Evolutionary genetic integration of behavioural and endocrine components of the stress response

Abstract

The vertebrate stress response comprises a suite of behavioural and physiological traits that must be functionally integrated to ensure organisms cope adaptively with acute stressors. The expectation that natural selection has favoured functional integration leads to a prediction of genetic integration: genetic variation in the stress response should include covariation between its component behavioural and physiological traits. Despite the implications of such genetic integration for our understanding of human and animal health, as well as evolutionary responses to natural and anthropogenic stressors, formal quantitative genetic tests of this prediction are lacking. Here we demonstrate that Trinidadian guppies (Poecilia reticulata) show genetic variation in a suite of behavioural and physiological components of the acute stress response, and that these are indeed integrated into a single major axis of genetic variation. This axis appears to reflect continuous variation in the magnitude of integrated stress responsiveness, rather than variation in ‘coping style’ (a verbal model that postulates equal levels of stress responsiveness will manifest differently across individuals). The genetic integration we find here could either facilitate or constrain evolutionary responses to selection, depending upon the extent to which the direction of selection aligns with this single major axis of genetic covariation among stress response traits. Such integration also suggests that, while stress-related disease typically arises from physiological components of the stress response, selection on the genetically correlated behavioural responses to stress could offer a viable non-invasive route to the genetic improvement of health and welfare in captive animal populations.

Introduction

Stress responses comprise suites of physiological and behavioural traits that enable individuals to cope with adverse conditions (Romero 2004; Øverli et al. 2007; McEwen & Wingfield 2010; Taborsky et al. 2020). Some individuals are likely better at coping with adverse conditions than others, and understanding the role played by underlying genetic variation could have important implications for managing stress-related disease in captive populations and predicting the evolutionary responses of free-living populations to both natural and anthropogenic stressors (Barton & Iwama 1991; Koolhaas et al. 1999; McEwen & Wingfield 2003; Romero 2004; Koolhaas 2008). For instance, if among-individual differences in stress response traits are a product of genetic variation (Koolhaas et al. 1999, 2007), then they may be a viable target for artificial selection strategies (Mignon-Grasteau et al. 2005). This could be used to reduce stress-related welfare issues in captive populations (e.g. in livestock; Broom & Johnson 1993; von Borell 1995; Möstl & Palme 2002). In free-living populations, variation in stress response traits is expected to cause fitness variation under stressful conditions (Wingfield 2003; Koolhaas 2008). Thus, exposure to stressors could lead to evolutionary changes in the distributions of traits that contribute to a population’s long-term resilience in the face of natural (e.g. predation risk; Clinchy, Sheriff & Zanette 2013) and/or anthropogenic (Tarlow & Blumstein 2007; Busch & Hayward 2009; Angelier & Wingfield 2012) challenges. However, as the evolutionary response of any trait to selection is determined in large part by its genetic variation, our ability to predict – and potentially harness – evolutionary changes in the stress response is currently hampered by limited understanding of the underlying genetics.

Natural selection does not act on single traits in isolation, but rather on multivariate phenotypes (Lande & Arnold 1983; Blows 2007). This is likely to be an important consideration for understanding the evolution of the stress response. For instance, while glucocorticoid (GC) levels are frequently used to measure the stress response (McEwen & Wingfield 2003; Korte et al. 2005), an individual’s first line of defence against acute environmental challenges will typically be behavioural (Moberg 2000). This may include risk avoidance strategies as well as the widely known ‘fight-or-flight’ responses. Subsequent GC release then serves to mediate physiological (and further behavioural) responses (Wingfield et al. 1998; Wingfield & Kitaysky 2002). Natural selection is therefore expected to favour combinations of behavioural and physiological stress response traits that act synergistically to maintain fitness under stressful conditions (Koolhaas et al. 1999; Øverli et al. 2007). Since evolutionary theory predicts that correlational selection will shape the structure of multivariate quantitative genetic variance (as represented by the genetic covariance matrix G; Blows 2007), we should expect genetic – as well as phenotypic – integration of behavioural and physiological stress response traits (McGlothlin & Ketterson 2008; Ketterson, Atwell & McGlothlin 2009; Cox, McGlothlin & Bonier 2016). However, there exists relatively little supporting empirical work for this prediction.

The most compelling evidence for genetic integration of behavioural and physiological stress response traits to date comes from artificial selection experiments on domestic animal populations (e.g., rainbow trout (Pottinger and Carrick, 1999), Japanese quail (Jones, Satterlee & Ryder 1994), house mice, Veenema et al. 2003b). For example, lines of rainbow trout selected for stress-induced plasma cortisol levels (Pottinger and Carrick, 1999) experienced correlated evolutionary changes in behaviour (Øverli, Winberg & Pottinger 2005). In a rare study of a wild-type population (albeit under captive conditions) species, cortisol levels were found to evolve in response to selection on behavioural ‘personality’ in great tits (Parus major; Carere et al. 2003). In these examples, the correlated responses of behavioural and physiological stress response traits to selection are consistent with some degree of genetic integration of these behavioural and physiological traits. However, the extent and ‘structure’ of this integration remains unclear, and some results were inconsistent with a hypothesized simple axis of genetic (co)variation among behavioural and physiological components of the stress response. For example, in the trout study, the ‘low-cortisol response’ selected lines actually showed a higher metabolic stress response under confinement (suggestive of opposing responses to selection by different physiological components of the stress response; Trenzado, Carrick & Pottinger 2003). Surprisingly (given the correlated selection response), in the great tit study cortisol levels were found to be phenotypically unrelated to behavioural ‘boldness’ (the trait selected on) under standardised testing (Thomson et al. 2011).

While selection experiments illustrate that genetic integration of behaviour and physiology can occur, estimation of the genetic variance-covariance matrix (G) through quantitative genetic modelling provides a complementary strategy that also allows investigation of exactly how (multivariate) genetic variation is structured within populations. In the context of the stress response, this should provide insights into both how selection has acted in the past (Ketterson et al. 2009), and whether responses to future selection are likely to be constrained (Blows & Walsh 2009; Walsh & Blows 2009). The former follows from the fact that strong correlational selection should lead to integration of traits in G over the long term, a phenomenon explored most commonly for suites of morphological traits (following, e.g., Cheverud 1982), but that is equally applicable to any aspect of phenotype (see, e.g., Hine, Chenoweth & Blows 2004; Hunt et al. 2007; Oswald, Singer & Robison 2013 for examples pertaining to behavioural evolution and mate choice). The latter follows from the fact that the direction (in multivariate trait space) and magnitude of a response to contemporary selection is limited by the amount of variance in G alignment with the vector of (directional) selection β (Blows & Walsh 2009; Walsh & Blows 2009).

