The Molecular Basis of Socially-Mediated Phenotypic Plasticity in a Eusocial Paper Wasp 2 3 Benjamin A

bioRxiv preprint doi: https://doi.org/10.1101/2020.07.15.203943; this version posted July 15, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

1 The molecular basis of socially-mediated phenotypic plasticity in a eusocial paper wasp 2 3 Benjamin A. Taylor1,2, Alessandro Cini1,2,3, Christopher D. R. Wyatt1,2, Max Reuter2,4* and Seirian 4 Sumner1,2* 5 6 7 Addresses

8 1 Centre for Biodiversity & Environment Research, University College London, Gower Street,

9 London WC1E 6BT, United Kingdom

10 2 Department of Genetics, Evolution & Environment, University College London, Gower Street,

11 London WC1E 6BT, United Kingdom

12 3 Dipartimento di Biologia, Università degli Studi di Firenze, Via Madonna del Piano 6, 50019 Sesto

13 Fiorentino, Italy

14 4 Centre for Life's Origins and Evolution, University College London, Gower Street, London WC1E

15 6BT, United Kingdom

16 * Authors contributed equally

18 Corresponding author:

19 Benjamin A Taylor

20 Tel: +447527477538

21 Email: [email protected]

22 Address: Centre for Biodiversity & Environment Research, University College London, Gower

23 Street, London WC1E 6BT, United Kingdom

24 25

1 bioRxiv preprint doi: https://doi.org/10.1101/2020.07.15.203943; this version posted July 15, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

26 Abstract 27 28 Phenotypic plasticity, the ability to produce multiple phenotypes from a single genotype, represents 29 an excellent model with which to examine the relationship between gene expression and 30 phenotypes. Despite this, analyses of the molecular bases of plasticity have been limited by the 31 challenges of linking individual phenotypes with individual-level gene expression profiles, 32 especially in the case of complex social phenotypes. Here, we tackle this challenge by analysing 33 the individual-level gene expression profiles of Polistes dominula paper wasps following the loss of 34 a queen, a perturbation that induces some individuals to undergo a significant phenotypic shift and 35 become replacement reproductives. Using a machine learning approach, we find a strong 36 response of caste-associated gene expression to queen loss, wherein individuals’ expression 37 profiles become intermediate between queen and worker states. Importantly, this change occurs 38 even in individuals that appear phenotypically unaffected. Part of this response is explained by 39 individual attributes, most prominently age. These results demonstrate that large changes in gene 40 expression may occur in the absence of detectable phenotypic changes, resulting here in a socially 41 mediated de-differentiation of individuals at the transcriptomic but not the phenotypic level. Our 42 findings also highlight the complexity of the relationship between gene expression and phenotype, 43 where transcriptomes are neither a direct reflection of the genotype nor a proxy for the molecular 44 underpinnings of the external phenotype. 45

2 bioRxiv preprint doi: https://doi.org/10.1101/2020.07.15.203943; this version posted July 15, 2020. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.

46 Introduction 47 48 The relationship between gene expression and external phenotype is complex and unresolved. 49 Much research in behavioural and evolutionary ecology is based on the implicit assumption that 50 phenotypic traits can be modelled as though they directly reflect gene expression patterns, and 51 that evolutionary trajectories can therefore be studied while remaining agnostic with regard to the 52 underlying molecular mechanisms (Hadfield et al 2007; Rittschof & Robinson 2014; Rubin 2016). 53 This ‘phenotypic gambit’ has proven a useful rule of thumb, permitting the establishment of a rich 54 body of literature surrounding the evolution of complex traits, despite a lack of data relating to the 55 genetic basis of these traits (e.g. Réale et al 2010; Chapman et al 2011; Fowler-Finn & Rodríguez 56 2012). In the past decade, however, advances in the affordability of ‘omic’ data and availability of 57 powerful bioinformatic methods have greatly enhanced our ability to assess the assumptions made 58 by the phenotypic gambit (Rittschof & Robinson 2014; Heyes 2016). The time is right to 59 disentangle the molecular bases of complex phenotypic traits. 60 61 Phenotypic plasticity, the ability of an individual to effect phenotypic changes in response to 62 external cues, is an ideal phenomenon with which to study the relationship between gene 63 expression and phenotype because it involves the production of multiple phenotypes without gene 64 sequence changes. Of particular value are species in which adult individuals can be experimentally 65 induced to transition between distinct, measurable phenotypes. By comparing the gene expression 66 profiles of groups of individuals that differ in their phenotypes as a result of plasticity, it is possible 67 to isolate phenotypic effects of gene expression while controlling for genetic differences. Using this 68 approach, significant progress has been made in unravelling the molecular underpinnings of 69 sequential sex changes in hermaphroditic fish (Horiguchi et al 2013; Casas et al 2016; Todd et al 70 2019), the distinct gregarious social phenotype of desert locusts (Cullen et al 2017; Lo et al 2018), 71 and the reproductive castes of social insects (Simola et al 2016; Libbrecht et al 2018; Rehan et al 72 2018; Shell & Rehan 2018). Such studies typically rely on comparisons between groups of 73 individuals with well-differentiated phenotypes, however. As a result, little is known about more 74 subtle effects during the transition from one morph to the other, and the relationship between 75 expression patterns and phenotypic traits at the individual level. 76 77 The reproductive castes found in the colonies of social insects provide excellent model systems for 78 determining the extent to which fine-scale changes in phenotype are reflected at the molecular 79 level. With a few exceptions (reviewed in Schwander et al 2010), the distinct queen and worker 80 phenotypes found in such colonies are plastically determined either during development or in 81 adulthood. Workers in some species can be experimentally induced to transition to a reproductive 82 role in response to the removal of a colony’s queen (e.g. Strassmann et al 2004; Tibbetts & Huang 83 2010) or as a result of exposure to varying levels of brood (e.g. Chandra et al 2018; Libbrecht et al

