlational selection has great potential to integrate multiple fields in 40 evolutionary biology, including developmental and functional biology, ecology, quantitative 41 genetics, phenotypic polymorphisms, hybrid zones and speciation processes. 42 43 44 45 2 46 Organisms are functionally integrated adaptive systems, where interactions among traits make the 47 whole more than the sum of its parts. How and why did such functional integration evolve, and 48 what are the evolutionary consequences of genetic correlations between traits? These questions 49 have occupied evolutionary biologists for decades, resulting in a rich but scattered scientific 50 literature on topics such as modularity1, evolvability1–3, multivariate selection on trait 51 combinations4–8 and the evolution of genetic correlation structure9–12. Early theoretical work by 52 Cheverud4 and Lande13 predicted that genetic correlations between traits should become aligned 53 with the direction of selection on trait combinations. This important insight made it possible to 54 connect correlational selection (selection on trait combinations rather than traits in isolation; see 55 formal definition in Box 1) to the field of evolutionary quantitative genetics, with its focus on 56 genetic correlation structures. A central testable prediction was adaptive alignment between 57 genetic correlations and the direction of correlational selection, although genetic correlations will 58 also be influenced by other evolutionary forces (e.g. mutation and genetic drift) and ecological 59 factors (e.g. fluctuating environmental conditions)9,10. 60 61 Correlational selection forms a nexus between several traditionally separate research fields, 62 including ecology and developmental biology (Fig. 1). Correlational selection links organismal 63 level features, such as function and development, both to population phenomena such as 64 modularity and genetic correlation structure and to underlying processes such as natural and sexual 65 selection, which typically arise from interactions with mates, predators, mutualists or the abiotic 66 environment (Fig. 2). These connections have not always been developed explicitly, with the result 67 that whole research fields have largely remained separate, partly due to different terminologies. 68 For instance, in a highly influential review about the evolution of modularity1, correlational 69 selection was not explicitly mentioned, and instead the authors used the terms modular selection 3 70 and a modular trait architecture as an expected outcome of selection. Correlational selection can 71 either strengthen or weaken correlations between traits, depending on ecological context. For 72 instance, plant evolutionary biologists studying floral pollination syndromes have noted that 73 mutualistic interactions between pollinators and plants may lead to adaptive de-coupling between 74 vegetative and floral parts, resulting in strong intramodule correlations but weak correlations 75 between modules14. Similarly, antagonistic interactions like predation can impose strong 76 correlational selection on behavioural traits, leading to tighter phenotypic integration and adaptive 77 multivariate phenotypic plasticity in stickleback fish15,16. Studies of the outcomes of artificial 78 selection and domestication processes have also revealed that correlations between animal 79 personality traits have sometimes become decoupled, compared to the ancestors where these traits 80 were more strongly genetically correlated17. 81 82 In light of the genomic revolution, time is now ripe to evaluate Cheverud4 and Lande13’s 83 predictions about the evolution of genetic architecture and to ask: have they been confirmed or 84 overturned by recent findings? In particular, are molecular signatures consistent with correlational 85 selection having shaped the genomic architecture of organisms6,18 and promoting functional 86 integration, e.g. through linkage or pleiotropy? Here, we review quantitative genetic theory and 87 data on correlational selection and link these to the partly separate literatures on modularity and 88 evolvability, as well as to recent genomic research. Our aim is to synthesize insights from these 89 different fields and to point out new directions for future research at their intersections. 90 91 Quantification and visualization of correlational selection 92 93 The first quantitative treatment of correlational selection was provided by Lande and Arnold19 94 (Box 1). These pioneers introduced statistical tools to measure selection on continuously 95 distributed phenotypic traits by estimating selection coefficients that could be incorporated into 4 96 the equations for predicting evolutionary responses. Below we discuss the interpretations of those 97 coefficients, and review the methods to estimate them. 98 99 Individual fitness surfaces are often complex, but can be analyzed to reveal the operation of 100 correlational selection (see definition in Box 1). Correlational selection is particularly likely when 101 the fitness surface resembles a ridge that is not parallel to either trait axis (Fig. 3A), as this form 102 of selection favors particular combinations of trait values over others and thereby selects for a non- 103 zero correlation between traits (Box 1). Correlational selection can also arise alongside disruptive 104 selection, for example when the fitness surface resembles a valley which is not parallel with either 105 trait axis (Box 1; Fig. 3). 106 107 The measurement of correlational selection requires data on the fitness and trait values of multiple 108 individuals (Fig. 3B). The major goals of such analyses are to visualize the fitness surface and 109 estimate coefficients that describe it5. In empirical studies, the true surface is unknown, but we can 110 deduce its properties by approximating the surface with simple functions. Quadratic surfaces are 111 often used to estimate coefficients corresponding to linear selection (β) and nonlinear selection (γ) 112 (Box 1)19. Unfortunately, it is difficult to visualize the fitness surface from the γ-coefficients alone. 113 However, the surface can be visualized by plotting it (Box 1) or by conducting a canonical analysis 114 that estimates the principal components (eigenvectors) of the surface (Box 1)5,7,8. Despite their 115 simplicity, quadratic coefficients can describe a wide variety of surfaces5. 116 117 When a quadratic surface does a poor job of approximating the actual fitness surface, the surface 118 can be visualized using non-parametric methods20. These techniques can reveal multiple peaks and 119 valleys in the fitness surface (Fig. 3), if they exist, whereas the quadratic approaches will always 120 depict a smooth and simple relationship, regardless of the ruggedness of the underlying fitness 5 121 surface. However, non-parametric approaches have the shortcoming that they usually do not 122 produce coefficients that are well-integrated into the equations of evolutionary change. 123 124 Our understanding of the empirical importance of correlational selection has lagged behind our 125 understanding of the prevalence and consequences of directional selection21, with only one meta- 126 analysis of correlational selection published to date22. There are good reasons to expect 127 correlational selection in a wide variety of ecological circumstances, and it might be particularly 128 strong when fitness is affected by biotic interactions, which can generate strong and chronic 129 selection on trait combinations6. Intraspecific interactions that have been shown to result in 130 correlational selection often involve sexual or social selection6. Prime examples include selection 131 on signaling traits such as colour8,23,24 as well as selection on territorial behaviours, which can 132 favor genetic coupling between traits like aggression, dispersal and colonization ability25(Fig. 2). 133 Interspecific interactions linked to correlational selection include predation based on colouration, 134 morphology and behaviour traits15,26, herbivory on plants27 and mutualistic interactions between 135 plants and their pollinators28. In many cases, the fitness surfaces are simple ridges or saddles, but 136 sometimes the surface is more complex. Indeed, complex fitness surfaces could be common20. A 137 priori we might expect to see multiple fitness peaks in organisms with discrete sympatric 138 morphs6,8,26 or between ecotypes29 or newly formed species30. 139 140 Evolution of genetic architecture in response to correlational selection 141 142 Correlational selection is central to our understanding of how genetic architecture evolves. 143 Correlational selection is also closely connected, albeit not identical, to the concept of fitness 144 epistasis in evolutionary genetics31(Box 1). Importantly, although the single-generation effects of 145 correlational selection on the genetic and phenotypic composition are readily understood, the 146 transmission of these changes across generations is a complex theoretical and empirical issue. 6 147 148 To address how the effects of correlational selection are transmitted across generations, we must 149 first define two parameters. The first is the additive genetic variance-covariance matrix (G), 150 summarizing additive genetic variance for a set of traits9,10. The diagonal elements of G are the 151 additive genetic variances, and the off-diagonal elements are additive genetic covariances (Fig. 152 3C; see also Section 1 in Supplementary Material). Additive genetic variances and covariances 153 describe patterns of trait inheritance, and depend on the frequency and effects of alleles. The 154 additive genetic covariances are critical from a multivariate standpoint, because they describe the 155 extent to which inheritance of different traits tends to be non-independent. In the bivariate case, G 156 can be represented as an ellipse containing 95% of the genetic values of the individuals in a 157 population32 (Fig. 3C). If two traits are strongly genetically correlated, the ellipse will be eccentric 158 and oriented such that it is not parallel to either trait axis. That is, genetic covariances between 159 traits result in directions of multivariate trait space with high (major axis of the correlation) and 160 low (minor axis of the correlation) genetic variance, even if genetic variance is high in all 161 individual traits33 (Fig 3). Importantly, the long-axis of the G-matrix (gmax) represents a genetic 162 line of least resistance34, the direction in phenotypic space which harbors the most genetic variance 163 and along which the population most easily evolves (see “Consequences for pleiotropy, 164 evolvability, modularity and phenotypic plasticity”). 