Here we estimate G for behavioural and physiological components of the acute stress response in Trinidadian guppies (Poecilia reticulata). This enables us to determine not only (i) whether these components are genetically integrated into a single major axis of genetic (co)variation, but also (ii) whether the structure and orientation of this axis suggests variation in overall stress responsiveness and/or ‘coping style’ (explained further below; Koolhaas et al. 2010; Boulton et al. 2015). We use fish from a captive colony of guppies derived from wild ancestors sampled from the Aripo River, Trinidad in 2008 and subsequently maintained at high population size (with no deliberate inbreeding or selection). We have validated the use of standardised ‘open field trials’ (OFTs) for testing (acute) behavioural stress responses in this species (Houslay et al. 2018), and demonstrated significant additive genetic (co)variance underpinning variation in risk-taking, exploratory, and ‘flight’ type components of the behavioural stress response using this testing paradigm (White & Wilson 2018; White, Houslay & Wilson 2018). We have also demonstrated, using a non-invasive waterborne hormone sampling method, that individuals differ significantly in their GC (specifically, free circulating cortisol) response to an acute stressor (handling, coupled to short term isolation and confinement; Houslay et al. 2019) and that, on average, this physiological response declines with repeated stressor exposure (consistent with habituation). Nothing is known about the genetic basis of variation in these physiological traits, or about their integration (phenotypically or genetically) with behavioural components of the stress response.

First, we combine OFT results with complementary ‘emergence trials’ (ET) and ‘shoaling trials’ (ST) to characterise among-individual and genetic variation in the behavioural stress response. Second, we characterise the physiological stress response and its rate of habituation by assaying GC levels following first and third exposure to a handling and confinement stressor (see methods). Utilising repeated behavioural and physiological testing of individual fish within a known pedigree structure, we are able to estimate the repeatable (among-individual) component of phenotypic (co)variance in these stress response traits, and then determine the additive genetic contribution to this (G; the genetic variance-covariance matrix for this suite of stress response traits). We predict that individual traits will be heritable and that G will contain strong genetic correlation structure between behavioural and physiological components of the stress response consistent with genetic integration. We also predict that G will be dominated by a single major axis of genetic variation in multivariate trait space, but are more circumspect about how that might look. The ‘stress coping style’ model (Koolhaas et al. 1999) predicts variation in the type of response to stress. Simplistically, as originally proposed this verbal model posits that individuals (or genotypes) perceive equal degrees of stress but differ in how this manifests phenotypically: genotypes at one end of the axis having ‘reactive’ behavioural phenotype (e.g., freezing behaviour) coupled to lower GC levels, while the ‘proactive’ end is characterised by more active ‘fight or flight’ behaviour coupled to higher GC levels. However, previous analyses of this population suggest variation may be more in the stress responsiveness than coping style (Houslay et al. 2018; Prentice et al. 2020; White, Pascall & Wilson 2020). That is to say, some individuals (or genotypes) perceive the trial as a more severe stimulus and exhibit more characteristic stress behaviours (e.g., flight and/or freezing, thigmotaxis) while others show more typical ‘unstressed’ behavioural profiles (e.g., exploration of the arena). In this scenario we predict high GC levels to co-occur with characteristic stress behaviours.

Results

In total we obtained (multivariate) behavioural data from 7,637 trials (3,812 OFTs, 1,548 ETs and 1,039 STs) on 1,518 individual fish. The number of individuals phenotyped (OFTs = 1,487, ETs = 806, STs = 532) and the mean number of observations per fish (OFTs = 2.6, ETs = 1.9, STs = 2.0) varied across the behavioural data types. All fish were contained within a genetic pedigree structure comprising maternal full-sibships nested within paternal half-sibships. This structure was produced via multiple round of breeding work and has a maximum depth of 5 ‘generations’. Some of the OFT data have already been used in studies of the evolutionary genetics of personality (White & Wilson 2018; White et al. 2018), but here we extend that dataset and use it in conjunction with other behavioural and physiological measures for different purposes. We also obtained 1,238 waterborne assays of cortisol levels for 629 fish (almost all from the final generation). The handling and confinement stressor applied for this assay was performed 3 times (at 48h intervals) for all fish tested, but the holding water sample was only processed for GC content at two time points (the first and last confinement, subsequently Cortisol₁ and Cortisol₃). Full details of husbandry, phenotyping and analysis are provided in Materials and Methods.

Genetic variance in behavioural components of the stress response

Behavioural data were extracted from OFTs, ETs and STs using video tracking of fish (as described in White, Kells & Wilson 2016; Houslay et al. 2018). Time to emerge from the shelter (‘emergence time’) was extracted from ETs and natural log (ln) transformed for analysis, while shoaling tendency was calculated from STs as the time spent in the third of the tank closest to a same-sex shoal (which was visible but physically separated) minus the time spent in the third of the tank farthest from the shoal. The OFT, ET and ST testing paradigms are all considered to assay behavioural components of the stress response in the broad sense, as each test starts with the capture and transfer of the focal fish into a novel, brightly lit, arena away from their home tank and familiar tank mates. Three traits were defined from the OFT and measured by videotracking for 4m 30s after an initial 30 second acclimation period on transfer into arena. Track length (distance swum), area covered (as a proportion of the arena floor area), and time in the middle (i.e., in the central area of the open field arena away from the tank walls, which is assumed to be perceived as riskier; e.g., Houslay et al. 2018) were recorded. Note low values of time in the middle imply thigmotaxis (i.e., tendency to avoid exposure to potential threats by hugging walls). All three of these OFT traits are repeatable and heritable in this population (White et al. 2016, 2018; Houslay et al. 2018; White & Wilson 2018). However, the absence of a strong positive correlation between track length and area covered (Fig 1A) is notable and potentially biologically informative; if fish moved randomly with respect to direction in the arena then area covered would increase monotonically (to an asymptote at 100%) with track length. A possible explanation is that a long track length can sometimes arise from a (putatively) less stressed fish exploring the arena (fish 1 in Fig 1B) and sometimes a (putatively) more stressed fish exhibiting a typical ‘flight’ response (Fish 4 in Fig 1B). These can be discriminated based on whether, in a given trial, high track length is associated with high area covered and time in the middle (exploration) or the converse (flight response).