84 2018), allowing changes in the behavioural, physiological and molecular traits that define caste 85 identity to be tracked. An additional benefit—and challenge—of studying social insect colonies is 86 that they involve complex social structures. Such interactions can be hard to study, but offer the 87 opportunity to assess the effects of social interactions upon phenotypes and transcriptomes. 88 89 In the present study, we explore the relationship between transcriptome and phenotype in the 90 context of plastic caste expression in the European paper wasp Polistes dominula (Christ 1791), 91 which is a model organism often used in studies of social insect behaviour (reviewed in Starks & 92 Turillazzi 2006; Jandt et al 2014) and, more recently, for analyses of caste gene expression (e.g. 93 Patalano et al 2015; Standage et al 2016; Geffre et al 2017; Manfredini et al 2018). In this species, 94 removing the established queen from a single-foundress colony induces a queen succession 95 process in which one (or very few) workers transition to a queen phenotype, with age playing a key 96 role in predicting which individual will do so (Pardi 1948; Tsuji &Tsuji 2005): almost invariably, the 97 new queen is one of the oldest individuals, and there is little conflict over succession (Strassmann 98 et al 2004; Taylor et al 2020). In a recent paper (Taylor et al 2020) we followed responses to queen 99 removal in P. dominula on a fine scale by measuring individual-level behavioural and physiological 100 traits and generating a univariate measure of caste identity (‘queenness’) to describe individuals’ 101 phenotypic profiles (Box 1). We found that phenotypic responses were limited to the subset of 102 individuals that transitioned to the queen role, while other individuals exhibited little or no 103 measurable change in measured traits. This system, in which individuals within a controlled 104 environment vary strongly and predictably in their phenotypic response to a shared stimulus, 105 affords an excellent opportunity to track the relationship between gene expression and phenotypic 106 expression at the individual level. 107 108 To do so, here we complement the individual-level phenotypic data of Taylor et al. (2020) by 109 assaying the brain gene expression profiles of 107 individuals for which fine-scale behavioural and 110 ovarian data have been obtained, including queens and workers from stable colonies and 111 individuals from colonies that had their queens experimentally removed. We then assess the 112 relationship between the transcriptome and the phenotypic changes induced by queen removal. To 113 match the univariate measure of caste identity that we applied to our phenotypic data (Box 1), we 114 use a support vector classification approach to generate a univariate measure of caste-associated 115 gene expression. Support vector classification is a powerful tool with which to transform complex 116 patterns in multidimensional data into a continuous classification score, allowing the detection of 117 subtle, widespread signals of differential expression between phenotypic states that are likely to be 118 missed in conventional differential expression analyses. Support vector machines (SVMs) have 119 become a key tool in the early identification of phenotypically indistinguishable cancer subtypes 120 (Segal et al 2003; Abeel et al 2010; Huang et al 2018) and their value has recently been 121 demonstrated in animal behaviour studies (e.g. Chakravarty et al 2019; Rast et al 2020). Given

122 their sensitivity, SVMs should be ideally suited to quantify gene expression variation across the 123 spectrum between differentiated worker and queen roles, but to our knowledge this is the first time 124 SVMs have been used to interrogate the molecular basis of behavioural traits. 125 126 Applying an SVM approach along with standard differential expression and gene co-expression 127 analyses, we show that brain gene expression responses to queen removal in P. dominula include 128 a colony-wide response that does not match that observed at the phenotypic level. Our results 129 indicate that gene expression in P. dominula colonies reflects both a generalised response to 130 queen loss which is independent of phenotype, and a phenotype-specific response that tracks 131 individuals’ expression of plastic phenotypic changes. To our knowledge, this study provides the 132 most comprehensive analysis to date of the ways that plastic phenotypes are reflected at the 133 transcriptomic level; our results expose the complexity of the relationship between individual-level 134 gene expression and individual phenotypes and call into question the assumptions of the 135 phenotypic gambit. 136

137 138 Box 1. Summarized experimental design describing the generation of phenotypic data pertaining to individual-level responses to queen 139 removal in Polistes dominula (Taylor et al 2020) used in this study. (A) Early-season nests were transferred to the lab prior to the 140 eclosion of workers and subordinate foundresses were removed. After at least 4 workers had eclosed, nests were filmed for three days 141 and then assigned to either control or queen removal treatments. Following treatment (sham or queen removal) nests were filmed for a 142 further three or twelve days and then all individuals were dissected to generate ovarian development indices following Cini et al (Cini et 143 al. 2013). Footage of nests was used to generate dominance indices in the form of Elo ratings (Elo 1978; Neumann et al. 2011). (B) 144 Ovarian and dominance indices from queens and control workers were used to produce a logistic regression model for caste 145 classification (0=worker, 1=queen). Data from individuals on queenless post-removal nests were then passed through this model to fit

146 caste estimates and thereby identify individuals with high ‘queenness’, i.e. those that exhibited strongly queen-like phenotypes following 147 queen removal and thus represented possible replacement queens. Of the individuals for which data are shown here, 27 queens, 12 148 workers from control nests, and 62 individuals from queen removal nests were subsequently sequenced to generate the data discussed 149 in the present study. 150 151 Results 152 153 Support vector classification reveals consistent patterns of caste gene expression 154 differentiation involving many genes 155 156 Our support vector approach allowed us to reduce complex variation in brain gene expression data 157 down to a single dimension of caste identity, matching the Bayesian logistic regression model we 158 used in Taylor et al. (2020) to condense ovarian and behavioural data into a unidimensional metric 159 of phenotypic caste (‘queenness’; Box 1). We trained an SVM using gene expression data from 27 160 queens and 12 workers from stable, unmanipulated colonies. An initial model, based on all 10734 161 genes annotated in our experiment, achieved a root mean squared validation error of 0.065 in 162 three-fold cross-validation, i.e. a model trained on a random subset of two thirds of workers and 163 queens classified the remaining third within 25.5% of their true values (worker=0, queen=1). By 164 applying a process of iterative feature selection we were able to remove genes that had low 165 weights in the classification and contributed to overfitting (Figure S1A; Table S1). Doing so 166 allowed us to identify an optimal model containing 1992 genes with a substantially reduced root 167 mean squared classification error of 0.021 (Table S2). This model classified queens and workers 168 very consistently, with strong separation of queens from workers (Figure 1A). Thirty-four gene 169 ontology (GO) terms were significantly enriched among these 1992 genes, including a number of 170 terms associated with translation such as rRNA processing, tRNA aminoacylation, and ribosomal 171 large subunit biogenesis (Figure S1B). 172