165 166 Multivariate phenotypic effects of new mutations constitute a second set of key parameters, which 167 are summarized in the mutational variance-covariance matrix (M)11,12. Theory often assumes that 168 mutational effects are normally distributed. In the univariate case, when a locus affects only one 169 trait, this distribution can be described by a mean and a variance, and if mutations are unbiased, 170 the mean will be zero. In the multivariate case, some loci might be pleiotropic13, meaning that they 171 affect more than one trait. In this case, the mutational effects are modeled as draws from a 172 multivariate normal distribution. This distribution is described by mutational variances for each 7 173 trait (diagonal elements of M) and mutational covariances between traits (off-diagonal elements 174 of M). Positive mutational covariances mean that a mutation tends to affect both traits in the same 175 direction, whereas negative mutational covariances indicate that mutations tend to affect traits in 176 opposite directions. 177 178 Our analytical understanding of how correlational selection shapes the evolution of genetic 179 variances and covariances comes from evolutionary quantitative genetic theory, particularly from 180 the pioneering work by Russell Lande13,35,36, and Wagner and Altenberg’s2 ideas about how 181 selection on pleiotropic patterns could lead to parcelation or integration between traits. Lande’s 182 work suggested that inheritance should become aligned with the shape of the selection surface in 183 well-adapted populations. Later, Cheverud used Lande’s model of selection on pleiotropic 184 mutations to predict that genetic correlations should match functional interactions among traits4. 185 Recently, this suggestion was extended to predict a three-way alignment among selection, 186 inheritance and mutation12. 187 188 How short term responses to correlational selection are transmitted across generations depends on 189 the distribution of allelic effects and the persistence of selection. Correlational selection can create 190 genetic correlation by promoting linkage disequilibrium between alleles that affect two different 191 traits6. However, such changes are expected to be eroded rapidly due to recombination if selection 192 is relaxed in subsequent generations6, suggesting that changes in genetic architecture due to this 193 kind of correlational selection may be transient37, unless correlational selection is persistent6. More 194 realistically, if correlational selection acts on traits whose expression is affected by alleles with 195 pleiotropic effects, then correlational selection will alter the frequencies of those pleiotropic 196 alleles. Therefore, the distribution of mutational effects has important consequences for the 197 efficacy of selection on genetic covariances. 198 8 199 Two recent advances have increased our general understanding of the evolution of genetic 200 architecture. First, increasingly powerful computer simulations have enabled researchers to 201 explore the long-term effects of correlational selection and mutation on the evolution of genetic 202 covariances9–12, expanding our knowledge beyond the case of mutation-selection balance under 203 the classical infinitesimal model35,38. Second, a rapid increase in genomic data has provided 204 insights into the empirical distributions of allelic effects in real populations. Combining both 205 approaches provides exciting opportunities to understand how selection and genetics jointly shape 206 the evolution of trait variation (see next section). 207 208 Simulation-based studies have verified the prediction by Lande13 and Cheverud4 that selection 209 will cause standing genetic variation to become aligned with the fitness surface9. For instance, if 210 the fitness surface is ridge-shaped, then populations will tend to harbor more variation in the 211 direction of phenotypic space aligned along the crest of the ridge and less variation perpendicular 212 to the ridge. However, other factors also influence the genetic architecture of traits: genetic drift 213 can cause G to fluctuate over evolutionary time9, a moving optimum stretches G in the direction 214 of the movement10, Migration also increases the genetic variance in the direction of phenotypic 215 space pointing toward the mean of the migrant source population in an island-mainland model39. 216 Recently, it has also been emphasized that the mutational variance is aligned with the direction of 217 phenotypic plasticity3,40, affecting both G and M. One interpretation of such alignment between 218 plasticity and mutational variance is that developmental systems might respond similarly to 219 environmental novelty as they do to genetic mutation40. Moreover, all else being equal, 220 correlations in M should generate correlations in G, because standing genetic variation ultimately 221 arises via mutation. 222 223 Interestingly, influences between the fitness surface, G and M can flow in both directions. While 224 M can influence the shape of G, the fitness surface in turn can shape both G and M11,12. Thus, if 9 225 the fitness surface is a ridge in phenotypic space (Fig. 3), selection will cause the long axis of G 226 to align with the ridge. If such a selective regime is stable over evolutionary time, selection can 227 cause alignment between the fitness surface, M and G12,41,42. Simulations show that evolution of 228 the mutational distribution is especially plausible when different loci interact epistatically12. 229 Recent progress in molecular biology, development and genomics suggests that such epistatic 230 interactions are extremely common43. Epistasis can therefore permit the evolution of the 231 mutational architecture because selection maintains variation at loci that have favorable 232 interactions under the prevailing selection regime. 233 234 A growing number of studies suggest that G can or has evolved in response to correlational 235 selection (Fig. 2). For instance, Delph et al.44 imposed artificial correlational selection on 236 combinations of male and female floral traits in the dioecious flower Silene latifolia (Fig. 2I) to 237 test whether the between-sex genetic correlation was evolvable. High between-sex genetic 238 correlations would potentially constrain the evolution of sexual dimorphism. Between-sex genetic 239 correlations broke down after a few generations of selection44, however, suggesting that these 240 correlations are due to linkage disequilibrium which is expected to break down rapidly under 241 artificial correlational selection or recombination. In another plant study, however, genetic 242 correlations were remarkably stable across several generations, suggesting that pleiotropy caused 243 these correlations45. 244 245 Genomic architecture of traits and consequences for multi-character evolution 246 247 The development of next-generation sequencing (NGS) provides new opportunities to investigate 248 correlational selection beyond what has been possible with classical quantitative genetics. 249 Genomic data has allowed us to pinpoint the genetic basis and architecture of traits, to estimate 10 250 empirical distributions of allelic effects in real populations, to reconstruct the evolution of genome 251 architecture relevant for trait evolution and to detect correlational selection from molecular 252 footprints (Box 2). 253 254 Recent studies using quantitative trait loci (QTL) mapping, genome-wide association studies 255 (GWAS), and whole genome sequencing of population samples (Box 2), have revealed that most 256 genotype-phenotype maps46 are complex. Most traits are determined by a large number of genes 257 of small effect, consistent with the so-called ‘polygenic model’ of inheritance38 allowing efficient 258 quantitative genetics modelling ignoring details of multilocus inheritance by assuming the 259 infinitesimal model47. However, empirical effect sizes distributions are often exponentially 260 distributed48,49 , with a few genes of major effect controlling a minority of traits for which the 261 infinitesimal model is violated50 and which often have an important role in adaptation and 262 speciation51,52. Molecular studies have further revealed that many functional genetic variants are 263 pleiotropic and affect multiple traits53. Multiple-mapping approaches, enabling joint estimation of 264 effects on multiple traits, hold great promise to further improve our understanding of pleiotropy54. 265 Molecular studies have also revealed that epistasis is common55, with genotype-phenotype maps 266 typically being highly nonlinear56, suggesting pervasive epistasis in genotype networks43. The 267 importance of epistasis is controversial because linear quantitative genetic models are rarely 268 improved by the addition of interaction terms57. However, genomic quantitative genetic studies 269 that incorporate a more precise estimate of the shared proportion of genome have revealed that 270 higher order variance components are not negligible58. Interestingly, a recent study of Timema 271 stick insects has shown that correlational selection arose from fitness epistasis due to ecological 272 factors (predation), in spite of the underlying traits (colour) having an additive genetic basis59. 