Figure 1:

The lack of a strong positive relationship between observed track length and area covered (panel A), is initially puzzling given expected autocorrelation and that both are used as positive indicators of exploratory (or ‘bold’) behaviour. Inset examples of OFT tracks from 4 individuals (panelB) shed light on this. Fish 1 and 2 appear to be exploring the tank, while 3 and 4 are engaging in stereotypical ‘flight’ behaviour characterised by strong thigmotaxis (remaining close to tank walls) and/or rapid movement along tank walls. As a consequence, individuals 2 and 3 have similar area covered during the OFT, but very different track lengths. We simulated random movements to define an expected null relationship between area covered and track length (panel C; dashed red line shows the fourth order polynomial model fit; see Appendix A). The polynomial regression was then used to predict the expected area covered under random movement for each trial’s observed track length, and the ‘relative area covered’ was calculated as the observation minus this prediction. Panel D shows the resultant relative area covered plotted against track length for all trials (dashed red line at relative area covered = 0, shows where individuals of any track length are expected to lie if they move randomly with respect to direction). From this it is apparent that fish 1 and 2 have high relative area covered, while 3 and 4 do not.

To quantitatively discriminate between exploratory behaviour and flight responses we derived a new trait, ‘relative area covered’. We used a simple simulation procedure (see Methods) to predict expected area covered for any given track length under a null ‘random swim’ within the arena (Fig 1C). Relative area covered is then calculated as observed area covered – expected area covered given the track length (Fig 1D) and will be high for fish engaging in exploration, and low for an obvious ‘flight’ response manifest as rapid swimming around the tank walls.

Pedigree-based ‘animal models’ (Wilson et al. 2009) were used to test for and estimate additive genetic variation in each of the five behavioural traits while controlling statistically for social housing group and non-genetic sources of among-individual variance (as well as several fixed effects; see methods for full details). These confirmed the presence of significant additive genetic variation for the relative area covered trait, as well as for track length and time in the middle (as expected from previous findings; White & Wilson 2018; White et al. 2018) as well as emergence time (Table 1). With the exception of shoaling tendency, heritabilities (conditional on fixed effects; see methods) are low to moderate (range of 8-17%; Table 1) but within the expected range for behaviours (Stirling, Réale & Roff 2002). We detected no additive genetic variance for shoaling tendency (Table 1), despite there being repeatable differences among individuals; R = 0.19 ± 0.04; χ²_0,1 = 20.01, P < 0.001).

Table 1:

Estimated variance components, along with adjusted heritability, for each trait as estimated in a univariate model (± standard error). Chi-square test statistics and p-values are provided for the pedigree term, testing for the presence of significant additive genetic variance (V_a).

Genetic variance in physiological components of the stress response

Using a series of nested bivariate animal models, we tested for additive genetic variance in cortisol levels (ln-transformed) following stressor exposure (handling and confinement) and for genotype-by-environment interaction (GxE). In this context, the environment (E) is the trial number in each fish’s stress trial series (i.e. cortisol level following stressor trial 1 or 3). Any GxE present can therefore be interpreted as genetic variance for habituation to the stressor, given that the average cortisol level was lower following exposure to the third stressor than the first (ln transformed ng/hr, mean ± SE; Cortisol₁ = 8.50 ± 0.05, Cortisol₃ = 8.05 ± 0.06, Wald F_1,12.9 = 120.5, P < 0.001; see methods for explanation of units). We first modelled Cortisol₁ and Cortisol₃ as distinct response variables in a bivariate framework assuming no GxE (such that we constrain V_A-Cortisol1 = V_A-Cortisol3 and the cross context additive genetic correlation r_{A-Cortisol1,Cortisol3} = 1). This model revealed a significant additive genetic component to variation among individuals in their cortisol levels following stressor exposure (χ²_0,1 = 6.58, P = 0.005).

Expanding the model to allow GxE (by estimating separate genetic variances for Cortisol₁ and Cortisol₃ and allowing the cross-context genetic correlation to deviate from +1) provides a significantly better fit to the data (χ²₂ = 9.65, P = 0.008), meaning GxE is present. This can be viewed as genetic variance for habituation to repeated stressor exposure, or as a change in genetic variance for cortisol from the first to the third sampling (Figure 2). These are two perspectives of the same phenomenon; a reduction in additive genetic variance between the first stressor (V_A-Cortisol1 = 0.076 ± 0.028) and the third (V_A-Cortisol3 = 0.047 ± 0.029) arises because genotypes with higher than average levels for Cortisol₁ habituate more rapidly (i.e. have more negative reaction norm slopes). Note however that the rank order of the genotypes does not appreciably change across the two contexts (i.e. genetic reaction norms show little crossing; Figure 2), so there is a strong positive cross-context genetic correlation (r_{A-Cortisol1,Cortisol3} ± SE = 0.74 ± 0.25).

Figure 2:

Guppies habituate to the waterborne sampling procedure, as shown by a decline in average log-transformed cortisol level (ng/hr) following stressor exposure between first and third exposures. Black circles and associated bars denote predicted population means (± standard error) from mixed model analysis. Grey lines depict the predicted genetic reaction norms across repeated stressor exposure for each individual. Weak, but statistically significant GxE is manifest as variance in the genetic reaction norm slopes (i.e. lack of parallelism) and results in a slight reduction of genetic variance for cortisol at the third exposure relative to the first.

In this model we also find that variance in cortisol explained by housing group effects is similar across contexts (V_{Group-Cortisol1} = 0.034 ± 0.013, V_{Group-Cortisol3}= 0.045 ± 0.016), but that residual (unexplained) variance is greater after the third stressor exposure (V_R-Cortisol1 = 0.166 ± 0.021, V_R-Cortisol3 = 0.229 ± 0.025). In combination, the changes in both additive genetic and residual variance between the two contexts lead to appreciably higher heritability for cortisol levels following the first stressor exposure relative to the third (h²_Cortisol1= 0.275 ± 0.093, h²_Cortisol3= 0.146 ± 0.088).

Testing for genetic integration and identifying the major axis of genetic (co)variance

There is strong evidence for phenotypic integration of Cortisol with behaviour at the among-individual levels (see Table S5). To test for and characterise the hypothesised genetic integration between behavioural and physiological components of the stress response, we built a multivariate animal model to estimate G. We excluded shoaling tendency given the absence of detectable genetic variance in the univariate model. We also elected to treat cortisol as a single trait (allowing for a fixed effect of stressor exposure number (1 vs 3) on the mean). Although the above analysis demonstrates GxE for cortisol, the strong positive cross-context genetic correlation justifies collapsing Cortisol₁ and Cortisol₃ into a single trait to maximise statistical power to detect any genetic covariance with behaviour.

Our final model contained 5 response traits: relative area covered, time in the middle and track length, (ln-transformed) emergence time, and (ln-transformed) Cortisol (now treated as two repeats of a single trait). We standardised all (transformed) traits to standard deviation units, to assist multivariate model fitting and to prevent eigenvectors of G (see below) being dominated by traits with higher variance in observed units. To simplify interpretation of G we also multiplied emergence time by −1 after transformation. Thus high values denote rapid emergence from the shelter.