173 174 Figure 1. Classifier estimates generated by the lowest-error SVM using 1992 caste-informative genes for (A) queens 175 and workers from control colonies and (B) individuals from experimental colonies following queen removal. Mean and 176 standard deviation for each group are shown in black. 177 178 The 1992 genes in our optimised SVM represent a much larger set than that which is found to 179 differentiate queens and workers based on a standard differential expression approach. When we 180 applied a DESeq2 analysis with a 1.5 fold-change threshold to the same set of queens and 181 workers used to train the SVM, we identified just 81 differentially expressed genes (with no 182 associated GO terms), a small number that is typical for similar analyses in Polistes (e.g. Toth et al 183 2014; Patalano et al 2015; Geffre et al 2017). All but four of these genes were present in the larger 184 SVM set (Figure S1C; Table S3), suggesting that the SVM captures almost all the information 185 obtained from these standard methods. Conversely, an SVM trained using just these 81 186 differentially-expressed genes exhibited a cross-validation error rate of 0.065, no better than the 187 original un-optimised model. Thus, the picture of caste differentiation provided by standard 188 differential expression analysis appears to miss a great number of subtle differences in individuals’ 189 gene expression profiles that contribute to caste differentiation. 190 191 192 Colony-wide brain gene expression responses to queen removal 193

194 Following the loss of a queen from a Polistes colony, typically one or a few individuals undergo a 195 phenotypic transition to become a replacement queen while the rest of the colony members remain 196 workers (Miyano 1991; Dapporto et al 2005; Taylor et al 2020). If this response is reflected at the 197 level of gene expression, then we would expect to observe significant changes in the 198 transcriptional profiles of individuals that exhibit a phenotypic transition, but not in those that do 199 not. Our individual-level expression data allowed us to track responses in 62 individuals’ gene 200 expression profiles following queen removal across 19 nests. We analysed individuals’ gene 201 expression profiles at three days after queen removal, when queen replacement is ongoing (n=24), 202 and at twelve days after queen removal, when succession is largely settled at the phenotypic level 203 (Strassmann et al 2004; Taylor et al 2020; n=38). Individuals for sequencing were selected to cover 204 a wide range of phenotypes, including those that remained entirely worker-like, those that had 205 transitioned to highly queen-like phenotypes, and those with intermediate phenotypes at the time of 206 sampling (Figure S1D-F). 207 208 Analysing shifts in SVM estimates of individual wasps allowed us to assess the degree to which 209 these concurred with the changes visible at the phenotypic level. Doing so, we found that the SVM 210 estimates of individuals from post-removal nests were intermediate between those of queens and 211 workers from control nests (Figure 1B), a finding which concurs with the placement of these 212 individuals according to principal component analysis (Figure S1G). This result is surprising given 213 that the majority of individuals on queen removal nests are phenotypically indistinguishable from 214 workers on control nests in terms of ovarian development and behavioural dominance (Taylor et al 215 2020). Thus, the large majority of individuals on queen removal nests exhibited perturbation of their 216 caste-associated gene expression, even though only a few of these individuals exhibited 217 responses to queen loss at the level of physiology or behaviour and transitioned from a worker to a 218 queen role. 219 220 Furthermore, we found no evidence that the degree of transcriptional perturbation declined over 221 time following queen removal. SVM classification estimates of individuals from queen removal 222 nests did not differ significantly between days three and twelve following queen removal (QR3 223 mean 0.331±0.111; QR12 mean 0.345±0.082; Wilcoxon W=369, p=0.55). The effects of queen 224 removal on individuals’ gene expression profiles therefore appear to be both widespread and 225 persistent, affecting all individuals in a nest and lasting beyond the point at which a new queen has 226 already become phenotypically established. 227 228 Interestingly, the strong colony-wide perturbation following queen loss that this SVM approach 229 identifies would have been entirely missed using a standard differential expression approach: 230 DESeq2 identified just five genes as differentially expressed between control workers and 231 individuals from manipulated nests, with no associated GO enrichment (Table S4).

232 233 234 Age and queenness explain variation in individual-level molecular responses to queen loss 235 236 While SVM classification indicates that queen removal causes colony-wide perturbation to brain 237 expression profiles, classifier estimates varied substantially between individuals following queen 238 removal, spanning a much greater range of values (0.116-0.540) than those of queens (0.900- 239 1.100) or workers (0.057-0.100) from queenright colonies (Figure 1). To better understand this 240 variation, we examined whether the classifier estimates for individuals from manipulated colonies 241 were predicted by those individuals' phenotypic traits—specifically ovarian development, 242 behavioural dominance, and age (Box 1; Taylor et al 2020). 243 244 Our phenotypic measure of caste identity (queenness) was a significant predictor of expression- 245 based SVM classifier estimates when fitted using a linear model (slope±SE = 0.0414±0.0114, p = 246 6.0×10-4; Figure 2A). The individual components of queenness, ovarian development and 247 dominance were also significant or near-significant predictors of SVM classification individually 248 (ovarian development: slope±SE = 0.0426±0.0113, p = 4.0×10-4; dominance: slope±SE = 249 0.0231±0.0122, p = 0.06). Notably, however, because phenotypic queenness was strongly 250 correlated with age among post-removal individuals (cor = 0.4832, p = 8.0×10-5; Taylor et al 2020), 251 age was an equally strong predictor of caste estimates when fitted in a separate linear model 252 (slope±SE = 0.0516±0.0106, p = 9.8×10-6; Figure 2B). Thus, the significance of queenness as a 253 predictor of caste estimates might have been an artefact of the fact that both are correlated with 254 age. To test whether queenness had an effect over and above that accounted for by age, we 255 calculated the residuals of phenotypic queenness on age and fitted the caste estimates against 256 these. These residuals were not significantly predictive of SVM classification (slope±SE = 257 0.0202±0.0123, p = 0.11; Figure 2C), nor were the residuals of ovarian development alone on age 258 (slope±SE = 0.0223±0.0123, p = 0.07) or the residuals of dominance alone (slope±SE = 259 0.0114±0.0126, p = 0.37). Age is thus the strongest determinant both of individuals’ caste 260 phenotypes and of their caste-associated gene expression.