11 273 274 These insights about the genetic architecture of traits have implications for the evolutionary 275 response to correlational selection. One emerging insight from experimental evolution studies is 276 that evolutionary changes in traits can often be achieved via many alternative "genomic solutions", 277 suggesting important roles for redundancy and historical contingency in evolution60. Further, 278 parallel evolution is often frequent for fitness itself, but less common for phenotypes, and less 279 common still at the levels of genes or individual mutations61. For example, Therkildsen et al.62 280 sequenced genomes of Atlantic silverside fish (Menidia menidia) selected for small and large size. 281 Despite highly parallel phenotypic changes and several parallel allele frequency shifts in growth- 282 related genes, genomic adaptation in one line was contingent on the presence of a large inversion 283 with moderate phenotypic effect62. On the other hand, pleiotropy, functional constraints and the 284 presence of major effect loci may limit the number of redundant genomic solutions in response to 285 correlational selection63. For example, threespine stickleback adapting to freshwater habitats show 286 highly repeatable evolution at a pleiotropic major effect locus64,65. 287 288 Correlational selection changes the genetic covariances among traits and thereby ultimately shapes 289 the evolution of genome architecture. Although many different mechanisms underly genetic 290 covariances66, the two basic causes are linkage disequilibrium and pleiotropy38. Linkage 291 disequilibrium captured by physical linkage is one genomic cause of trait correlations, in which 292 recombination is suppressed in heterochromatic regions, in genomic rearrangements, on sex 293 chromosomes, or due to a high density of transposable elements67. For example, Choudhury et al.68 294 sequenced 304 Arabis alpina genomes, and found that the S-locus supergene responsible for strict 295 outcrossing was in a linkage disequilibrium block that included high levels of polymorphic 296 transposable elements. Genomic rearrangements such as gene duplications, translocations, 12 297 chromosomal fusions or inversions can also maintain linkage disequilibrium, due to disrupted 298 meiotic chromosome pairing reducing recombination or the joint forces of physical linkage and 299 selection69–71. Linkage disequilibrium may be preserved in deep evolutionary time, forming so- 300 called ‘supergenes', some of which might resemble sex chromosomes69,72–74 (Fig. 4A). There are 301 several empirical examples of how co-selected complex trait combinations, bound to supergenes, 302 cause behavioural and morphological differences between discrete morphs with different 303 reproductive tactics (Fig. 4B)72–74. For example, a recent study in the heterostylic plant genus 304 Primula revealed the build-up of an S-locus supergene controlling style, anther and pollen grains 305 via gene duplications and neofunctionalization75. 306 307 Even in the absence of physical recombination suppression, genetic covariances among traits can 308 arise through linkage disequilibrium between loci6. Such linkage disequilibrium can potentially be 309 maintained by assortative mating among individuals with the same correlated trait combinations76 310 and by strong divergent or disruptive selection favoring several correlated trait optima within6,77 311 or between populations78. Correlational selection can also theoretically lead to speciation through 312 reinforcement of assortative mating by the evolution of genomic coupling between preference and 313 trait loci, even if they are initially unlinked6,76,79. There are several examples of ecotype or species 314 pairs where inter-chromosomal linkage disequilibrium is maintained either by strong selection80 315 or a combination of strong selection and assortative mating81, with some studies demonstrating 316 genomic coupling between unlinked preference and sexually-selected trait loci76. 317 13 318 Genomic tools in combination with quantitative genetic approaches also enable us to obtain better 319 estimates of G (Box 2)58,82. In addition, comparisons between the genetic variances and 320 covariances of M (the mutation matrix) and G (the matrix of standing genetic variation) can reveal 321 the presence of correlational selection and how it operates during the organismal life-cycle and 322 shapes both mutational pleiotropy and pleiotropy of the standing genetic variation83–85 (Box 2). 323 Mutation accumulation experiments (MA) have revealed strong mutational pleiotropy85–87 and 324 have indicated that correlational selection on such pleiotropy leads to a reduction in the 325 corresponding genetic correlations in G83(Box 2). Thus, correlational selection might operate 326 against mutational pleiotropy, resulting in a discrepancy between M and G. Consequently, 327 spontaneous and positive mutational correlations among traits could largely be maladaptive and 328 reflect the input from mutation-selection balance86. This contrived example underscores the point 329 that correlational selection can not only strengthen adaptive genetic correlations among traits, it 330 can also weaken and break up maladaptive genetic correlations44,83 331 332 Finally, genomic information from several populations can be used to address how multiple traits 333 co-evolve, using information from a coancestry matrix, which can be estimated with a handful of 334 marker loci88. Csilléry et al.89 recently used such an approach and found evidence for correlated 335 character evolution in the timing and growth rate across 16 silver fir (Abies alba) populations. 336 While this methodology is limited to describing the average effect across all causal loci, such 337 approaches could enable us to describing the genomic architecture of trait correlations in terms of 338 individual loci, their physical distribution across the genome, and their effect sizes90. 339 340 Consequences for pleiotropy, evolvability, modularity and phenotypic plasticity 341 342 The intuition that pleiotropy slows and constrains evolution was well articulated by Orr91, who 343 updated Fisher’s classical geometric model92 to show that phenotypic complexity slows adaptation 14 344 when pleiotropy is universal. However, this link between pleiotropy and constraints on 345 evolvability - the ability of a population to respond to selection93,94 - has recently been challenged. 346 First, pleiotropy may be largely confined within functional, integrated trait modules46, allowing 347 traits in separate modules to adapt semi-independently2. That is, as predicted by Cheverud, 348 correlational selection will select for congruent phenotypic covariances4,13. Moreover, individual- 349 based simulations with divergent multivariate directional selection, pushing groups of traits in 350 opposite directions, showed that phenotypic variation can indeed evolve to become more 351 modular95. Increased modularity may also evolve when environmental fluctuations favour new 352 combinations of conserved functions96 or when selection across multiple environments favors the 353 expression of partially overlapping sets of genes97 . While these studies suggest that circumstances 354 favoring high evolvability can drive the evolution of modularity, theory has also shown that highly 355 integrated and pleiotropic genetic architectures can have high evolvability93. There is still 356 considerable room for development of theory to predict when we expect modularity to emerge as 357 a solution to adaptive challenges (Section 2 in Supplementary Material). 358 359 A common feature for the evolutionary origin of modularity is directional selection, though 360 modularity can also evolve as a consequence of selection for robustness to environmental 361 perturbations98, and merely adding selection to models with universal pleiotropy does not produce 362 stable variational modules99. The responsiveness of modularity to directional selection also limits 363 its use as a predictor of long-term evolutionary responses, perhaps explaining why functional 364 modularity is only a modest predictor of co-evolutionary rates of evolution among genes100. 365 Another potential explanation for this pattern is that functional and variational modularity only 366 partially overlap. Empirical evidence of co-expression of genes is strong101, but whether these co- 367 adapted gene modules are organized as variational modules is controversial. Some studies using 368 transcriptional data showed that genetically correlated transcripts tend to share developmental 369 pathways, reflecting transcriptional modules that are mostly enriched with functionally related 15 370 genes102,103, whereas other studies could not find significant overlap between gene expression and 371 functional groupings104. Modular functional capacities do not require structural modularity105, and 372 modularity at the level of gene regulation may better predict evolvability106. The mismatch 373 between variational modules and functional gene groupings can complicate the semi-independent 374 evolution of phenotypic modularity. 375 376 Multivariate perspectives show that additive genetic variation within populations is distributed 377 very unevenly across traits, with some linear combinations of traits accounting for most of the 378 variance (i.e., gmax), while other trait combinations are associated with very little variance33. This 379 pattern can stem from genetic variation being funneled through a few central developmental 380 pathways, mediated by few developmental genes of large effect107. There has been considerable 381 interest in gmax, (see section “Evolution of genetic architecture in response to correlational 382 selection”) because it can either facilitate or bias evolutionary responses to selection depending 383 on its alignment to the selective surface108. Additive genetic variance is determined by the effects 384 and frequencies of contributing alleles, and at the genomic level, the initial response to selection 385 should be dominated by loci with relatively high intra-locus variance and large effect. Although 386 genomic studies, such as GWAS, have a tendency to detect loci with high frequencies of minor 387 alleles and larger effects109, empirical evidence of the contribution of variants to additive genetic 388 variance points to mostly rare variants with mainly small but highly pleiotropic effects110. If gmax 389 reflects the most common empirical pattern, selection aligned with gmax should promote adaptation 390 through minor allele frequency changes at many loci111. 