The resultant estimate of G (Table 2) contains significant additive genetic covariance structure overall (Likelihood Ratio Test of the full model vs. a reduced model G that contains variances but not covariances:χ²₁₀ = 36.79, P < 0.001). Strong genetic covariance/correlation estimates, both positive and negative, were found for a number of trait pairs and were deemed statistically significant (based on the bootstrapped 95% confidence intervals not crossing zero). We find strong genetic covariance among all 3 OFT traits: track length shows significant negative genetic covariance with both relative area covered and time in the middle, and relative area covered and time in the middle show significant positive genetic covariance with one another. Track length also has a significant positive genetic covariance with –ln(emergence time). Ln-transformed Cortisol shows significant negative genetic covariance with time in the middle (Figure 3). Ln-transformed Cortisol also covaries negatively with both relative area covered and – ln(emergence time), and positively with track length, although these covariances were not significantly different from zero (based on 95% confidence interval).

Figure 3:

The additive genetic relationship between log-transformed cortisol (y-axis) and four behaviours (a, relative area covered; b, time in the middle; c, track length; d, -log emergence time). Points show (predicted) bivariate genetic deviations from the population means, plotted for those individuals in the pedigree with cortisol data. In each case the black ellipse depicts the ‘shape’ of the relationship as given by the point estimate of G. Specifically it encompasses the area expected to contain 95% of the bivariate genetic distribution for the population. Grey ellipses denote the corresponding areas defined from 5000 bootstrapped replicates of G, and so highlight the uncertainty around these bivariate distributions.

Table 2:

Additive genetic covariance-correlation matrix (G) from the full multivariate animal model. Genetic variances provided on the shaded diagonal, with genetic covariances below and genetic correlations above. 95% confidence intervals in parentheses are estimated from 5000 bootstrapped replicates. Where the confidence intervals for any estimate do not cross zero the estimate is considered statistically significant (at the0.05 alpha level) and are shown in bold.

Eigen decomposition of G provides a more holistic view of the genetic covariance structure and the level of integration among traits. Here the major axis (first principal component, PC1, with 95% confidence intervals from 5000 bootstrap replicates) explains 57.8% (47.3%, 79.6%) of the genetic variance in multivariate phenotype (PC2 = 25.7% [14.3, 37.6]; PC3 = 9.3% [4.7, 16.5]; PC4 = 5.6% [0, 9.2]; PC5 = 1.6% [0, 3.6]). All traits except emergence time load significantly on this axis (Figure 4). Relative area covered and time in the middle load in one direction, while track length and Cortisol load in the other direction. This structure is suggestive of a single major axis of genetic variation in integrated stress response, where genotypes at one end of this axis can be considered to have ‘weaker’ behavioural stress responses to the OFT assay (i.e., swim shorter distances, spend more time in the central area of the tank, and exhibit exploratory swimming patterns that cover greater areas relative to their distance swum) and ‘weaker’ physiological responses to stress (i.e., produce lower cortisol levels in response to the stressor). Meanwhile, genotypes at the other end of this axis can be considered to have ‘stronger’ behavioural responses to stress (i.e., swimming further, while spending more time close to the tank edges, and covering less area relative to their distance travelled in OFTs) and ‘stronger’ physiological responses to stress (i.e., produce higher cortisol levels in response to the stressor).

Figure 4:

Trait loadings from the first eigen vector (principal component) of G. This axis explains 57.8% of the genetic (co)variation found in the focal behavioural and physiological components of the stress response in our guppy population. Points show trait loadings from the first eigen vector of our estimate of G, with bars representing 95% confidence intervals on each loading (calculated from 5000 bootstrapped replicates of the model).

Discussion

In this study we sought to determine whether – and to what extent – there exists genetic variation for, and integration between, behavioural and physiological (endocrine) components of the stress response. Our results provide three main novel insights. First, we find that genetic variation does underpin individual differences in both behavioural and physiological components of the stress response. Second, we find genetic covariance structure among these behavioural and physiological traits, indicating that they are indeed genetically integrated. Thirdly, having identified the structure of the major axis of G we suggest that it is more readily interpreted as an axis of genetic variation in stress responsiveness than in stress coping style (although we acknowledge the distinction may be somewhat subjective). Overall, by estimating the genetic covariance structure among traits we find the first quantitative genetic support to date for the hypothesis of evolutionary integration between behavioural and endocrine components of the stress response.

We find heritable (co)variation in and among behaviours assayed in the open field trial (OFT), including the derived trait relative area covered. The latter trait, derived by considering an appropriate biological null model of the relationship between track length and (absolute) area covered serves as a useful proxy for exploratory behaviour. Here we demonstrate an axis of repeatable and heritable variation that spans from less active but more exploratory swimming patterns (lower track length but higher relative area covered and time in the middle) through to a ‘flight’ type response, characterised by higher activity coupled to thigmotaxis and low exploration (higher track length, lower relative area covered and time in the middle). Although the distinction is both subjective and arguable, we interpret this axis as more consistent with variation in the magnitude of stress responsiveness than in coping style. This is because, while low track lengths could arise from stressed fish exhibiting ‘reactive’ freezing behaviour, we would not necessarily expect it to be associated with high time in the middle (i.e. reduced thigmotaxis). Given the wide use of OFTs in biomedical research (e.g., Rex et al. 1998; von Horsten, Karl & Pabst 2003) as well as in animal behaviour, our phenotyping approach may have broad applicability for discriminating between exploration and stress/anxiety-related behaviours. At the slight risk of introducing semantic confusion, we also note that the OFT paradigm is widely applied to studies of ‘shy-bold’ type personality variation in fishes (Toms, Echevarria & Jouandot 2010) and other vertebrates (Carter et al. 2013; Perals et al. 2017). The extent to which behavioural differences deemed characteristic of a ‘shy—bold’ personality axis (commonly, if not universally defined as repeatable variation in response to perceived risk; Wilson et al. 1994) should be viewed as equivalent to variance in behavioural stress responsiveness, or coping style, is a matter of debate (see Boulton et al. 2015). We view these as overlapping – if not necessarily identical – concepts, and suggest it will generally be prudent to empirically assess the structure of OFT variation for any system rather than assume a priori that it will match a preferred verbal model (see also White et al. 2016; Houslay et al. 2018; White, Pascall & Wilson 2020).