261 262 263 Figure 2. (A-C) Scatterplots of SVM classifier estimates for individuals from post-removal colonies plotted against (A) 264 phenotypic queenness, (B) age, and (C) the residuals of queenness on age respectively. (D-F) Module-trait correlations

265 for consensus modules identified by WGCNA in queen removal individuals, control workers and queens respectively. 266 Within each cell, the Pearson correlation coefficient is given above and the FDR-adjusted p-value is given below in 267 parentheses. (G-H) Enriched gene ontology terms at p<0.01 for gene module 11 and gene module 8. 268 269 In order to further discern the factors that shaped individuals’ gene expression profiles following 270 queen removal, we performed weighted gene co-expression network analysis (WGCNA) using the 271 full set of annotated genes. We generated 22 consensus modules across all samples, and then 272 determined which traits significantly predicted the expression of a given module within each group 273 (Figure 2D-F). In doing so, we identified three modules of particular interest. 274 275 Module 11 (Red) consists of 519 genes associated with 14 enriched GO terms, including a number 276 of terms associated with molecular binding (Figure 2G). This module is notable in that its 277 expression in individuals from queen removal colonies is positively associated with SVM 278 classifications but not with phenotypic correlates of caste. A second module, Module 2 279 (Lightyellow), consists of 102 genes and appears to be associated with many caste-related traits in 280 the queen removal condition, being negatively correlated with queenness, age, SVM classification 281 and (more weakly) with ovarian development among queen removal individuals. We also found 282 weak evidence that this module is negatively associated with age among workers from control 283 nests. Unexpectedly, despite the seeming importance of Module 2 in predicting caste identity 284 among workers, not a single GO term was enriched among the genes in the module at p<0.01. 285 Finally, Module 8 (Yellow), consists of 614 genes and has a strongly negative correlation with age 286 among both control workers and queen removal individuals. This module is associated with 21 GO 287 terms, including a number of terms associated with molecular binding, such as protein binding, 288 DNA binding, RNA binding and chromatin binding (Figure 2H). The latter term is particularly 289 notable: regulation of chromatin accessibility is one of several epigenetic processes that have been 290 implicated in the control of caste expression in social insects (Simola et al 2016; Wojciechowski et 291 al 2018; Duncan et al 2020). Downregulation of this module with age might plausibly contribute to 292 the increased phenotypic plasticity exhibited by older individuals in P. dominula. 293 294 295 296 Discussion 297 298 This study provides one of the most comprehensive analyses to date regarding the transcriptomic 299 signatures of plastic phenotypes, and the first to employ an SVM approach to interrogate the 300 molecular basis of behaviour. We have applied this approach to identify caste-specific gene 301 expression profiles in P. dominula and to explore the relationship between transcriptomic and 302 plastic phenotypic changes following a major social perturbation in paper wasp colonies. We 303 identify a set of nearly 2000 genes that optimally capture gene expression differences between

304 established queens and workers. Using a caste classifier based on these genes, we find that 305 queen removal leads to a colony-wide shift in expression, where the expression profiles of all 306 individuals move towards a state intermediate between those of established queens and workers. 307 Individual variation around this intermediate state is related to age and phenotypic attributes, with 308 older (and therefore more queen-like) individuals showing expression profiles that are closer to that 309 expected for established queens. Our results show that molecular responses to queen removal in 310 P. dominula consist of both a general colony-wide response independent of phenotypic change and 311 a response that reflects the plastic phenotypic transition. 312 313 Our study contributes to the enormous progress in our understanding of the relationship between 314 molecular changes and changes in phenotypic expression that has been made in the past decade, 315 facilitated by the increased availability of 'omic’ data and complex bioinformatic analyses. Recent 316 studies have started to challenge the view that there is a direct correspondence between 317 transcriptomic states and external phenotypes. Libbrecht et al (2018), for example, show that gene 318 expression responses associated with a reversible phenotypic change differ qualitatively based on 319 the directionality of the change (from reproductive to non-reproductive or vice versa). Meanwhile, 320 molecular manipulations have revealed a surprising degree of plasticity in canonically implastic 321 traits such as mammalian sex (Matson et al 2011) or ant castes (Simola et al 2016). Our results go 322 further, showing a shift in caste-specific brain gene expression profiles among individuals whose 323 phenotypic caste expression remains otherwise apparently unchanged. This shows that the 324 expectation of a close match between expression profiles and phenotypes is excessively simplistic, 325 or at least that detecting such a match requires detailed knowledge of relevant genes and/or an 326 exhaustive phenotypic characterisation. This, in turn, suggests that the use of expression data to 327 infer the molecular basis of phenotypes is more challenging than hitherto appreciated. 328 329 A major advantage of this study is the use of individual-level gene expression data from a large 330 number of subjects, including individuals reared in a shared social environment but exhibiting very 331 different phenotypic responses to perturbation. By sequencing individuals rather than pools, we 332 were able to match each gene expression profile to high-resolution phenotypic data that captures 333 the scale of naturally-occurring variation in features such as age, ovarian development and 334 dominance behavior. This resolution allows us to address questions that are otherwise 335 inaccessible in gene expression analyses. For example, we have been able to show that caste 336 identity, but not the residuals of caste identity on age, are significantly predictive of individuals’ 337 change in transcriptomic caste identity following queen loss. 338 339 Our discovery of colony-wide responses to queen loss suggests that this social perturbation 340 provokes a significant reaction even from individuals that have little hope of attaining the vacant 341 reproductive role. This is a surprising finding given that P. dominula is thought to express a