391 392 The mere presence of additive genetic variation is not sufficient to predict evolutionary outcomes, 393 as the response to selection depends on the orientation of selection relative to the distribution of 394 genetic variation94,112. Only a few studies have measured both multivariate linear and quadratic 395 selection and the distribution of genetic variation for those phenotypes. These studies typically 16 396 demonstrate relatively low genetic variance in the multivariate trait combinations associated with 397 fitness variation, which slows down phenotypic evolution113. However, the causes of low genetic 398 variance for multivariate phenotypes currently under selection, and thus the longer-term 399 consequences of the covariance patterns, remain poorly resolved33. 400 401 Because there is substantial genetic variance in other directions of trait space besides gmax, changes 402 in the orientation of selection could result in relatively unbiased, rapid, adaptation94. Moreover, 403 pleiotropy can be context-dependent101,114, meaning that apparent pleiotropic constraints may shift 404 in novel environments or evolve through epistatic interactions. For instance, changes in the 405 selective environment could remove bias by changing the orientation of genetic variation40, 406 potentially through context-dependent pleiotropic effects of alleles114, or through rapid evolution 407 of G, which might be particularly likely if trait covariances are generated through opposing 408 pleiotropic effects across contributing loci93. For example, a recent experimental study in yeast 409 demonstrated that while alleles had pleiotropic effects on two life-history traits, variation in effects 410 across environments resulted in genetic correlations ranging from -0.5 to +0.5115. Finally, 411 mutational pleiotropy is only one of several factors influencing standing genetic variation, which 412 also depends on multivariate selection and linkage disequilibrium. 413 414 The relationship between how genetic variance changes across contexts and how phenotypes 415 respond to the environment directly (i.e., phenotypic plasticity) can determine the longer-term 416 outcomes of correlational selection. A recent meta-analysis of published estimates of G and plastic 417 responses to novel environments suggested that multivariate phenotypic plasticity might 418 correspond to axes of genetic variation associated with substantial standing genetic variation40. 419 This study and theory3 suggest that bias in evolutionary response generated through G can become 420 recapitulated through phenotypic plasticity. Clearly, more work is needed to understand how 17 421 environmental and genetic information are interpreted through the developmental systems (Section 422 2 in Supplementary Material). 423 424 Conclusions 425 426 Here, we have re-visited the early suggestions by Cheverud and Lande that correlational selection 427 can shape genetic and phenotypic architecture in light of the recent genomic revolution. These 428 early insights are consistent with increasing empirical evidence of genomic coupling and 429 recombination suppression that could have arisen by correlational selection, although direct 430 evidence for this process, in most cases, lacking. A remaining challenge is therefore to integrate 431 organismal-level research on correlational selection on phenotypes with genomics and 432 developmental biology. Below, we point to some promising new avenues for future integrative 433 research in this exciting area. 434 435 First, despite empirical evidence that correlational selection can build up or eliminate genetic 436 correlations between co-selected traits (Fig. 2), our knowledge of the mechanistic (i.e. genomic 437 and developmental) underpinnings of such changes is still limited. To what extent are such changes 438 caused by transient changes in linkage disequilibrium or the evolution of adaptive pleiotropy, and 439 what is the relationship between modularity and correlational selection? We are only just 440 beginning to understand the genomic mechanisms involved in adaptive recombination suppression 441 caused by correlational selection, including the roles of supergenes69,71, structural genomic 442 rearrangements70 such as gene duplications, chromosomal fusions or inversions and other 443 mechanisms including TEs67,68. Promising future research directions in the study of the genomic 444 consequences of correlational selection are to use genomic tools to study how correlational 445 selection might lead to the gradual buildup of supergenes75 and how such selection might operate 446 on mutational pleiotropy across the organismal life-cycle, using a combination of mutation 18 447 accumulation (MA) experiments, quantitative genetics and quantification of gene expression 448 changes during ontogeny83,83,85–87 449 450 Second, the relationship between phenotypic plasticity and correlational selection is largely 451 unknown. The traditional perspective has been that correlational selection would primarily shape 452 genetic correlation structure, by either strengthening or weakening genetic correlations between 453 traits4,6,44. Research on stickleback fish has found that predation results in changed phenotypic 454 correlation structures15, but some of these phenotypic changes might reflect multivariate 455 phenotypic plasticity rather than changes in genetic integration16. How genetic covariances and 456 multivariate phenotypic plasticity jointly evolve under correlational selection is therefore a largely 457 unexplored research area with great potential16,40. More generally, since correlational selection in 458 the past might have shaped either phenotypic or genetic correlations (or both), it leaves an 459 evolutionary “memory” of past selective environments116 which can reveal itself in the form of 460 alignment between the selective surface, P, G and M12,41,42. 461 462 Third, the importance of correlational selection in speciation and macroevolution is largely 463 unknown, despite early work on evolutionary allometry and the idea of evolution along “lines of 464 least resistance”34,117. Recent research on shape-size allometry118, brain-body size allometry119 and 465 metabolic allometry120 have revealed that allometric relationships are not static evolutionary 466 constraints, but can be altered by selection. Specifically, correlational selection could maintain 467 adaptive allometric slopes, either due to internal causes related to deleterious pleiotropy118 or 468 because external ecological factors make certain slopes more beneficial than others119,120. 469 470 Finally, we also see a great potential for research on the genomic consequences of correlational 471 selection in the fields of animal and plant domestication17, and in the context of dispersal strategies, 472 social behaviours and personalities16,25,121. Humans might have consciously or unconsciously 19 473 either eliminated or strengthened genetic correlations between traits during domestication of plants 474 and animals, through artificial correlational selection on suites of traits, which in some cases has 475 led to adaptive introgression back into wild relatives122. One result of domestication is the 476 formation of suites of co-inherited traits with distinct genomic signatures17. In natural populations, 477 co-adaptation between social behaviours and dispersal121 could frequently have been driven by 478 correlational selection, resulting in increased genetic integration25. Artificial correlational 479 selection to either strengthen118 or eliminate genetic correlations44 is a promising experimental 480 approach in this context. 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 20 499 References 500 501 1. 921–931 (2007). 502 503 2. Wagner, G. P. & Altenberg, L. Perspective: Complex adaptations and the evolution of evolvability. Evolution 50, 967–976 (1996). 504 505 Wagner, G. P., Pavlicev, M. & Cheverud, J. M. The road to modularity. Nat. Rev. Genet. 8, 3. Draghi, J. A. & Whitlock, M. C. Phenotypic plasticity facilitates mutational variance, 506 genetic variance, and evolvability along the major axis of environmental variation. 507 Evolution 66, 2891–2902 (2012). 508 4. selection. J. Theor. Biol. 110, 155–171 (1984). 509 510 5. Phillips, P. C. & Arnold, S. J. Visualizing multivariate selection. Evolution 43, 1209–1266 (1989). 511 512 Cheverud, J. M. Quantitative genetics and developmental constraints on evolution by 6. Sinervo, B. & Svensson, E. Correlational selection and the evolution of genomic architecture. Heredity 16, 948–955 (2002). 513 514 7. Blows, M. W. & Brooks, R. Measuring nonlinear selection. Am. Nat. 162, 815–820 (2003). 515 8. Blows, M. W., Brooks, R. & Kraft, P. G. Exploring complex fitness surfaces: multiple ornamentation and polymorphism in male guppies. Evolution 57, 1622–1630 (2003). 516 517 9. Jones, A. G., Arnold, S. J. & Bürger, R. Stability of the G-matrix in a population 518 experiencing pleiotropic mutation, stabilizing selection, and genetic drift. Evolution 57, 519 1747–1760 (2003). 520 521 10. Jones, A. G., Arnold, S. J. & Bürger, R. Evolution and stability of the G-matrix on a landscape with a moving optimum. Evolution 58, 1639–1654 (2004). 21 522 523 524 525 526 527 528 11. Jones, A. G., Arnold, S. J. & Bürger, R. The mutation matrix and the evolution of evolvability. Evolution 61, 727–745 (2007). 12. Jones, A. G., Bürger, R. & Arnold, S. J. Epistasis and natural selection shape the mutational architecture of complex traits. Nat. Commun. 5, 3709 (2014). 