We also find significant heritable variation in emergence time, although this is not tightly integrated with OFT traits in G in the manner we had expected. Faster emergence in typically interpreted as reflecting a lack of fear of the open arena (i.e., greater ‘boldness’; see Burns 2008). In our study the qualitative pattern in G runs counter to this, if we consider that boldness and behavioural stress responsiveness are broadly analogous. Thus genotypes predisposed to shorter emergence time are associated with greater track length and, albeit non-significantly, lower relative area covered and time in the middle during OFT trials. That is, shorter emergence time is associated with the putatively more stressed ‘flight’ end of the behavioural axis revealed in the OFT. Our interpretation of this result is that at least some genotypes (and individuals) likely perceive the shelter area as less safe than the open arena. This possibility was also suggested by an earlier finding that some individual guppies decrease (rather than increase) shelter use following simulated predation events (Houslay et al. 2018). If general to other systems, unexpected results such as these may have important consequences for interpretation of personality tests, and correlations (or lack thereof) between behaviours assayed across test paradigms (Carter et al. 2013).

Tendency to shoal varies among individuals but is not detectably heritable. Though not generally considered a stress-response trait per se, shoaling is an anti-predator behaviour in guppies (Herbert-Read et al. 2017). We had therefore predicted that heightened perception of risk in the open field might also be associated with increased shoaling tendency. This was not the case at the among-individual level (Table S5), while the absence of detectable genetic variance meant that we could not test this prediction in G.

We find strong evidence of significant additive genetic variance in a key physiological component of the stress response: waterborne cortisol concentrations following exposure of the fish to a handling stressor. Our findings suggest that previously detected differences among individuals in cortisol response to a stressor (Houslay et al. 2019) are primarily attributable to genetic effects, with the estimated heritability (h² = 0.26) being almost 75% of the individual–level repeatability (R = 0.35) for ln-transformed Cortisol. In addition, by adopting a reaction norm approach to modelling stress physiology, as recently advocated by ourselves (Houslay et al. 2019) and others (e.g., Fürtbauer et al. 2015; Hau & Goymann 2015; Taff & Vitousek 2016), we detect GxE reflecting genetic differences in the extent of habituation to the stressor over repeated exposures. This result is potentially important since poor habituation of the hypothalamic-pituitary-adrenal/interrenal (HPA/I) response to repeated or ongoing stressors can lead to well documented health problems in human and animal populations (Segerstrom & Miller 2004; Koolhaas 2008; Romero, Dickens & Cyr 2009; Mason 2010). Our detection of heritable variation in the degree of habituation to stressors raises the possibility of developing targeted selection strategies to improve welfare in captive populations (e.g., Frankham et al. 1986; Muir & Craig 1995; Oltenacu & Algers 2009).

Our findings also highlight that there is greater additive genetic variance (and heritability) for cortisol levels following the first exposure to the stressor than following the third. This pattern, which occurs because genotypes that produce the highest cortisol response at first exposure also show the most marked habituation, is consistent with the idea of cryptic genetic variance (Paaby & Rockman 2014) being ‘released’ by exposure to novel, and so potentially stressful, environments (Ledón-Rettig, Pfennig & Crespi 2010; Ledón-Rettig et al. 2014; Berger et al. 2011). All else being equal, it also means that selection on cortisol levels following stressor exposure should induce a stronger evolutionary response in naïve relative to habituated fish. However, the strong positive cross-environment correlation means that the ranking of genotypes with regard to their cortisol responses is consistent across repeated stressor exposures. Thus selection on the (average) GC response would result in a correlated evolutionary response of habituation rate, and vice versa.

Considering all traits together, G shows evidence of genetic integration between behavioural and endocrine components of the stress response. As noted with respect to OFT behaviours, we consider that the major axis of G is best interpreted as genetic variance in the magnitude of stress responsiveness. Accordingly, genotypes tending to show (putatively) more stressed ‘flight’ type behaviour in the OFT (i.e. thigmotaxis, high track length, low relative area covered) also produce higher levels of cortisol following the handling and confinement stressor. The fact that the only significant bivariate correlation including cortisol is with time in the middle (negative correlation) suggests that thigmotaxis is a particularly strong indicator of high stress responsiveness. We note that our interpretation that G is dominated by variation in the magnitude of stress responsiveness does not mean the ‘coping style’ model has no merit. Indeed, a subsequent ‘two-tier’ iteration of the coping style model proposed that variation in the magnitude of the stress response (termed ‘stress reactivity’) could be viewed as a second axis, distinct from variation in the ‘type’ (or style) of response (Koolhaas et al. 2010).

Here G certainly contains substantial variation not explained by its leading axis and this may point towards interesting avenues for further study. For instance, while relative area and tracklength are negatively correlated and both load significantly on the major axis of G, there is also much variation in the former at low to moderate track lengths (illustrated by e.g., fish 1 and 4 in Figure 1). It seems plausible, if speculative, that this could be caused by a tendency of some individuals to show initial immobility (i.e. freezing) in response to the stressor but then recover quickly and begin explorative swimming. Others may show an initial flight response before relatively rapid recovery to explorative swimming. Such differences, if present, would not be detectable by simple principal components analysis of covariance structures based on whole trial data (as here and, for example, Van Reenen et al. 2005). Rather this would require more detailed modelling of within-trial behavioural dynamics.

The genetic integration of behaviour and physiology detected here is consistent with the idea that correlational selection in the past has led to the coevolution of these stress response components. Covariance structure in G will modify, and potentially constrain, evolutionary responses to selection - whether natural or artificial. Here we have no direct knowledge of how contemporary selection is acting in the wild. Nor do we know whether it might be changing as a consequence of anthropogenic stressors. Thus we cannot comment directly on how G will shape future evolution of the guppy stress response beyond noting that selection on behaviour will cause correlated evolution of endocrine physiology (and vice versa). Nonetheless, while it seems reasonable to expect that current integration of stress response in natural populations should be broadly adaptive, this seems less likely in captive populations (at least for species without a long history of domestication and opportunity for adaptation to artificial environments). We know that prolonged, chronic activation of stress response pathways (notably the HPA(I) axis) frequently disrupts health and survival in captive animals (Huether 1996; Boonstra 2013). It may be that more stress-responsive genotypes are disadvantaged in novel artificial conditions (e.g., if acute stress responsiveness positively predicts susceptibility to chronic stress). However, even if true this would not imply high (acute) stress-responsiveness was also disadvantageous in the wild. Since natural selection should purge alleles that are universally detrimental, it seems more plausible that genetic variation along the major axis described here is maintained by some form of selective trade-off (as widely hypothesised for maintenance of personality variation; e.g., Stamps 2007; Wolf et al. 2007; Réale et al. 2010). For instance, genotypes susceptible to harm under chronic stressor exposure will likely persist in populations if they also confer advantages under an acute stress challenge. In natural populations not only is exposure to acute stressors more common than to chronic stressors, but also selection through chronic stress exposure may be conditional on (and subsequent to) surviving acute challenges (such as predator attacks).