342 ‘conventional’ gerontocratic mechanism of dominance and queen succession that mitigates the 343 need for costly intragroup conflicts over the identity of the replacement queen (Pardi 1948; Tsuji & 344 Tsuji 2005; Monnin et al 2009), which should greatly reduce the need for young, low-ranking 345 workers to respond to queen loss (Taylor et al 2020). The gene expression responses of lower- 346 ranked workers to queen removal might plausibly represent a form of safeguard against queen 347 loss: if queen loss sometimes occurs multiple times in quick succession or is frequently associated 348 with a general decimation of the nest population (i.e. through predation), there might be kin- 349 selected benefits of a colony-wide ‘de-differentiation’ of individuals that facilitates a quicker 350 succession process. 351 352 Replacement queens in our colonies did not have access to males and therefore remained 353 unmated even after queen succession. This may partially explain the fact that even individuals with 354 fully developed ovaries and very high dominance ratings did not transition to a fully queen-like 355 gene expression profile, as mating can induce significant gene expression changes in insects (e.g. 356 Gomulski et al 2012; Zhou et al 2014). The lack of immediate mating opportunities for new queens 357 in our experiment is not necessarily unrealistic, however: unmated Hymenopteran females can lay 358 unfertilized eggs, which develop as males. Moreover, in naturally-occurring early P. d o m i nu l a 359 nests, replacement queens may be established a month or more before they are mated 360 (Strassmann et al 2004), presumably due to a scarcity of early males. The unmated replacement 361 queens analysed here are therefore representative of those that would be present on wild nests 362 shortly after queen loss. 363 364 To our knowledge, this study represents the first application of a support vector classification 365 approach to behavior-associated transcriptomic data. Using this approach we identified a large 366 group of genes as differing meaningfully between Polistes castes—over 10% of annotated genes, 367 a much larger set than those found using standard analytic approaches in this study and others 368 (e.g. Ferreira et al 2013; Toth et al 2014; Patalano et al 2015; Geffre et al 2017). Standard 369 approaches using packages such as edgeR (Robinson et al 2010), DESeq2 (Love et al 2014), or 370 NOISeq (Tarazona et al 2011), which typically include information-sharing between genes and 371 relatively strict fold change cutoffs, allow differential expression to be assessed with a high degree 372 of confidence at the level of individual genes. However, because such approaches assess genes 373 individually and do not reduce the dimensionality of the samples’ transcriptional profiles, they are 374 not well-suited to the tracking of subtle but consistent broad-scale changes. This is especially true 375 for gene expression data that are noisy and heterogenous, or those that are dominated by many 376 genes of small effect (Huang et al 2018). The molecular bases of caste in simple social insect 377 societies represent such a class of data. 378

379 Using a machine learning approach, here we have undertaken a detailed analysis of the 380 relationship between gene expression and socially-mediated phenotypic plasticity, revealing broad- 381 scale changes in caste-associated gene expression profiles following a major social disruption. Our 382 results reveal a hitherto unrecognized capacity for large scale disruption to caste-biased gene 383 expression profiles even in the absence of apparent changes in caste phenotype, a disconnect that 384 undermines simplistic models of the relationship between transcriptome and phenotype. Future 385 studies should continue to marry detailed phenotypic and gene expression data in order to assess 386 the prevalence and provenance of such discontinuities. 387

388 389 Acknowledgements 390 391 We would like to thank F. Cappa, I. Pepiciello, L. Dapporto and others at the Dipartimento di 392 Biologia, Università di Firenze for assistance with field work, and E. Favreau and M. Bentley for 393 their assistance with analyses. Special thanks to R. Cervo, without whose generous support this 394 study would not have been possible. This work was supported by the Natural Environment 395 Research Council (grant code NE/L002485/1). 396 397 398 Author contributions 399 400 BAT, SS and MR conceived of and designed the experiment. BAT and AC carried out the initial 401 experiment. BAT performed behavioural analyses, dissections and RNA extractions, and wrote the 402 initial draft of the manuscript. BAT conducted the analyses with advice from CDRW, MR and SS. 403 BAT and CDRW designed the figures. All authors read and contributed to the manuscript. 404 405 406 Data availability 407 408 Sequencing data associated with this paper have been deposited in the NCBI Gene Expression 409 Omnibus (www.ncbi.nlm.nih.gov/geo) under accession number GSE153532 with reviewer access 410 token adctccsahjgjbqt.