13. Lande, R. The genetic covariance between characters maintained by pleiotropic mutations. Genetics 94, 203–215 (1980). 14. Armbruster, W. S., Pélabon, C., Hansen, T. F. & Mulder, C. P. H. Floral integration, 529 modularity and accuracy: distinguishing complex adaptations from genetic constraints. in 530 Phenotypic integration: studying the ecology and evolution of complex phenotypes (eds. 531 Pigliucci, M. & Preston, K.) 23–49 (Oxford University Press, 2004). 532 533 15. Bell, A. M. & Sih, A. Exposure to predation generates personality in threespined sticklebacks (Gasterosteus aculeatus). Ecol. Lett. 10, 828–834 (2007). 534 16. Dingemanse, N. J., Barber, I. & Dochtermann, N. A. Non-consumptive effects of predation: 535 does perceived risk strengthen the genetic integration of behaviour and morphology in 536 stickleback? Ecol. Lett. 23, 107-118. 537 17. Hansen Wheat, C., Fitzpatrick, J. L., Rogell, B. & Temrin, H. Behavioural correlations of 538 the domestication syndrome are decoupled in modern dog breeds. Nat. Commun. 10, 1–9 539 (2019). 540 541 542 543 544 18. Hurst, L. D., Pál, C. & Lercher, M. J. The evolutionary dynamics of eukaryotic gene order. Nat. Rev. Genet. 5, 299 (2004). 19. Lande, R. & Arnold, S. J. The measurement of selection on correlated characters. Evolution 37, 1210–1226 (1983). 20. Schluter, D. & Nychka, D. Exploring fitness surfaces. Am. Nat. 143, 597–616 (1994). 22 545 546 547 548 549 21. Siepielski, A. M. et al. Precipitation drives global variation in natural selection. Science 355, 959–962 (2017). 22. Roff, D. A. & Fairbairn, D. J. A test of the hypothesis that correlational selection generates genetic correlations. Evolution 66, 2953–2960 (2012). 23. Svensson, E. I., McAdam, A. G. & Sinervo, B. Intralocus sexual conflict over immune 550 defense, gender load, and sex-specific signaling in a natural lizard population. Evolution 63, 551 3124–3135 (2009). 552 24. McGlothlin, J. W., Parker, P. G., Nolan, V. & Ketterson, E. D. Correlational selection leads 553 to genetic integration of body size and an attractive plumage trait in dark-eyed juncos. 554 Evolution 59, 658–671 (2005). 555 556 557 558 559 560 561 25. Duckworth, R. A. & Kruuk, L. E. B. Evolution of genetic integration between dispersal and colonization ability in a bird. Evolution 63, 968–977 (2009). 26. Brodie III, E. D. Correlational selection for color pattern and antipredator behavior in the garter snake Thamnophis ordinoides. Evolution 46, 1284–1298 (1992). 27. Wise, M. J. & Rausher, M. D. Costs of resistance and correlational selection in the multiple-herbivore community of Solanum carolinense. Evolution 70, 2411–2420 (2016). 28. Fenster, C. B., Reynolds, R. J., Williams, C. W., Makowsky, R. & Dudash, M. R. 562 Quantifying hummingbird preference for floral trait combinations: The role of selection on 563 trait interactions in the evolution of pollination syndromes. Evolution 69, 1113–1127 564 (2015). 565 566 29. Arnegard, M. E. et al. Genetics of ecological divergence during speciation. Nature 511, 307–311 (2014). 23 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 30. Martin, CH. & Wainwright, P. C. Multiple fitness peaks on the adaptive landscape drive adaptive radiation in the wild. Science 339, 208–211 (2013). 31. Phillips, P. C. Epistasis - the essential role of gene interactions in the structure and evolution of genetic systems. Nat.Rev.Gen. 9, 855–867 (2008). 32. Steppan, S. J., Phillips, P. C. & Houle, D. Comparative quantitative genetics: evolution of the G matrix. Trends Ecol. Evol. 17, 320–327 (2002). 33. Blows, M. W. & McGuigan, K. The distribution of genetic variance across phenotypic space and the response to selection. Mol. Ecol. 24, 2056–2072 (2015). 34. Schluter, D. Adaptive radiation along genetic lines of least resistance. Evolution 50, 1766– 1774 (1996). 35. Lande, R. The maintenance of genetic variability by mutation in a polygenic character with linked loci. Genet. Res. 26, 221–235 (1976). 36. Lande, R. The genetic correlation between characters maintained by selection, linkage and inbreeding. Genet. Res. 44, 309–320 (1984). 37. Bulmer, M. G. The effect of selection on genetic variability: a simulation study. Genet. Res. 28, 101–117 (1976). 38. Lynch, M. & Walsh, B. Genetics and analysis of quantitative traits. (Sinauer Associates, Inc., 1998). 39. Guillaume, F. & Whitlock, M. C. Effects of migration on the genetic covariance matrix. Evolution 61, 2398–2409 (2007). 40. Noble, D. W. A., Radersma, R. & Uller, T. Plastic responses to novel environments are 588 biased towards phenotype dimensions with high additive genetic variation. Proc. Natl. 589 Acad. Sci. 116, 13452–13461 (2019). 24 590 591 592 593 594 41. Houle, D., Bolstad, G. H., van der Linde, K. & Hansen, T. F. Mutation predicts 40 million years of fly wing evolution. Nature 548, 447–450 (2017). 42. Svensson, E. I. & Berger, D. The role of mutation bias in adaptive evolution. Trends Ecol. Evol. 34, 422–434 (2019). 43. Schweizer, G. & Wagner, A. Genotype networks of 80 quantitative Arabidopsis thaliana 595 phenotypes reveal phenotypic evolvability despite pervasive epistasis. bioRxiv 596 2020.02.26.966390 (2020) doi:10.1101/2020.02.26.966390. 597 44. Delph, L. F., Steven, J. C., Anderson, I. A., Herlihy, C. R. & Brodie, E. D., III. Elimination 598 of a genetic correlation between the sexes via artificial correlational selection. Evolution 65, 599 2872–2880 (2011). 600 601 602 603 604 605 606 607 608 609 610 45. Conner, J. K. Genetic mechanisms of floral trait correlations in a natural population. Nature 420, 407–410 (2002). 46. Wagner, G. P. & Zhang, J. The pleiotropic structure of the genotype–phenotype map: the evolvability of complex organisms. Nat. Rev. Genet. 12, 204–213 (2011). 47. Barton, N. H., Etheridge, A. M. & Véber, A. The infinitesimal model: Definition, derivation, and implications. Theor. Popul. Biol. 118, 50–73 (2017). 48. Orr, H. A. The population genetics of adaptation: the distribution of factors fixed during adaptive evolution. Evolution 52, 935–948 (1998). 49. Flint, J. & Mackay, T. F. C. Genetic architecture of quantitative traits in mice, flies, and humans. Genome Res. 19, 723–733 (2009). 50. Stinchcombe, J. R., Weinig, C., Heath, K. D., Brock, M. T. & Schmitt, J. Polymorphic 611 genes of major effect: Consequences for variation, selection and evolution in Arabidopsis 612 thaliana. Genetics 182, 911–922 (2009). 25 613 51. Orr, H. A. The genetics of species differences. Trends Ecol. Evol. 16, 343–350 (2001). 614 52. Nadeau, N. J. et al. The gene cortex controls mimicry and crypsis in butterflies and moths. 615 616 617 618 619 620 621 622 623 624 Nature 534, 106–110 (2016). 53. Visscher, P. M. et al. 10 years of GWAS discovery: Biology, function, and translation. Am. J. Hum. Genet. 101, 5–22 (2017). 54. Pitchers, W. et al. A multivariate genome-wide association study of wing shape in Drosophila melanogaster. Genetics 211, 1429–1447 (2019). 55. Mackay, T. F. C. Epistasis and quantitative traits: using model organisms to study gene– gene interactions. Nat. Rev. Genet. 15, 22–33 (2014). 56. Sailer, Z. R. & Harms, M. J. Detecting high-order epistasis in nonlinear genotypephenotype maps. Genetics 205, 1079–1088 (2017). 57. Hill, W. G. “Conversion” of epistatic into additive genetic variance in finite populations 625 and possible impact on long-term selection response. J. Anim. Breed. Genet. 134, 196–201 626 (2017). 627 628 629 630 631 632 633 634 58. Gienapp, P. et al. Genomic quantitative genetics to study evolution in the wild. Trends Ecol. Evol. 32, 897–908 (2017). 59. Nosil, P. et al. Ecology shapes epistasis in a genotype–phenotype–fitness map for stick insect colour. Nat. Ecol. Evol. 1–12 (2020) doi:10.1038/s41559-020-01305-y. 60. Blount, Z. D., Lenski, R. E. & Losos, J. B. Contingency and determinism in evolution: Replaying life’s tape. Science 362, (2018). 61. Bolnick, D. I., Barrett, R. D. H., Oke, K. B., Rennison, D. J. & Stuart, Y. E. (Non)Parallel Evolution. Annu. Rev. Ecol. Evol. Syst. 49, 303–330 (2018). 26 635 636 637 638 639 640 641 62. Therkildsen, N. O. et al. Contrasting genomic shifts underlie parallel phenotypic evolution in response to fishing. Science 365, 487–490 (2019). 63. Stern, D. L. & Orgogozo, V. Is genetic evolution predictable? Science 323, 746–751 (2009). 64. Colosimo, P. F. et al. Widespread parallel evolution in sticklebacks by repeated fixation of Ectodysplasin alleles. Science 307, 1928–1933 (2005). 65. Archambeault, S. L., Bärtschi, L. R., Merminod, A. D. & Peichel, C. L. Adaptation via 642 pleiotropy and linkage: Association mapping reveals a complex genetic architecture within 643 the stickleback Eda locus. Evol. Lett. 4, 282–301 (2020). 644 66. van Rheenen, W., Peyrot, W. J., Schork, A. J., Lee, S. H. & Wray, N. R. Genetic 645 correlations of polygenic disease traits: from theory to practice. Nat. Rev. Genet. 20, 567– 646 581 (2019). 647 67. Stapley, J., Feulner, P. G. D., Johnston, S. E., Santure, A. W. & Smadja, C. M. Variation in 648 recombination frequency and distribution across eukaryotes: patterns and processes. Philos. 649 Trans. R. Soc. B Biol. Sci. 372, 20160455 (2017). 650 68. Choudhury, R. R., Rogivue, A., Gugerli, F. & Parisod, C. Impact of polymorphic 651 transposable elements on linkage disequilibrium along chromosomes. Mol. Ecol. 28, 1550– 652 1562 (2019). 653 654 655 656 69. Thompson, M. J. & Jiggins, C. D. Supergenes and their role in evolution. Heredity 113, 1–8 (2014). 70. Yeaman, S. Genomic rearrangements and the evolution of clusters of locally adaptive loci. Proc. Natl. Acad. Sci. U. S. A. 110, E1743-1751 (2013). 27 657 658 659 660 661 662 663 664 665 71. Faria, R., Johannesson, K., Butlin, R. K. & Westram, A. M. Evolving inversions. Trends Ecol. Evol. 34, 239–248 (2019). 72. Tuttle, E. M. et al. Divergence and functional degradation of a sex chromosome-like supergene. Curr. Biol. 26, 344–350 (2016). 73. Kupper, C. et al. A supergene determines highly divergent male reproductive morphs in the ruff. Nat. Genet. 48, 79–83 (2016). 