Conclusions and future directions

Here we find evidence for genetic variation in – and integration of – behavioural and physiological (endocrine) components of the stress response. Overall we consider the structure of G to be more consistent with a continuous axis of variation in acute stress responsiveness than with the widely invoked ‘reactive – proactive’ model of variation in stress coping style (Koolhaas et al. 2007). This interpretation rests largely on the structure of behavioural variation revealed by the OFT, which is dominated by an axis running from genotypes that are more stress responsive (rapid ‘flight’ behaviour coupled to thigmotaxis) to those that show more exploratory behaviour expected from a (putatively) unstressed fish. Endocrine traits align with this axis: genotypes exhibiting ‘flight’ behaviour show higher cortisol levels (and exhibit faster habituation of GC physiology) when subject to repeated handling and confinement stressors.

Our results suggest that correlational selection in the past has likely shaped the multivariate stress response, and that continued evolution of stress-related behaviour will have consequences for glucocorticoid physiology and vice versa. Determining contemporary selection on the stress response, and testing the possibility that genetic variation is maintained by fitness trade-offs, is thus an obvious – if empirically challenging – next step to understanding the functional importance of genetic variation in wild populations. In a more applied context, integration of behavioural and endocrine stress-response components at the genetic level has potential utility for genetic improvement of managed populations. Specifically, it may be possible to identify non-invasive, high throughput, behavioural biomarkers and target them in selection schemes to reduce chronic activation of the HPA/I axis and its attendant deleterious effects.

Materials and methods

Husbandry and breeding

We used fish taken from our captive population housed at the University of Exeter’s Penryn campus, which is descended from wild fish collected in 2008 from the lower Aripo River in Trinidad. This population has been maintained at a population size of several thousand, and has undergone no deliberate selection or inbreeding. All fish are fed to satiation twice daily (0800 – 1000h and again at 1600 – 1800h) using commercial flake food and live Artemia nauplii. Water temperature is maintained at 23-24°C in well-aerated closed system tank stacks that undergo 25% water changes each week and with weekly tests for ammonia, nitrate and nitrite levels. Lighting is kept at a 12:12 light/dark cycle. The experiment described in this study was carried out in accordance with the UK Animals (Scientific Procedures) Act 1986 under licence from the Home Office (UK), and with local ethical approval from the University of Exeter.

To create our pedigreed sub-population, female fish were sampled haphazardly from the stock tanks and kept in female-only groups for 3 months. Isolation from male contact minimised the chance of females carrying viable sperm from previous matings. For the first generation of offspring, we used a group breeding design (as detailed in White and Wilson, 2018); briefly, females were tagged under anaesthetic (buffered MS222 solution) using visible implant elastomer (VIE) to allow individual identification. We then assigned groups of 4 females to 1 male in 15L breeding tanks (18.5cm × 37cm × 22cm), and inspected females daily for high gravidity (swollen abdomens and enlarged ‘gravid spots’). Heavily gravid females were then isolated in 2.8L brood tanks to give birth (and were returned to the breeding tanks either after producing a brood or two weeks of isolation). Any offspring produced in the breeding tanks were excluded from the experiment as maternal identity could not be positively identified. For the following generations, after 3 months of isolation from males we moved females into individual 2.8L tanks, with 1 male then circulated among 3 females. Males were moved between females every 5-8 days. In this way, females did not have to be moved to brood tanks, and any offspring could be assigned to mothers definitively. In this setup, offspring were moved to a separate brood tank on the day of birth. Note that as the gestation period for guppies is approximately 1 month, any brood produced by a female less than one month after exposure to their designated male was recorded in the pedigree as having unknown paternity.

Within 24h of a female producing a brood we recorded her weight (g) and brood size. We kept juvenile fish in full-sib family groups in 2.8L tanks before moving them to 15L ‘growth’ tanks at an average age of 56 days. At an average age of 133 days (range 59-268) we tagged individuals and placed them into mixed family groups of 16-20 adults (with an even mix of males and females), kept in 15L tanks. Note that variation in tagging age arose largely because groups were necessarily established sequentially as sufficient individuals from multiple families reached a large enough size that we deemed the procedure to be safe. Each adult group comprised a mix of fish from different families, reducing the potential for common environment effects to upwardly bias our genetic parameter estimation.

Overview of behavioural phenotyping

Behavioural phenotyping commenced at least one week after tagging. In all trials, we filmed movement behaviour of individual fish using a Sunkwang video camera equipped with a 6-60mm manual focus lens suspended over the tank. We used the tracking software Viewer II (BiObserve) to extract behavioural data from each recording (detailed below). The tank was lit from below using a light box, and screened with a cardboard casing to prevent external visual disturbance. After each behavioural trial, the individual tested was weighed and then moved to a temporary ‘holding tank’. Once a full group (as described above) had been tested, all were moved from the holding tank back to their home tank. We replaced the water in the testing and holding tanks between groups to reduce the build-up of hormones or other chemicals. The first offspring generation experienced 4 repeat open field trials (OFTs) over a 2-week period, with at least 48h between trials. Subsequent generations experienced 4 repeat behavioural trials, alternating 2 OFTs with 2 emergence trials (ETs). For the final 2 generations, we extended the OFTs by including a shoaling trial (ST) at the end of each OFT.

Open field trials (OFT) followed the methodology described by White et al (2016). Briefly, we assessed individual behaviour in a 20cm × 30cm tank, filled to a depth of 5cm with room-temperature water from the main supply. We caught fish individually from their home tank, examined them quickly for identification tags, then placed them immediately into the centre of the OFT tank. After allowing 30s for acclimation, we filmed behaviour for 4m30s. Behaviours characterised from the tracking software were track length (the total distance the fish moved during the trial; cm), area covered (the percentage of 1cm × 1cm grid squares through which the fish moved during the trial; %), and time in middle (time spent in a rectangular inner zone which was defined as being the same size as an outer area; seconds).

Shoaling trials (ST) were appended to a subset of OFTs, by positioning a small tank containing 10 stock fish (of same sex as the test subject) next to one end of the OFT tank but with visual cues blocked by a cardboard divider. At the end of the normal OFT, we removed this divider slowly, allowing the focal animal to have visual contact with the shoal. We began recording the shoaling trial 30s after removing the divider in order to limit any artefacts of slight disturbance. (Note that we used a further cardboard casing around the shoaling tank to avoid any additional external visual stimulus). We then recorded behaviour of the test fish for an additional 3 minutes. We characterised shoaling tendency via the tracking software by subdividing the tank area into 3 equal-sized rectangular areas: one next to the tank holding the group of same-sex fish, one farthest from this group, and the central area. We then calculated shoaling tendency as the time spent in the 1/3 area closest to the same-sex group after subtracting the time spent in the 1/3 area farthest away. The decision to use a single-sex shoal aimed to reduce any effects of (potential) mate preference and/or avoidance, but also this necessitated replicate arena setups allowing male and female individuals from each group to be tested in the OFT/ST concurrently. We randomised which tank was used for each sex in each group and recorded this information.