411 Supplementary methods 412 413 Gene expression quantification 414 415 Brain tissue was extracted from the heads of individual samples and RNA was extracted using the RNeasy 416 Mini Kit (Qiagen) according to manufacturer’s instructions. Library preparation was performed by Novogene 417 Co. followed by sequencing on an Illumina HiSeq 2000 platform with 150-base pair paired-end reads. Reads 418 were filtered with SortMeRNA (Kopylova et al 2012) using default options to remove ribosomal sequences. 419 Trimmomatic (Bolger et al 2014) was then used to perform quality trimming. First, we trimmed adapter 420 sequences and leading and trailing bases with low phred scores (<3). We then used the MAXINFO option 421 with target length 36 and strictness 0.7 to trim low-quality sequences from the remaining reads. Reads were 422 next mapped to 11313 transcripts from the P. dominula genome annotation 1.0 (Standage et al. 2016) using 423 STAR (Dobin et al 2013) with default options. All 101 samples produced >85% uniquely mapped reads. 424 Reads were assembled into transcripts using StringTie2 (Kovaka et al 2019) before being passed on as raw 425 counts to downstream analyses. Prior to downstream analysis, transcripts were filtered to remove any gene 426 which did not have >20 counts across all samples in at least one of the experimental groups (queens, control 427 workers, and day three and day twelve post-manipulation workers). Following this filtering, 10734/11313 428 (94.9%) of transcripts remained. 429 430 Support vector machine classification 431 432 Support vector classification was performed in R using the e1071 package (Meyer et al 2017) using the gene 433 expression profiles for all available queens (coded with a value of 1; n=26) and control workers (coded with a 434 value of 0; n=12). Classifiers were assessed via their 3-fold cross-validation error rates; classifiers with lower 435 classification error were considered to be superior. Initially, we tested classifiers using a variety of kernel 436 functions (radial, linear, sigmoid and polynomial) combined with grid searches across a wide range of kernel 437 parameters. A radial kernel with γ=10-6 and cost parameter C=25 was found to produce the lowest error rate 438 of all combinations, so all subsequent classifiers were fit using this kernel and a more focused grid search in 439 a parameter space of 24

453 calculated relative to a baseline fold change of 1.5, i.e. p-values refer to the probability that absolute change 454 between two groups was greater than 50%. Genes were considered differentially expressed between 455 conditions if p<0.05 after false discovery rate correction according to the Benjamini-Hochberg procedure. 456 457 Gene co-expression network analysis 458 459 Weighted gene co-expression network analysis was performed in R using the WGCNA package (Langfelder 460 & Horvath 2008). As WGCNA is particularly sensitive to genes with low expression, data were first subjected 461 to a second round of filtering in which genes that had <10 reads in >90% of samples were removed. 462 Consensus gene modules across all samples were then constructed using a soft-threshold power of 9. 463 Initially, 26 gene modules were identified. Modules whose eigengene correlation was >0.75 were 464 subsequently merged, after which 22 consensus modules remained. Finally, the Pearson correlation of each 465 module with each phenotypic trait within each group (queens, control workers and individuals from queen 466 removal nests) was calculated and subjected to Benjamini-Hochberg FDR correction. 467 468 469 Gene ontology (GO) enrichment analysis 470 471 In order to perform GO enrichment analysis, we first used OrthoFinder (Emms & Kelly 2019) to identify 472 orthologues for each P. dominula gene in D. melanogaster, a model species for which GO annotations are 473 much more complete. GO annotations for each D. melanogaster gene were acquired from BioMart (Smedley 474 et al 2015) and each P. dominula gene was then assigned GO terms permissively, i.e. a given P. dominula 475 gene was assigned a GO term if that term appeared as an annotation to any of its orthologues. GO 476 enrichment analysis was then performed in R via the topGO package (Alexa & Rahnenfuhrer 2009) using 477 TopGO’s weight01 algorithm and Fisher’s exact test to identify GO terms that were significantly 478 overrepresented (p<0.01) in a focal set of genes against a background consisting of all genes that appeared 479 in the relevant analysis.

480 References 481 482 Abeel, T., Helleputte, T., Van de Peer, Y., Dupont, P., & Saeys, Y. Robust biomarker identification for cancer 483 diagnosis with ensemble feature selection methods. Bioinformatics 26, 392-398 (2010). 484 485 Alexa, A., & Rahnenführer, J. Gene set enrichment analysis with topGO. Bioconductor Improv 27 (2009). 486 487 Bolger, A. M., Lohse, M., & Usadel, B. Trimmomatic: a flexible trimmer for Illumina sequence 488 data. Bioinformatics 30, 2114-2120 (2014). 489 490 Casas, L., Saborido-Rey, F., Ryu, T., Michell, C., Ravasi, T., & Irigoien, X. Sex change in clownfish: molecular 491 insights from transcriptome analysis. Sci. Rep. 6, 35461. (2016). 492 493 Chakravarty, P., Cozzi, G., Ozgul, A., & Aminian, K. A novel biomechanical approach for animal behaviour 494 recognition using accelerometers. Methods Ecol. Evol. 10, 802-814 (2019). 495 496 Chandra, V., Fetter-Pruneda, I., Oxley, P. R., Ritger, A. L., McKenzie, S. K., Libbrecht, R., & Kronauer, D. J. 497 Social regulation of insulin signaling and the evolution of eusociality in ants. Science 361, 398-402 (2018). 498 499 Chapman, B. B., Brönmark, C., Nilsson, J. Å., & Hansson, L. A. The ecology and evolution of partial 500 migration. Oikos 120, 1764-1775 (2011). 501 502 Cini, A., Meconcelli, S., & Cervo, R. Ovarian indexes as indicators of reproductive investment and egg-laying 503 activity in social insects: a comparison among methods. Insectes Soc. 60, 393-402 (2013). 504 505 Cullen, D. A., Cease, A. J., Latchininsky, A. V., Ayali, A., Berry, K., Buhl, J., ... & Rogers, S. M. From 506 molecules to management: mechanisms and consequences of locust phase polyphenism. Adv. Insect 507 Physiol 53, 167-285 (2017). 508 509 Dapporto, L., Sledge, F. M., & Turillazzi, S. Dynamics of cuticular chemical profiles of Polistes dominulus 510 workers in orphaned nests (Hymenoptera, Vespidae). J. Insect Physiol. 51, 969-973 (2005). 511 512 Dobin, A., Davis, C. A., Schlesinger, F., Drenkow, J., Zaleski, C., Jha, S., ... & Gingeras, T. R. STAR: ultrafast 513 universal RNA-seq aligner. Bioinformatics 29, 15-21 (2013). 514 515 Elo, A. E. (1978). The rating of chess players, past and present. New York: Arco Pub. 516 Emms, D. M., & Kelly, S. OrthoFinder: phylogenetic orthology inference for comparative genomics. Genome 517 Biol. 20, 1-14 (2019). 518 519 Ferreira, P. G., Patalano, S., Chauhan, R., Ffrench-Constant, R., Gabaldon, T., Guigo, R., & Sumner, S. 520 Transcriptome analyses of primitively eusocial wasps reveal novel insights into the evolution of sociality and 521 the origin of alternative phenotypes. Genome Biol. 14, R20 (2013). 522 523 Fowler‐Finn, K. D., & Rodríguez, R. L. The evolution of experience‐mediated plasticity in mate 524 preferences. J. Evol. Biol. 25, 1855-1863 (2012). 525 526 Geffre, A. C., Liu, R., Manfredini, F., Beani, L., Kathirithamby, J., Grozinger, C. M., & Toth, A. L. 527 Transcriptomics of an extended phenotype: parasite manipulation of wasp social behaviour shifts expression 528 of caste-related genes. Proc. R. Soc. B 284, 20170029 (2017). 529 530 Gomulski, L. M., Dimopoulos, G., Xi, Z., Scolari, F., Gabrieli, P., Siciliano, P., ... & Gasperi, G. Transcriptome 531 profiling of sexual maturation and mating in the Mediterranean fruit fly, Ceratitis capitata. PLoS One 7, 532 (2012). 533

