Proceedings of the Royal Society B: Biological Sciences
You have accessResearch articles

Immediate predation risk alters the relationship between potential and realised selection on male traits in the Trinidad guppy Poecilia reticulata

Alexandra Glavaschi

Alexandra Glavaschi

Department of Biology, University of Padova, Via Ugo Bassi 58/B, 35131 Padova, Italy

[email protected]

Contribution: Conceptualization, Formal analysis, Investigation, Methodology, Writing – original draft, Writing – review & editing

Google Scholar

Find this author on PubMed

Silvia Cattelan

Silvia Cattelan

Department of Biology, University of Padova, Via Ugo Bassi 58/B, 35131 Padova, Italy

Leibniz Institute on Aging - Fritz Lipmann Institute, Beutenbergstraße 11, 07745 Jena, Germany

Contribution: Conceptualization, Formal analysis, Methodology, Writing – review & editing

Google Scholar

Find this author on PubMed

Alessandro Devigili

Alessandro Devigili

Department of Biology, University of Padova, Via Ugo Bassi 58/B, 35131 Padova, Italy

Contribution: Formal analysis, Software, Validation, Visualization, Writing – review & editing

Google Scholar

Find this author on PubMed

Andrea Pilastro

Andrea Pilastro

Department of Biology, University of Padova, Via Ugo Bassi 58/B, 35131 Padova, Italy

Contribution: Conceptualization, Formal analysis, Funding acquisition, Methodology, Supervision, Writing – review & editing

Google Scholar

Find this author on PubMed



Imminent predation risk affects mating behaviours in prey individuals in a multitude of ways that can theoretically impact the strength of sexual selection, as well as its operation on traits. However, empirical studies of the effects of imminent predation risk on sexual selection dynamics are still scarce. Here we explore how perceived predation affects: (1) the relationship between the opportunity for selection and the actual strength of selection on male traits; and (2) which traits contribute to male fitness and the shape of selection on these traits. We simulate two consecutive reproductive episodes, under control conditions and perceived predation risk using experimental populations of Trinidad guppies. The opportunity for selection is higher under predation risk compared to the control condition, but realised selection on traits remains unaffected. Pre- and postcopulatory traits follow complex patterns of nonlinear selection in both conditions. Differences in selection gradients deviate from predictions based on evolutionary and non-lethal effects of predation, the most notable being strong disruptive selection on courtship rate under predation risk. Our results demonstrate that sexual selection is sensitive to imminent predation risk perception and reinforce the notion that both trait-based and variance-based metrics should be employed for an informative quantification.

1. Introduction

Predation is a ubiquitous selective force that can impact shape prey populations in a variety of ways, the most obvious being through direct killing, or lethal effects, that shape prey phenotype on evolutionary timescales. Predators affect prey not just in the long term by consuming it but also in the short term by intimidating it and disrupting its activity [1,2]. Unlike evolutionary (or background) predation risk, immediate predation risk can vary over very short timescales and exerts non-lethal effects, which are increasingly more often recognized as significant drivers of changes in prey populations [35].

Some of the best studied non-lethal effects of predation consist of adjustments in mating behaviour in both males and females. For instance, changes in female choosiness and/or preference for conspicuous male traits following predator threat perception have been reported in variable field crickets Gryllus lineaticeps [6], lesser waxmoths Achroia grisella [7] and swordtails Xiphophorus helleri [8,9]. Exposure to predator cues is associated with a higher number of failed copulations in Pardosa milvina wolf spiders [10] and with a reduction in male courtship in Schizocosa wolf spiders [11]. Predation threat perception can intensify sexual conflict if females engaged in antipredator behaviours are not able to evade unwanted mating attempts [12]. For example, female water striders perceiving a risk of predation are more likely to accept matings they would otherwise avoid, particularly with large males [13,14]. Similarly, females of pygmy squid Idiosepius paradoxus remove fewer forcibly inserted spermatangia in the presence of predator cues [15], with possible effects on sperm competition.