Emergence trials (ET) followed the methodology described by White et al. (2016). Briefly, we tested individuals in a 20cm × 40cm tank, filled to a depth of 8cm with room-temperature water from the main supply. A 10cm section of the tank length was walled off creating a shelter area (20cm x10cm), the walls and floor of which were painted black. The focal fish was placed into the shelter area and allowed to acclimate for 30s, at which point we opened a sliding door to allow access to the rest of the tank, which was brightly lit from below and otherwise bare. Time to emerge (in seconds) was recorded by the tracking software automatically as the fish exited the shelter area and emerged into the open tank section. Trials were ended either at emergence or at 15 min if the fish had not emerged by that point; in the case of non-emergence, fish were given the maximum value (i.e., 900s).

Derivation of ‘relative area’ from OFT trials

The ‘area covered’ variable assayed in the OFT is calculated in BiObserve by dividing the arena (i.e., the total area of the tank as viewed from the camera) into 1cm × 1cm grid squares. The path taken by the fish during observation is then used to determine what proportion of these grid squares the fish entered. However, we sought to derive a measure of ‘relative area’ that describes whether a fish covers a large, or small area relative to its observed track length.

To do this we simulated ‘random swims’ within the arena across the observed range of track lengths. We first selected 40 OFT results at random from our total data set and extracted the coordinates of the fish in each frame from the raw tracking file, creating a set of x and y movements and their associated distances. As original coordinates were recorded in pixels we used the calibration of the software to convert to cm units. We then use a ‘random walk’ algorithm to select a movement (i.e., step size and direction) from this observed distribution at random, and calculate the new coordinates. If the movement keeps the ‘fish’ within the bounds of the ‘tank’ (i.e., defined as a 20cm × 30cm arena), the movement is accepted and coordinates added to a movement matrix; if not, a new movement is drawn from the distribution. If the movement is greater than 1cm in distance, we break the movement into a number of smaller parts to be added to the matrix (such that we capture the coordinates of grid squares through which the 'fish' moved along the way). Once the total distance of the random walk reached or exceeded the track length set as the simulation objective, the path is terminated and the area covered is calculated by counting the number of unique grid squares in the matrix of coordinates and dividing by the total number possible.

After simulating random walks across 500 values of track length (using a vector of 100 values evenly spaced across the range of true data, repeated 5 times), we modelled (simulated) area covered as a fourth order polynomial function of track length. Using this regression model (which explained 97.8% of the variance in simulated data), we calculated the relative area for each actual OFT trial as the observed area covered minus the area covered under a random swim, as predicted from our regression model and the observed track length.

Waterborne hormone sampling

On completion of behavioural data collection, individuals entering the endocrine testing program were left undisturbed for a minimum of two weeks. Waterborne hormone sampling was then conducted over a 5-day period that included three handling and confinement stressor exposures with 48h between each. We followed the method described by Houslay et al (2019) to obtain repeated non-invasive GC measures of individuals using holding water samples from the first and third confinements. Note that only two samples per fish were analysed because the financial and time costs of doing three was deemed prohibitive. We nonetheless applied the stressor stimulus three times as our prior study showed this was sufficient to produce a strong habituation response, i.e., a significant decrease in water-borne cortisol over the three sampling periods (Houslay et al. 2019).

We collected samples between 1200 – 1400h to control for diel fluctuations in GC levels. For each sample, we netted an entire group from their home tank quickly using a large net, transferring them to 2 holding tanks (containing water from the home tank supply) for moving to an adjacent quiet room (performed within 20s of the net first hitting the water). We then transferred fish to individual Pyrex beakers containing 300ml of clean water from the main supply (which serves the main housing units), which has been warmed to the appropriate temperature (mean = 24.1°C, range 23-24.9°C). Beakers were placed within cardboard ‘chambers’ to prevent fish from seeing each other or experiencing outside disturbance. One fish was transferred every 30s, alternating across holding tanks, such that all fish were in their beakers within 10min of the initial netting. After 60 mins in the beaker, each fish was removed by pouring its sample through a clean net into a second beaker, with the fish then quickly checked to confirm ID and returned to the holding tank until the entire group could be returned to its home tank.

We immediately filtered each water sample using Grade 1 filter paper (Whatman), then passed them slowly through solid phase C18 extraction columns (Sep-Pak C18, 3cc, Waters) via high-purity tubing (Tygon 2474, Saint Gobain) under vacuum pressure (Earley et al. 2006). Columns were primed beforehand with 2 × 2ml HPLC-grade methanol followed by 2 × 2ml distilled water, and were washed afterwards with a further 2 × 2ml distilled water to purge salts. We then covered both ends of each column with film (Parafilm M, Bemis) and stored them at −20C for future analysis. We washed all beakers, tubes and funnels with 99% ethanol and rinsed them with distilled water prior to each sampling procedure. The remainder of the endocrine assay procedure involved elution, resuspension, separation and quantification of free cortisol by enzyme immunoassay (EIA) using Cayman Chemicals, Inc EIA kits. Detailed methods are exactly as described by Houslay et al (2019) and so not repeated here (note that here we assayed the free fraction of cortisol only). To validate the cortisol kits, we examined whether the kit standard curve was parallel to a serial dilution curve derived from pooled guppy water-borne hormone extract. 20μl was taken from each of the male samples and pooled; 20μl was taken from each of the female samples and combined into a separate pool. 400μl of the pools was serially diluted from 1:1 to 1:128 and these samples were assayed alongside the kit standard curve on two occasions (June and December 2017, marking the start and finish of sample processing). All dilution curves were parallel to the standard curve (slope comparison test, Zar 1996, p.355; June, male: t₁₂ = 0.029, P = 0.97; June, female: : t₁₂ = 0.343, P = 0.74; December, male: : t₁₂ = 0.119, P = 0.91; December, female: : t₁₂ = 0.224, P = 0.83). The serial dilution curves also identified 1:32 as an appropriate dilution to ensure that all samples fell on the linear phase of the standard curve. A total of 37 96-well plates were used and the pooled sample was included at the beginning and end of each plate. Intra-assay coefficients of variation ranged from 0.12-19.83% with a median of 3.08%; the inter-assay coefficient of variation was 19.22%. Cortisol is presented and modelled in (ln-transformed) units of ng/hr to reflect the 1 hour sampling duration.