534 Hadfield, J. D., Nutall, A., Osorio, D., & Owens, I. P. F. Testing the phenotypic gambit: phenotypic, genetic 535 and environmental correlations of colour. J. Evol. Biol. 20, 549-557 (2007). 536 537 Heyes, C. Blackboxing: social learning strategies and cultural evolution. Philos. Trans. R. Soc. B 371, 538 20150369 (2016). 539 540 Horiguchi, R., Nozu, R., Hirai, T., Kobayashi, Y., Nagahama, Y., & Nakamura, M. Characterization of gonadal 541 soma‐derived factor expression during sex change in the protogynous wrasse, Halichoeres 542 trimaculatus. Dev. Dynam. 242, 388-399 (2013). 543 544 Huang, S., Cai, N., Pacheco, P. P., Narrandes, S., Wang, Y., & Xu, W. Applications of support vector machine 545 (SVM) learning in cancer genomics. Cancer Genom. Proteom. 15, 41-51 (2018). 546 547 Jandt, J. M., Tibbetts, E. A., & Toth, A. L. Polistes paper wasps: a model genus for the study of social 548 dominance hierarchies. Insectes Soc. 61, 11-27 (2014). 549 550 Kopylova, E., Noé, L., & Touzet, H. SortMeRNA: fast and accurate filtering of ribosomal RNAs in 551 metatranscriptomic data. Bioinformatics 28, 3211-3217 (2012). 552 553 Kovaka, S., Zimin, A. V., Pertea, G. M., Razaghi, R., Salzberg, S. L., & Pertea, M. Transcriptome assembly 554 from long-read RNA-seq alignments with StringTie2. Genome Biol. 20, 1-13 (2019). 555 556 Langfelder, P., & Horvath, S. WGCNA: an R package for weighted correlation network analysis. BMC 557 Bioinformatics 9, 559 (2008). 558 559 Libbrecht, R., Oxley, P. R., & Kronauer, D. J. Clonal raider ant brain transcriptomics identifies candidate 560 molecular mechanisms for reproductive division of labor. BMC Biol. 16, 89 (2018). 561 562 Lo, N., Simpson, S. J., & Sword, G. A. Epigenetics and developmental plasticity in orthopteroid insects. Curr. 563 Opin. Insect. Sci. 25, 25-34 (2018). 564 565 Love, M., Anders, S., & Huber, W. Differential analysis of count data–the DESeq2 package. Genome Biol. 15, 566 10-1186 (2014). 567 568 Manfredini, F., Brown, M. J., & Toth, A. L. Candidate genes for cooperation and aggression in the social wasp 569 Polistes dominula. J. Comp. Physiol. A 204, 449-463 (2018). 570 571 Matson, C. K., Murphy, M. W., Sarver, A. L., Griswold, M. D., Bardwell, V. J., & Zarkower, D. DMRT1 572 prevents female reprogramming in the postnatal mammalian testis. Nature 476, 101-104 (2011). 573 574 Meyer, D., Dimitriadou, E., Hornik, K., Weingessel, A. & Leisch, F. e1071: Misc Functions of the Department 575 of Statistics, Probability Theory Group (Formerly: E1071), TU Wien. R package version 1.6-8. 576 https://CRAN.R-project.org/package=e1071. (2017).

577 Miyano, S. Colony development, worker behavior and male production in orphan colonies of a Japanese 578 paper wasp, Polistes chinensis antennalis Pérez. Popul. Ecol. 68, 347–361 (1986).