74. Lamichhaney, S. et al. Structural genomic changes underlie alternative reproductive strategies in the ruff (Philomachus pugnax). Nat.Genet. 48, 84–88 (2016). 75. Huu, C. N., Keller, B., Conti, E., Kappel, C. & Lenhard, M. Supergene evolution via 666 stepwise duplications and neofunctionalization of a floral-organ identity gene. Proc. Natl. 667 Acad. Sci. 117, 23148–23157 (2020). 668 669 670 671 672 673 674 675 676 76. Merrill, R. M. et al. Genetic dissection of assortative mating behavior. PLOS Biol. 17, e2005902 (2019). 77. Whitlock, M. C., Phillips, P. C., Moore, F. B.-G. & Tonsor, S. J. Multiple fitness peaks and epistasis. Annu. Rev. Ecol. Syst. 26, 601–629 (1995). 78. Dudley, S. A. The response to selection on plant physiological traits: evidence for local adaptation. Evolution 50, 103–110 (1996). 79. Kirkpatrick, M. & Ravigné, V. Speciation by natural and sexual selection: Models and experiments. Am. Nat. 159, S22–S35 (2002). 80. Hohenlohe Paul A., Bassham Susan, Currey Mark & Cresko William A. Extensive linkage 677 disequilibrium and parallel adaptive divergence across threespine stickleback genomes. 678 Philos. Trans. R. Soc. B Biol. Sci. 367, 395–408 (2012). 28 679 81. Hench, K., Vargas, M., Höppner, M. P., McMillan, W. O. & Puebla, O. Inter-chromosomal 680 coupling between vision and pigmentation genes during genomic divergence. Nat. Ecol. 681 Evol. 3, 657 (2019). 682 683 82. Gienapp, P., Calus, M. P. L., Laine, V. N. & Visser, M. E. Genomic selection on breeding time in a wild bird population. Evol. Lett. 3, 142–151 (2019). 684 83. McGuigan, K., Collet, J. M., Allen, S. L., Chenoweth, S. F. & Blows, M. W. Pleiotropic 685 mutations are subject to strong stabilizing selection. Genetics 197, 1051-+ (2014). 686 687 84. McGuigan, K. et al. The nature and extent of mutational pleiotropy in gene expression of male Drosophila serrata. Genetics 196, 911-+ (2014). 688 85. Hine, E., Runcie, D. E., McGuigan, K. & Blows, M. W. Uneven distribution of mutational 689 variance across the transcriptome of Drosophila serrata revealed by high-dimensional 690 analysis of gene expression. Genetics (2018) doi:10.1534/genetics.118.300757. 691 86. Estes, S., Ajie, B. C., Lynch, M. & Phillips, P. C. Spontaneous mutational correlations for 692 life-history, morphological and behavioral characters in Caenorhabditis elegans. Genetics 693 170, 645–653 (2005). 694 87. Houle, D. & Fierst, J. Properties of spontaneous mutational variance and covariance for 695 wing size and shape in Drosophila melanogaster. Evolution 67, 1116–1130 (2013). 696 88. Ovaskainen, O., Karhunen, M., Zheng, C., Arias, J. M. C. & Merilä, J. A new method to 697 uncover signatures of divergent and stabilizing selection in quantitative traits. Genetics 189, 698 621–632 (2011). 699 89. Csilléry, K. et al. Adaptation to local climate in multi-trait space: evidence from silver fir ( 700 Abies alba Mill.) populations across a heterogeneous environment. Heredity 124, 77–92 701 (2020). 29 702 703 90. Berg, J. J. & Coop, G. A Population genetic signal of polygenic adaptation. PLOS Genet. 10, e1004412 (2014). 704 91. Orr, H. A. Adaptation and the cost of complexity. Evolution 54, 13–20 (2000). 705 92. Fisher, R. A. The Genetical Theory of Natural Selection. (Clarendon Press, 1930). 706 93. Pavlicev, M. & Hansen, T. F. Genotype-phenotype maps maximizing evolvability: 707 708 709 710 711 712 713 714 715 716 modularity revisited. Evol. Biol. 38, 371–389 (2011). 94. Hine, E., McGuigan, K. & Blows, M. W. Evolutionary constraints in high-dimensional trait sets. Am. Nat. 184, 119–131 (2014). 95. Melo, D. & Marroig, G. Directional selection can drive the evolution of modularity in complex traits. Proc. Natl. Acad. Sci. 112, 470–475 (2015). 96. Kashtan, N. & Alon, U. Spontaneous evolution of modularity and network motifs. Proc. Natl. Acad. Sci. 102, 13773–13778 (2005). 97. Espinosa-Soto, C. & Wagner, A. Specialization can drive the evolution of modularity. PLoS Comput. Biol. 6, e1000719 (2010). 98. Ancel, L. W. & Fontana, W. Evolutionary lock-in and the origin of modularity in RNA 717 structure. in Modularity: Understanding the Development and Evolution of Natural 718 Complex Systems (eds. Callebaut, W. & Rasskin-Gutman, D.) 129–141 (The MIT Press, 719 2009). 720 99. Wagner, G. P. & Mezey, J. G. The role of genetic architecture constraints in the origin of 721 variational modularity. in Modularity in Development and Evolution (eds. Schlosser, G. & 722 Wagner, G. P.) 338–358 (University of Chicago Press, 2004). 723 724 100. Fokkens, L. & Snel, B. Cohesive versus flexible evolution of functional modules in Eukaryotes. PLOS Comput. Biol. 5, e1000276 (2009). 30 725 726 727 101. Huang, W. et al. Genetic basis of transcriptome diversity in Drosophila melanogaster. Proc. Natl. Acad. Sci. 112, E6010–E6019 (2015). 102. Schweizer, R. M. et al. Physiological and genomic evidence that selection on the 728 transcription factor Epas1 has altered cardiovascular function in high-altitude deer mice. 729 PLOS Genet. 15, e1008420 (2019). 730 731 732 103. Hämälä, T. et al. Gene expression modularity reveals footprints of polygenic adaptation in Theobroma cacao. Mol. Biol. Evol. 37, 110–123 (2020). 104. Collet, J. M., McGuigan, K., Allen, S. L., Chenoweth, S. F. & Blows, M. W. Mutational 733 Pleiotropy and the strength of stabilizing selection within and between functional modules 734 of gene expression. Genetics 208, 1601–1616 (2018). 735 736 737 738 739 740 741 742 743 744 745 105. Jiménez, A., Cotterell, J., Munteanu, A. & Sharpe, J. A spectrum of modularity in multi‐ functional gene circuits. Mol. Syst. Biol. 13, (2017). 106. Verd, B., Monk, N. A. & Jaeger, J. Modularity, criticality, and evolvability of a developmental gene regulatory network. eLife 8, e42832 (2019). 107. Pallares, L. F. et al. Mapping of craniofacial traits in outbred mice identifies major developmental genes involved in shape determination. PLOS Genet. 11, e1005607 (2015). 108. Arnold, S. J., Pfrender, M. E. & Jones, A. G. The adaptive landscape as a conceptual bridge between micro- and macroevolution. Genetica 112–113, 9–32 (2001). 109. Rockman, M. V. The QTN program and the alleles that matter for evolution: all that’s gold does not glitter. Evol. Int. J. Org. Evol. 66, 1–17 (2012). 110. Shikov, A. E., Skitchenko, R. K., Predeus, A. V. & Barbitoff, Y. A. Phenome-wide 746 functional dissection of pleiotropic effects highlights key molecular pathways for human 747 complex traits. Sci. Rep. 10, 1–10 (2020). 31 748 749 111. Sella, G. & Barton, N. H. Thinking About the evolution of complex traits in the era of genome-wide association studies. Annu. Rev. Genomics Hum. Genet. 20, 461–493 (2019). 750 112. Walsh, B. & Blows, M. W. Abundant genetic variation plus strong selection = Multivariate 751 genetic constraints: A geometric view of adaptation. Annu. Rev. Ecol. Evol. Syst. 40, 41–59 752 (2009). 753 754 755 113. Teplitsky, C. et al. Assessing multivariate constraints to evolution across ten long-term avian studies. PLOS ONE 9, e90444 (2014). 114. Pavlicev, M. & Cheverud, J. M. Constraints evolve: context dependency of gene effects 756 allows evolution of pleiotropy. Ann. Rev. Ecol. Evol. Syst. 46, 413–434 (2015). 757 115. Wei, X. & Zhang, J. Environment-dependent pleiotropic effects of mutations on the 758 maximum growth rate r and carrying capacity K of population growth. PLOS Biol. 17, 759 e3000121 (2019). 760 116. Parter, M., Kashtan, N. & Alon, U. Facilitated variation: How evolution learns from past 761 environments to generalize to new Environments. PLOS Comput. Biol. 4, e1000206 (2008). 762 117. Lande, R. Quantitative genetic analysis of multivariate evolution, applied to brain:body size 763 764 allometry. Evolution 33, 402–416 (1979). 118. Bolstad, G. H. et al. Complex constraints on allometry revealed by artificial selection on the 765 wing of Drosophila melanogaster. Proc. Natl. Acad. Sci., USA. 112, 13284–13289 (2015). 766 119. Tsuboi, M. et al. Breakdown of brain–body allometry and the encephalization of birds and 767 768 769 mammals. Nat. Ecol. Evol. 2, 1492–1500 (2018). 120. White, C. R. et al. The origin and maintenance of metabolic allometry in animals. Nat. Ecol. Evol. 3, 598 (2019). 32 770 771 772 773 774 121. Mullon, C., Keller, L. & Lehmann, L. Social polymorphism is favoured by the co-evolution of dispersal with social behaviour. Nat. Ecol. Evol. 2, 132–140 (2018). 122. Schweizer, R. M. et al. Natural selection and origin of a melanistic allele in North American gray wolves. Mol. Biol. Evol. 35, 1190–1209 (2018). 123. Brodie III, E. D. Genetic correlations between morphology and antipredator behaviour in 775 natural populations of the garter snake Thamnophis ordinoides. Nature 342, 542–543 776 (1989). 777 124. Auinger, H.-J. et al. Model training across multiple breeding cycles significantly improves 778 genomic prediction accuracy in rye (Secale cereale L.). Theor. Appl. Genet. 129, 2043– 779 2053 (2016). 780 781 782 783 784 125. Xie, K. T. et al. DNA fragility in the parallel evolution of pelvic reduction in stickleback fish. Science 363, 81–84 (2019). 126. Slate, J. INVITED REVIEW: Quantitative trait locus mapping in natural populations: progress, caveats and future directions. Mol. Ecol. 14, 363–379 (2005). 127. Brieuc, M. S. O., Waters, C. D., Drinan, D. P. & Naish, K. A. A practical introduction to 785 Random Forest for genetic association studies in ecology and evolution. Mol. Ecol. Resour. 786 18, 755–766 (2018). 787 788 789 128. Nielsen, R. Molecular Signatures of Natural Selection. Annu. Rev. Genet. 39, 197–218 (2005). 129. Barghi, N., Hermisson, J. & Schlötterer, C. Polygenic adaptation: a unifying framework to 790 understand positive selection. Nat. Rev. Genet. 1–13 (2020) doi:10.1038/s41576-020-0250- 791 z. 33 792 793 794 795 796 797 798 799 800 130. Lemos, B., Araripe, L. O. & Hartl, D. L. Polymorphic Y chromosomes harbor cryptic variation with manifold functional consequences. Science 319, 91–93 (2008). 131. Haddad, R., Meter, B. & Ross, J. A. The genetic architecture of intra-species hybrid mitonuclear epistasis. Front. Genet. 9, 481 (2018). 