The adjustments in mating behaviour in response to an imminent predator threat have wide potential to influence the operation of sexual selection (e.g. its strength, shape, targeted traits), but whether or not they do so remains underexplored [16]. This is perhaps surprising, given that ecological factors, in general, affect sexual selection dynamics [1720] and predation pressure is one such factor that is virtually omnipresent in the wild. Differences in the operation of sexual selection in populations subject to different levels of predation on evolutionary timescales are relatively well understood [21]. However, immediate predation risk can vary over short time scales irrespective of the background predation intensity characteristic of the habitat [22,23] and it is possible for the same population to experience consecutive reproductive episodes under different levels of predation risk, with potentially different outcomes for the shape, strength and targets of sexual selection.

One species in which the effects of predation risk, both evolutionary and immediate, on aspects of reproduction are extensively documented is the guppy Poecilia reticulata. On its native island of Trinidad, this small freshwater fish with internal fertilization inhabits rivers and pools along a predation gradient, with consequences for ecology and life-history traits [21,24]. Males from low-predation localities reach maturation at a larger body size, are more brightly coloured and perform courtship behaviours (sigmoid displays; SDs hereafter) at relatively high frequencies, while males from high-predation populations are smaller, duller and rely more heavily on forced copulation attempts (gonopodial thrusts; GTs hereafter) [21]. In addition, predation regime indirectly affects postcopulatory traits linked to guppy sperm performance [25]. Observations in the laboratory in benign environments indicate that more colourful males with a higher courtship rate are usually preferred by females and have higher reproductive success [24,26]. Following perception of an immediate predation risk, males reduce the frequency of SDs while increasing the rate of GTs [27]. This change in male mating tactic is partly mediated by a reduction in female receptiveness [28] and preference for conspicuous male coloration [29]. At the same time, if given the chance to observe male behaviour in the presence of predators, females prefer bolder males (that show a higher propensity to take risks, i.e. [30]), who indeed benefit from a higher reproductive success compared to their shier counterparts [31].

We have previously shown that simulation of immediate predation risk increases the total opportunity for sexual selection (an indicator of the strength of sexual selection, expressed as the standardized variance in reproductive success, IRS) on guppy males [32]. This was mainly driven by a higher variance in mating success, suggesting that, at least in our experimental conditions, predation risk may be associated with stronger sexual selection on male precopulatory traits.

While IRS is a useful summary statistic that captures variance originating from all stages of sexual selection, it has been criticized for failing to distinguish between contributions from male traits and random variation in reproductive success not attributable to sexual selection [33]. Indeed, IRS does not represent the realised selection of traits, but an estimate of the upper limit of the strength of sexual selection [34,35]. The association between traits and fitness is better quantified by selection gradients, which have the added advantage of describing both linear and nonlinear selection [36]. A downside of selection gradients, however, consists of the fact that their interpretation relies on the assumption that traits under selection have been correctly identified and quantified. Despite ongoing debate [37,38], proponents of both trait-based and variance-based indices agree that both should be quantified whenever possible, to capture an accurate picture of sexual selection [35,39]. At the same time, direct comparisons between the two methods have been largely based on simulated datasets [40], while empirical tests have yielded mixed results [39,41,42]. Here we build on our previous findings by exploring the effects of immediate predation risk on (1) the relationship between the total opportunity for sexual selection (standardized variance in reproductive success, IRS) and the actual strength of selection on male traits, and (2) the targets and shape of selection. We conduct multivariate selection analyses followed by canonical rotations, focusing on male traits known to contribute to pre and postmating success in the guppy [19,30,4347].

Choosiness and polyandry are two important components of female mating behaviour, whose variation under immediate predation risk will influence the strength and shape of selection on male traits, although exact patterns are not easily predictable. Choosiness comprises the time and effort a female is prepared to expend towards making a choice [48] and it is expected to decrease in response to the perception of an immediate predation risk. In this case, mating should be more random with respect to male precopulatory traits (such as body size, coloration and courtship behaviour), leading to weaker sexual selection on these traits [49,50]. On the contrary, if polyandry decreases in response to predation risk (as observed in our guppy population), then the potential for selection on male precopulatory traits should increase [51] while the importance of postcopulatory traits for male reproductive fitness should decrease.