579 Miyano, S. Worker reproduction and related behavior in orphan colonies of a Japanese paper wasp, Polistes 580 jadwigae (Hymenoptera, Vespidae). J. Ethol. 9, 135-146 (1991). 581 582 Monnin, T., Cini, A., Lecat, V., Fédérici, P., & Doums, C. No actual conflict over colony inheritance despite 583 high potential conflict in the social wasp Polistes dominulus. Proc. R. Soc. B 276, 1593-1601 (2009). 584

585 Neumann, C., Duboscq, J., Dubuc, C., Ginting, A., Irwan, A. M., Agil, M., ... & Engelhardt, A. Assessing 586 dominance hierarchies: validation and advantages of progressive evaluation with Elo-rating. Anim. 587 Behav. 82, 911-921 (2011). 588 589 Pardi, L. Dominance order in Polistes wasps. Physiol. Zool. 21, 1-13 (1948). 590 591 Patalano, S., Vlasova, A., Wyatt, C., Ewels, P., Camara, F., Ferreira, P. G., ... & Sumner, S. Molecular 592 signatures of plastic phenotypes in two eusocial insect species with simple societies. Proc. Natl. Acad. Sci. 593 U.S.A. 112, 13970-13975 (2015). 594 595 Rast, W., Kimmig, S. E., Giese, L., & Berger, A. Machine learning goes wild: Using data from captive 596 individuals to infer wildlife behaviours. PLoS One 15, e0227317 (2020). 597 598 Réale, D., Dingemanse, N. J., Kazem, A. J., & Wright, J. Evolutionary and ecological approaches to the 599 study of personality. Philos. Trans. R. Soc. B. 365, 3947-3946 (2010). 600 601 Rehan, S. M., Glastad, K. M., Steffen, M. A., Fay, C. R., Hunt, B. G., & Toth, A. L. Conserved genes underlie 602 phenotypic plasticity in an incipiently social bee. Genome Biol. Evol. 10, 2749-2758 (2018). 603 604 Rittschof, C. C., & Robinson, G. E. Genomics: moving behavioural ecology beyond the phenotypic 605 gambit. Anim. Behav. 92, 263-270 (2014). 606 607 Robinson, M. D., McCarthy, D. J., & Smyth, G. K. edgeR: a Bioconductor package for differential expression 608 analysis of digital gene expression data. Bioinformatics 26, 139-140 (2010). 609 610 Rubin, H. The phenotypic gambit: selective pressures and ESS methodology in evolutionary game 611 theory. Biol. Philos. 31, 551-569 (2016). 612 613 Schwander, T., Lo, N., Beekman, M., Oldroyd, B. P., & Keller, L. Nature versus nurture in social insect caste 614 differentiation. Trends Ecol. Evol. 25, 275-282 (2010). 615 616 Segal, N. H., Pavlidis, P., Antonescu, C. R., Maki, R. G., Noble, W. S., DeSantis, D., ... & Cordon-Cardo, C. 617 Classification and subtype prediction of adult soft tissue sarcoma by functional genomics. Am. J. Pathol. 163, 618 691-700 (2003). 619 620 Shell, W. A., & Rehan, S. M. Behavioral and genetic mechanisms of social evolution: insights from incipiently 621 and facultatively social bees. Apidologie 49, 13-30 (2018). 622 623 Simola, D. F., Graham, R. J., Brady, C. M., Enzmann, B. L., Desplan, C., Ray, A., ... & Berger, S. L. 624 Epigenetic (re) programming of caste-specific behavior in the ant Camponotus floridanus. Science 351, 625 aac6633 (2016). 626 627 Smedley, D., Haider, S., Durinck, S., Pandini, L., Provero, P., Allen, J., ... & Bardou, P. The BioMart 628 community portal: an innovative alternative to large, centralized data repositories. Nucleic Acids Res. 43, 629 W589-W598 (2015). 630 631 Standage, D. S., Berens, A. J., Glastad, K. M., Severin, A. J., Brendel, V. P., & Toth, A. L. Genome, 632 transcriptome and methylome sequencing of a primitively eusocial wasp reveal a greatly reduced DNA 633 methylation system in a social insect. Mol. Ecol. 25, 1769-1784 (2016). 634 635 Starks, P. T., Turillazzi, S., & West-Eberhard, M. J. Polistes paper wasps: emergence of a model genus. 636 In Ann. Zool. Fenn. 43, 385-386 (2006). 637

638 Strassmann, J. E., Fortunato, A., Cervo, R., Turillazzi, S., Damon, J. M., & Queller, D. C. The cost of queen 639 loss in the social wasp Polistes dominulus (Hymenoptera: Vespidae). J. Kansas Entomol. Soc. 77, 343-355 640 (2004). 641 642 Tarazona, S., García, F., Ferrer, A., Dopazo, J., & Conesa, A. NOIseq: a RNA-seq differential expression 643 method robust for sequencing depth biases. EMBnet.journal 17, 18-19. (2011). 644 645 Taylor, B. A., Cini, A., Cervo, R., Reuter, M., & Sumner, S. Queen succession conflict in the paper wasp 646 Polistes dominula is mitigated by age-based convention. Behav. Ecol., araa045 (2020). 647 648 Tibbetts, E. A., & Huang, Z. Y. The challenge hypothesis in an insect: juvenile hormone increases during 649 reproductive conflict following queen loss in Polistes wasps. Am. Nat. 176, 123-130 (2010). 650 651 Todd, E. V., Ortega-Recalde, O., Liu, H., Lamm, M. S., Rutherford, K. M., Cross, H., ... & Godwin, J. R. 652 Stress, novel sex genes, and epigenetic reprogramming orchestrate socially controlled sex change. Sci. 653 Adv. 5, eaaw7006 (2019). 654 655 Toth, A. L., Tooker, J. F., Radhakrishnan, S., Minard, R., Henshaw, M. T., & Grozinger, C. M. Shared genes 656 related to aggression, rather than chemical communication, are associated with reproductive dominance in 657 paper wasps (Polistes metricus). BMC Genom. 15, 75 (2014). 658 659 Tsuji, K., & Tsuji, N. Why is dominance hierarchy age-related in social insects? The relative longevity 660 hypothesis. Behav. Ecol. Sociobiol. 58, 517-526 (2005). 661 662 West-Eberhard, M. J. The Social Biology of Polistine Wasps. Misc. Pub. Mus. Zool. Univ. Mich. 140, 1-101 663 (1969). 664 665 Zhou, S., Mackay, T. F., & Anholt, R. R. Transcriptional and epigenetic responses to mating and aging in 666 Drosophila melanogaster. BMC Genom. 15, 927 (2014).