132. Long, A., Liti, G., Luptak, A. & Tenaillon, O. Elucidating the molecular architecture of adaptation via evolve and resequence experiments. Nat. Rev. Genet. 16, 567–582 (2015). 133. Bollback, J. P., York, T. L. & Nielsen, R. Estimation of 2Nes from temporal allele frequency data. Genetics 179, 497–502 (2008). 134. Marques, D. A., Jones, F. C., Di Palma, F., Kingsley, D. M. & Reimchen, T. E. 801 Experimental evidence for rapid genomic adaptation to a new niche in an adaptive 802 radiation. Nat. Ecol. Evol. 2, 1128–1138 (2018). 803 804 805 135. Svensson, E. I. & Calsbeek, R. The Adaptive Landscape in Evolutionary Biology. (Oxford University Press, 2012). 136. Hämälä, T., Gorton, A. J., Moeller, D. A. & Tiffin, P. Pleiotropy facilitates local adaptation 806 to distant optima in common ragweed (Ambrosia artemisiifolia). PLOS Genet. 16, 807 e1008707 (2020). 808 137. Roda, F., Walter, G. M., Nipper, R. & Ortiz‐Barrientos, D. Genomic clustering of adaptive 809 loci during parallel evolution of an Australian wildflower. Mol. Ecol. 26, 3687–3699 810 (2017). 811 812 813 814 815 34 816 817 818 Acknowledgements 819 We are grateful to Debora Goedert for comments on the first draft of this manuscript. E.I.S. and 820 A.R. were funded by grants from the Swedish Research Council (VR: grant no. 2016-03356 and 821 2018-04537, respectively). D. M. was supported by the Swiss National Science Foundation 822 grant #31003A_163338 to Ole Seehausen, Laurent Excoffier and Rémy Bruggmann. J.A.D. 823 acknowledges support by NSF 13-510 Systems & Synthetic Biology, award #1714550. J.M.H. is 824 supported by the German Federal Ministry of Education and Research (BMBF). K.C. is supported 825 by an SNF grant (CRSK-3_190288). K.M. was funded by the Australian Research Council 826 (DP190101661). M.N.S. was supported by Fundação de Amparo à Pesquisa do Estado de São 827 Paulo (FAPESP), projects 2015/19556-4 and 2016/22159-0. The authors also wish to thank Butch 828 Brodie for kindly providing us with the photograph of the garter snakes in Fig. 2A, and the original 829 figure of his classic fitness surface in Box 1. 830 831 832 833 Author contributions 834 E.I.S. and A.R. conceived of the paper, organized the writing and put together the first draft. All 835 authors contributed to writing, improving and finalizing the manuscript. 836 837 838 839 840 Competing interests The authors declare no competing interests. 841 842 843 35 844 Box 1: What is correlational selection? 845 846 Correlational selection involves several interrelated concepts. Here, we define the most important 847 terms. 848 849 The individual fitness surface 850 Imagine a function relating an individual’s trait values to that individual’s expected lifetime fitness 851 (Fig. 3). Supposing fitness depends on two traits, we can depict the fitness function as a three- 852 dimensional surface. The two horizontal axes represent trait values and elevation represents 853 fitness. Fitness peaks and valleys represent regions in trait space with high and low fitness, 854 respectively. Such a surface will reveal regions where favorable trait combinations produce high 855 fitness, as well as unfavorable trait combinations that confer low fitness. Individual fitness surfaces 856 can take almost any shape, including single-peaked surfaces, multi-peaked surfaces or ridges (Fig. 857 3), and can involve any number of traits. 858 859 Brodie’s pioneering study26 of coloration and behavior in garter snakes provided one of the first 860 empirical examples of how such individual fitness surfaces can illustrate correlational selection in 861 a natural population (see inset). A snake’s color pattern could be either blotched or striped. 36 862 Moreover, snakes either crawled in a straight line or reversed directions repeatedly when evading 863 predators. Striped snakes had higher fitness when they fled predators in a straight line, whereas 864 blotched snakes had higher fitness when they reversed directions. Therefore, survival depended on 865 the interaction between two types of traits – colour and behavior – rather than on single traits. 866 Interestingly, colouration and behavior were also genetically correlated with each other123, 867 providing empirical support for the prediction4,13 that correlational selection can promote and 868 maintain genetic correlations. 869 870 An operational definition of correlational selection 871 Correlational selection occurs when the relationship between an individual’s trait value and 872 expected fitness for one trait depends on that individual’s trait values for other traits, and direct 873 selection acts in such a way as to establish or maintain genetic and hence phenotypic correlations 874 among traits6. One way to think about correlational selection is to imagine slicing the fitness 875 surface parallel to one of the trait axes. If the slices differ in shape as we proceed along the fitness 876 surface (Fig. 3A), then fitness is the result of interactions between traits. 877 878 Lande and Arnold19 showed that correlational selection could be measured by using simple 879 regression approaches. If we assume that traits have a multivariate normal distribution, then the 880 fitness surface can be estimated by a regression of the form (in the bivariate case): 881 882 883 884 885 886 𝑤(𝑧1 , 𝑧2 ) = 𝛼 + 𝛽1 𝑧1 + 𝛽2 𝑧2 + 12𝛾11 𝑧12 + 12𝛾22 𝑧22 + 𝛾12 𝑧1 𝑧2 + 𝜀, (1) where α is an intercept, 𝑧1 and 𝑧2 are trait values, and 𝜀 is a residual term. The parameters 𝛽1 and 𝛽2 are the linear selection gradients, which estimate directional selection on each trait. The matrix 𝛾11 of quadratic selection coefficients 𝜸 = [𝛾 12 37 𝛾12 𝛾22 ] estimates stabilizing, disruptive and 887 888 correlational selection. The diagonal elements of 𝜸 measure quadratic selection on each trait (i.e., stabilizing or disruptive selection), whereas the off-diagonal elements represent correlational 890 selection. Thus, non-zero off-diagonal elements of 𝜸 constitute evidence of correlational selection. 891 Fitness epistasis and epistatic selection 892 Much of the fitness variation in complex phenotypic traits originates from allelic variation at 893 individual loci, each with small fitness effects. Favourable trait combinations at the organismal 894 level often also reflect favourable allelic combinations at separate sets of loci. At the genomic 895 level, correlational selection occurs when the fitness effects of a particular locus depend on the 896 genotype at another locus or, more generally, depend on the genetic background. This situation is 897 often described as fitness epistasis or epistatic selection and is likely to have a big impact on 898 genome evolution6,18. 889 899 900 901 902 903 904 905 906 907 908 909 910 911 38 912 Box 2. Methods to study genomic signatures of correlational selection 913 914 Genomics can inform quantitative genetics 915 Due to the highly polygenic nature of most traits, quantitative genetics is a pragmatic method to 916 predict short term evolutionary change in phenotypes58. Genomic tools can however be integrated 917 with quantitative genetics methodology to expand our understanding38,58,82. For example, the so- 918 called GBLUP approach82 allows the pedigree-relatedness matrix of an “animal model” to be 919 replaced by a marker-based relatedness matrix to infer genetic variances and covariances, i. e. G. 920 By accurately determining the proportion of genome shared, such genomic approaches may 921 improve the estimates of G compared to using pedigree data alone, where relatedness is based on 922 a shallow pedigree124. 923 924 Genomic approaches can also provide information about mutation rates of SNPs and indel variants, 925 thereby improving our understanding of the role of mutation rates in evolution125 and the 926 importance of mutational pleiotropy and M-matrix evolution83–85. Of particular interest is the 927 effects of new mutations on genetic variances and covariances, i. e. M83–85. A promising approach 928 is the combination of mutation accumulation experiments (MA) with estimates of M and G85–87. 929 Studies on MA-lines have revealed strong mutational pleiotropy across the transcriptome85. Such 930 strong mutational pleiotropy in M contrasts with weaker pleiotropy in G, suggesting that 931 correlational selection operates against maladaptive strong mutational covariance, which results in 932 a weakening of pleiotropy during the course of the life cycle83,84. 39 933 934 To quantify outcomes of correlational selection, we need to identify the genetic loci under such 935 selection. Traditionally, this has been achieved by quantitative trait loci (QTL) mapping, admixture 936 mapping and genome wide association studies (GWAS)126 which have limited power to detect 937 small effect size genes. Newer approaches map pleiotropy by simultaneously associating genomic 938 loci with multiple traits54 and can also detect epistatic interactions using machine learning 939 algorithms127. 940 941 Detecting the genomic signatures of correlational selection 942 Correlational selection could potentially be inferred from signatures of selective sweeps at loci 943 under strong selection128 or, for highly polygenic traits, allele frequency shifts that are not 944 explainable by genetic drift90,129. Selection on polygenic traits often leads to small frequency 945 changes in many genes, which is more difficult to detect129. Since correlational selection favors 946 certain allele combinations, one outcome is deviations from Hardy-Weinberg equilibrium and the 947 build-up of linkage disequilibrium between alleles at unlinked loci, detectable both across 948 individuals and between age classes within populations. Genomic data may also indicate whether 949 recombination suppression leading to trait correlations67, such as in supergenes or genomic 950 rearrangements, has been favored by correlational selection. On longer time scales, genomic data 951 can reveal how such supergenes are gradually built up and assembled via gene duplications and 952 neofunctionalization 75. Experimental assays such as introgression lines130 or reciprocal crosses of 953 diverged lineages131 can be used to confirm whether combinations of alleles or genomic regions 954 are under correlational selection. Evolve and re-sequence experiments comparing populations 40 955 before and after selection132, or studies of allele frequency time series during an experiment133 can 956 give further, detailed insight into allelic interactions and both genomic and phenotypic responses 957 to experimentally imposed selection62,134. 