Statistical methods

All data handling and analysis was performed in R version 3.4.1 (R Core Team 2017). We used the ‘tidyverse’ set of packages for data handling and visualisation (Wickham 2017), and ASreml-R 3.0 (Butler 2009) for fitting linear mixed effects models (as described in full below). We also used ‘nadiv’ for pedigree preparation and to estimate (approximate) standard errors on linear functions of variance components as estimated from the mixed models (Wolak 2012). All models fitted assumed (multivariate) Gaussian error structures, and we visually assessed residuals to verify this was reasonable (after data transformation in some cases). To test for significance of among individual and/or genetic (co)variance components, we fitted nested models with different random effects structures and compared them using likelihood ratio tests (LRTs). We calculated χ²_nDF as twice the difference in model log likelihoods, with the number of degrees of freedom (n) equivalent to the number of additional parameters in the more complex model. When testing a single random effect (variance component), we assumed the difference to be asymptotically distributed as an equal mix of χ²₀ and χ²₁ (denoted χ²_0,1; Self and Liang, 1987; Visscher, 2006).

For each OFT and ST behaviour in turn (relative area, time in middle, track length, shoaling tendency, and emergence time), we used the random effects specification to partition phenotypic variation (V_p, conditional on fixed effects as described below) into the effects of additive genetics (V_a), permanent environment defined as the non-(additive) genetic component of among-individual differences, V_pe), and housing group (V_group), as well as residual variation (V_residual). We natural log-transformed emergence time prior to analysis to meet assumptions of residual normality and homoscedasticity. For all behavioural traits, we included fixed effects of assay repeat, the order within each group in which the fish was trialled (mean-centred continuous predictor), temperature (mean-centred and scaled to standard deviation units), time (in minutes from midnight, mean-centred and scaled to standard deviation units), age (mean-centred and scaled to standard deviation units), sex, and the generation from the breeding population. For shoaling tendency only, we incorporated an additional fixed effect of setup (as detailed above). We tested the significance of genetic variance for each behaviour by LRT comparison of the corresponding full model to one in which the (additive) genetic random effect was excluded.

Cortisol data were also natural log (ln) transformed for analysis. We formulated a bivariate model to test for both additive genetic variation and genotype-by-environment interaction (GxE) in cortisol levels across the two ‘contexts’ (i.e. samples retained for each individual at first and third confinement, denoted Cortisol₁, Cortisol₃). Random effects were first used to partition phenotypic (co)variance (conditional on fixed effects) into among-group and residual components. Fixed effects included the context-specific means, and overall effects of the order in which the fish was caught from each group for assay (mean-centred continuous predictor), temperature (mean-centred and scaled to standard deviation units), time of day (mean-centred and scaled to standard deviation units), age (mean-centred and scaled to standard deviation units), and sex. In addition, we included fixed covariates of body mass (mean-centred and scaled to standard deviation units) and a sex by body mass interaction (see Houslay et al. 2019 for rationale of controlling for body size effects on waterborne hormone levels in this way). Note that modelled in this way each individual is sampled only once for each context-specific cortisol trait so no random effect of individual identity is included. To test for additive genetic variation (V_a) we compared this first bivariate model to a second formulation that also included the (additive) genetic merit, but under the assumption that this is invariant with context within an individual (such that V_a1=V_a3 and r_a1,3=1 and there is no GxE). We then test for the GxE by comparing the second model to a third in which we allow GxE (i.e., the context-specific genetic variances are free to differ and the cross-context genetic correlation can be <+1).

Lastly, we built a multivariate animal model to estimate G and to test the hypothesised genetic integration among behavioural and physiological stress components. We retained only response traits that harboured significant V_a as shown in univariate models, and so the final model comprised response traits relative area, time in middle, track length, emergence time (log transformed), and Cortisol (log transformed). We multiplied (transformed) emergence time by −1 to simplify interpretation of estimated correlation structures (i.e., higher values for all behavioural traits then represent nominally ‘bolder’ behaviours). We also scaled all (transformed) response variables to standard deviation units. This was to facilitate model fitting, and also prevent scale effects complicating interpretation of eigenvectors of G. Fixed and random effects were fitted on each trait as specified for the univariate models. Note that one exception to this is that we elected to treat Cortisol as a single repeated-measures trait here (with two repeats, one per context) such that a permanent environment effect was now included. Fixed effects estimates are reported in the supplementary information (Table S1).

We specified additive genetic (G), permanent environment (PE), group (GROUP), and residual (R) covariance structures as unstructured matrices to be estimated. Note that R partitions observation-level covariances (as opposed to individual-level in PE) that are not definable or statistically identifiable if traits are not measured at the same time (i.e., all covariances relating to emergence time or Cortisol). Where this was the case we constrained specific covariance terms in R to equal zero. Estimates of PE, GROUP and R are provided in the supplementary information (Tables S2-S4). We tested for overall additive genetic covariance among the traits by comparing this model against a reduced one in which G was specified as a diagonal matrix (i.e., additive variances are estimated but covariances are assumed to equal zero). To aid the interpretation of covariance terms contained in G, we calculated the corresponding genetic correlations r_a from the full model. For any pair of traits . We also subjected our estimate of G to eigen decomposition to determine the proportion of additive genetic variation captured by each principal component and assess whether a single major axis of variation could indeed explain most of the genetic variance in the multivariate phenotype (consistent with a simple proactive-reactive coping style model). We estimated uncertainty on the trait loadings associated with each principal component (eigenvector) using a parametric bootstrap approach as described by Boulton et al (2014).

For visualisation of bivariate relationships at the additive genetic level, we used the R package ‘ellipse’ (Murdoch & Chow 2018) to determine the coordinates of an ellipse representing the approximate 95% confidence region of deviations based on the point estimate of G. We repeated this procedure for the corresponding regions defined from 5000 bootstrapped values of G (i.e., to indicate uncertainty arising from estimation of the genetic covariance structure itself). Best linear unbiased predictors (BLUPs) are used for visualisation only, not for any statistical analysis (Houslay & Wilson 2017).

To test for associations between all traits (i.e., including shoaling tendency) at the among-individual level, we also built a multivariate model as above with the addition of shoaling tendency and without estimating additive genetic effects. The estimates of all among-individual (co)variances are provided in the supplementary information (Table S5).

SUPPLEMENTARY MATERIAL

Table S1: Fixed effects estimates from the full multivariate animal model.

Table S2: Permanent environment (co)variance matrix from the full multivariate animal model.

Table S3: Group (co)variance matrix from the full multivariate animal model.

Table S4: Residual variance-correlation matrix from the full multivariate animal model.

Table S5: Among-individual (co)variance matrix from the multivariate model that excluded genetic effects.

Footnotes

• Revisions to introduction and discussion. Figures updated. Methods and results unchanged apart from a few cosmetic edits.