Based on observations of guppies in the wild and in the laboratory (see above), and considering the complex patterns of linear and nonlinear selection identified in our population [19,52], we can predict that the combinations of traits advantaged under perceived predation risk include boldness and GTs, while in control conditions they comprise orange coloration, SDs, gonopodium length, iridescence and GTs. In addition, in accordance with the reduction in female polyandry observed previously [32], we expect that postcopulatory traits would be less relevant for male fitness in the presence of predation compared to control conditions.

2. Material and methods

(a) Experiment overview

As described in [32], mating trials in the presence and absence of predation risk were carried out in populations consisting of six males and six virgin females (hereafter ‘replicates’, see below). Males were selected from stock tanks ensuring that, within the same experimental population, they could be individually recognizable by the human observer from colour patterns. The sequence of data collection is presented in the electronic supplementary material, figure S1. Briefly, males were isolated individually for three days, then subject to two boldness tests, photographed and stripped of sperm to standardize their initial sperm reserves (see electronic supplementary material for descriptions of these standard procedures). Five days after photography, males were subject to mating trials in the first treatment, with the second treatment following six days after. We tested a total of 20 male replicates both in the presence and absence of predation cues (i.e. a repeated-measure design), with treatment order balanced across the populations, while the groups of females differed between treatments. We determined the sample size based on the minimum number of individuals required for multivariate selection analyses comprising 10 traits (see below; [36]). Within each replicate, virgin females were size-matched by eye. Following mating trials, females were individually isolated into 4 L tanks equipped with nurseries until they produced a brood or for a maximum of 50 days. Average brood size was similar in the two treatments (control = 7.394, predation = 7.398; see [32]). Fin clips for the purpose of DNA extraction were obtained from males at the end of behavioural observations and from females after they produced a brood. Offspring were euthanized at 24–48 h of age with an overdose of MS-222 solution and preserved in pure ethanol at −20°C until processing. Paternity assignment was based on microsatellite loci analysis and is fully described in the electronic supplementary material. Following data collection, all adults were released into post-experimental tanks and not re-used in further experiments.

(b) Predation risk simulation and behavioural observations

The predator attack simulation followed an established procedure [32,53]. Males and females were housed separately in single-sex tanks (D × L × H: 39 × 89 × 40 cm) filled with water up to a depth of 25 cm, containing gravel and an aerator and covered by black paper on three sides. The predation treatment consisted of exposure to visual and olfactory cues: a predator model (fish bait resembling a pike cichlid Chrenicichla alta, 8 cm long) and alarm substance (guppy skin extract [54]). Fish were presented with the predator model in their holding tank, in the afternoon before the first mating trial, between 16.00 and 18.00. The model was submerged 2 cm above the tank floor and moved along the length at a constant speed for 10 min via a remotely operated pulley system. Halfway through the visual predator exposure, 2 ml alarm substance were gently pipetted into the tank water. These procedures were completed from behind a curtain to avoid the fish perceiving experimenter cues. The following morning, males were transferred into the corresponding female replicate tank, 4 ml alarm substance were added, and an observer blind to experimental treatment (AG) started live behavioural recordings. Three 30 min observations were carried out at 1 h intervals between 08.00 and 11.00, after which males were moved back to their original tank. These procedures were repeated for two consecutive days, amounting to a total of 180 min of observations per replicate. The fish were left undisturbed (i.e. no predator model was used and no alarm substance was added) in the control treatment.

During behavioural observations, the number of SDs and GTs were recorded for each male in each treatment. Female shoaling cohesion (measured by a second observer as part of a different study) increased under predation risk and males reduced their courtship rate [32] as observed in wild guppies exposed to natural predators [55,56], confirming that our setup successfully manipulated fish perception of predation risk. We found an effect of treatment order on the frequency of SDs, but the interaction between treatment and treatment order was not significant (see electronic supplementary material). One male died during the experiment, bringing the final sample size to 119 males.