958 959 Bridging the genotype-phenotype-fitness map 960 Ideally, the genotypic and phenotypic levels should be studied alongside the adaptive 961 landscape108,135 and integrated into a genotype-phenotype-fitness map. This integration has been 962 achieved for very few non-model organisms such as threespine stickleback29, Bahama pupfish30 963 and Tinema stick insects59 in which the fitness landscape was mapped experimentally with 964 information about the genomic architecture of traits. Experimental field studies on fitness epistasis 965 in natural populations combined with genomic data is a promising integrative approach to detect 966 the genomic consequences of correlational selection59. 967 968 41 969 Legends to figures 970 42 971 Figure 1. The scope of correlational selection and its links to different fields in evolutionary 972 biology. Correlational selection is relatively well-understood statistically and theoretically (Box 973 1), but we still do not know its prevalence in natural populations and the extent to which it has 974 shaped genome evolution in diverse organisms. In a few cases, correlational selection has been 975 studied and documented in natural field populations and in laboratory artificial experimental 976 studies (Fig. 2). The main effects of correlational selection are to strengthen or reinforce 977 phenotypic and/or genetic correlations between traits6,22,23,26, which may be governed by separate 978 sets of loci, or to break up non-adaptive or maladaptive genetic correlations, such as between the 979 sexes44. These effects of correlational selection on phenotypic and potentially also genetic 980 correlation structure have consequences for several organismal-level phenomena that are of great 981 interest in evolutionary genetics and developmental biology. These include G-matrix evolution, 982 phenotypic plasticity, modularity, evolvability and phenotypic integration (upper part of figure), 983 as discussed in this review. Theory suggests that correlational selection at the organismal level can 984 potentially drive the downstream evolution of genomic architecture4,6,18 (lower part of figure), 985 although here our knowledge is more limited. Correlational selection could preserve adaptive 986 genetic correlations between traits that are governed by different sets of loci by suppressing 987 recombination rates, thereby maintaining inversion polymorphisms and other structural genomic 988 variation that is often associated with balanced genetic polymorphisms (Fig. 4). In addition, 989 correlational selection could lead to adaptive pleiotropy, such as during range expansions when 990 populations are far away from their adaptive peaks136, and could shape patterns of epistasis 991 between loci12. Finally, correlational selection is likely to be involved in local adaptation, if 992 different sets of character combinations are favoured in different abiotic137 or biotic 43 993 environments27,28, but the consequences for speciation and other aspects of macroevolution remain 994 largely unexplored. 995 996 44 997 998 Figure 2. Phenotypic and quantitative genetics studies on organisms and traits in which 999 correlational selection has experimentally been demonstrated or inferred in the field or in 1000 the laboratory. A. Northwestern garter snake (Thamnophis ordinoides). B. Side-blotched lizard 1001 (Uta stansburiana). C. Australian fruit fly (Drosophila serrata). D. Western bluebird (Sialia 1002 mexicana). E. Dark-eyed junco (Junco hyemalis). F. Guppy (Poecilia reticulata). G. Three-spined 1003 stickleback (Gasterosteus aculeatus). H. Fire pink (Silene virginica). I. White campon (Silene 1004 latifolia). Correlational selection has been demonstrated and quantified for a number of different 1005 traits, including both discrete colour polymorphisms6,8,23,26 (A, B, F) and continuous, quantitative 1006 characters15,24,25,28,44,94 (C, D-E, G-I), both in animals and in plants. The ecological causes and 1007 selective agents driving such correlational selection have been shown to be predators (A,G), 1008 interspecific mutualists such as pollinators (H) and conspecific interactions, especially under 1009 sexual selection (B-C, E-F). In some of these studies, the phenotypic traits that were found or 1010 implicated to be under correlational selection were also genetically or phenotypically correlated 45 1011 with each other (A-B, D-E), suggesting that correlational selection can build up, promote or 1012 strengthen genetic integration between the traits in question. Conversely, artificial correlational 1013 selection has been demonstrated to be efficient in breaking up an intersexual genetic correlation in 1014 at least one case (I). Finally, traits that have been found to experience correlational selection 1015 include visual colouration traits (A,B,E,F), chemical communication traits (C), behavioural traits 1016 such as dispersal, aggression and personality (D,G) and structural traits such as size and shape (H). 1017 Photo credits: A: Butch Brodie. B. Barry Sinervo. C-I: Public domain. C. Antoine Morin: 1018 https://www.eurekalert.org/multimedia/pub/94488.php. 1019 https://en.wikipedia.org/wiki/Dark-eyed_junco#/media/File:Dark-eyed_Junco-27527.jpg E. 1020 Wikimedia 1021 eyed_Junco-27527.jpg 1022 https://www.nature.com/articles/nature12717/figures/1?draft=collection 1023 G. 1024 https://commons.wikimedia.org/wiki/File:GasterosteusAculeatusMaleHead.JPG 1025 H. 1026 https://en.wikipedia.org/wiki/Silene_virginica#/media/File:Silene_virginica_Arkansas.jpg 1027 I. 1028 https://en.wikipedia.org/wiki/Silene_latifolia#/media/File:Silene_latifolia_9631.JPG Commons (Ken Wikimedia Wikimedia Wikimedia Thomas): D. Wikimedia Commons: https://commons.wikimedia.org/wiki/File:DarkKimberly F. Commons Commons Commons 1029 1030 1031 1032 46 Hughes/Nature: (Piet Spaans): (Eric (Walter Hunt): Siegmund): 1033 47 1034 Figure 3. Illustration of correlational selection, along with important parameters used to 1035 quantify it and determine how its effects are carried across generations. A. Example fitness 1036 surfaces for hypothetical traits 𝑧1 and 𝑧2 (top row) and conditional fitness curves for 𝑧1 given fixed 1037 1038 1039 1040 values of 𝑧2 (colored lines in both rows). A. When selection is additive, the fitness effects of 𝑧1 are independent of the value of 𝑧2 . The conditional fitness curves are then identical aside from their height above the trait axes (bottom row, left of the dashed line). Under correlational selection, in contrast, the fitness effects of 𝑧1 depend on 𝑧2 , and so the shape of the conditional fitness curves 1042 changes with value of 𝑧2 (bottom row, right of the dashed line). B. Estimation of multivariate 1043 fitness surface (first column) is unobservable but can be sampled by measuring the relative fitness 1044 of individuals in a population (second column). The true surface can then be estimated via 1045 quadratic regression (third column) or by non-parametric smooth splines (fourth column). See 1046 Section 3 in Supplementary Material for full details. C. The 𝑮-matrix (orange, left) is the variance- 1041 1047 fitness surfaces from samples of individual trait values, 𝑧1 and 𝑧2 , and relative fitness, 𝑤. The true covariance matrix of additive genetic effects (i.e. breeding values) for a multivariate phenotype. 1049 The 𝑴-matrix (blue, right) is the variance-covariance matrix of additive mutational effects. Points 1050 If the distribution of point values is multivariate normal, it can be summarized via an ellipsoid. 1051 The principal axes of the ellipsoid (crossed lines) align with the eigenvectors and their lengths are 1052 proportional to the square roots of the eigenvalues. The major axis, associated with the largest 1053 eigenvalue, indicates the direction of maximum additive genetic or mutational variance. 1048 represent individual breeding values (orange) and additive mutational effects (blue) respectively. 1054 1055 48 1056 1057 Figure 4. Examples of genomic trait architectures that might reflect past or ongoing 1058 correlational selection. We focus here on empirical examples where multiple loci are involved in 1059 the adaptive traits in question, as these reflect the most challenging situations to maintain adaptive 1060 genetic correlations between traits, due to the eroding effects of recombination when traits are 1061 governed by multiple unlinked loci. However, we underscore that correlational selection could 1062 equally well lead to the evolution of adaptive pleiotropy136 as an alternative mechanism to maintain 1063 adaptive genetic correlations between traits. A. A complex mating polymorphism in male ruff 1064 (Philomachus pugnax) reproductive tactics involves multiple correlated morphological and 1065 behavioral traits, and the different character combinations in the male morphs are preserved 1066 because of the lack of recombination between different loci that are held together in a single large 1067 chromosomal inversion73,74.B. Assortative mating maintains linkage disequilibrium between 1068 unlinked color pattern loci under correlational selection in Heliconius butterfly species, facilitated 1069 by tight linkage between preference and trait loci on one chromosome76. C. In a multifarious 1070 selection experiment on threespine sticklebacks (Gasterosteus aculeatus), the predicted 1071 phenotypic changes in multiple traits were caused by widespread underlying genomic changes that 1072 could potentially be attributed to correlational selection for different character combinations in the 1073 different phenotypes134. 49