(c) Statistical analyses

We estimated the relationships between male relative fitness and phenotype using separate multivariate selection analyses [36] followed by canonical rotations [57] for each predation treatment. We calculated fitness as the proportion of offspring sired by each male out of the total number of offspring produced within each replicate. We included (i) body size (measured as surface area), (ii) gonopodium length, (iii) area of orange coloration, (iv) area of iridescent coloration, (v) sperm number, (vi) sperm velocity, (vii) sperm viability, (viii) number of SDs, (ix) number of GTs and (x) boldness as predictor variables in both models. Black coloration is another component of guppy male sexual phenotype, but we did not include it as its contribution towards male fitness is indirect and only reported in certain populations [58]. Values for boldness, morphological and ejaculate traits are the same for the control and predation treatments, while sexual behaviour was measured during mating trials, so values for SDs and GTs differed between treatments. Predation risk was associated with a reduced average courtship rate [32], but between-individual differences remained constant (see electronic supplementary material). We standardized response variables to a mean of one and trait values to a mean of zero and standard deviation of one [36].

First, we conducted linear regressions including all trait estimates to obtain linear selection gradients (β). We then fitted second-order regressions including all linear, quadratic and correlational terms to estimate the matrices of nonlinear selection gradients (hereafter referred to as gamma matrices). Statistical packages underestimate quadratic coefficients by 0.5, therefore we doubled these estimates to obtain the correct values [59].

We compared the linear, quadratic and correlational coefficients between treatments with a Monte Carlo simulation with 10 000 iterations. We compared the observed differences (predation – control) in the coefficients with a random distribution of differences obtained by shuffling each male's reproductive success across treatments. Significance was calculated as the proportion of iterations in which the observed difference exceeded the 95% distribution in the random differences. We used a similar procedure to estimate differences in standardized variance in reproductive success (IRS; see also [32]), proportion of variance explained by male traits (R2 from full quadratic regressions) and total amount of variance explained by traits (IRS × R2). We obtained the standard errors of these point estimators with a bootstrap procedure based on 10000 samples.

Interpreting the size and significance of individual coefficients can underestimate the strength of nonlinear selection [60]. To overcome this problem, we conducted canonical rotations of the gamma matrices by multiplying them with the matrices of standardized traits [57]. Canonical rotations produce new axes of nonlinear selection characterized by loadings of the original traits, similarly to loadings of original variables on principal components obtained by PCA, and identify combinations of traits under selection beyond pairwise comparisons [61,62]. The number of canonical axes obtained is equal to the number of traits included in the analysis (eigenvectors M1 – M10 in each treatment; see below). Each eigenvector has an associated eigenvalue (λ), equivalent to the quadratic selection coefficient along the new axis. The strength of selection (curvature) along each eigenvector is given by its eigenvalue and the shape of selection by its sign, with positive eigenvalues indicating disruptive selection and negative eigenvalues stabilizing selection. We also rotated original linear selection coefficients (β) onto the new traits in order to obtain estimates of linear selection along the new axes (θ) [63]. We used the permutation procedure proposed by [64] to calculate the significance of each eigenvector. Analyses were conducted with R 4.0.3 [65] and PopTools 3.2 [66] in Microsoft Excel. We visualized fitness surfaces using the ‘Tps’ function of the ‘fields’ package in R [67].

3. Results

We previously demonstrated that the standardized variance in male reproductive fitness is significantly higher in the predation treatment (IRS = 1.073) compared to control (IRS = 0.633; figure 1; see also [32]). Of this total variance, the proportion explained by traits, as estimated by the multiple regression analyses, was significantly higher in the control treatment (R2 = 0.709) compared to predation (R2 = 0.546; delta ± SE = −0.162 ± 0.096, p = 0.026; figure 1). By multiplying the standardized variance in reproductive fitness observed in the two treatments by the proportion of variance explained by traits, we obtained an estimate of the strength of overall sexual selection on the male traits considered in this study. We found that sexual selection on traits was higher in the predation treatment (IRS × R2 = 0.586) compared to control (IRS × R2 = 0.449), although this difference was not statistically significant (delta ± SE = 0.138 ± 0.103, p = 0.127; figure 1).

Figure 1.

Figure 1. Differences between treatments in standardized variance in fitness (opportunity for selection, IRS), amount of standardized variance explained by male traits (IRS × R2 from second-order regressions), and proportion of variance (R2 from second-order regressions) explained by male traits. Asterisks indicate significant values (p < 0.05).

When reproductive fitness was analysed separately for each treatment, we did not find any significant linear selection gradients (β) (electronic supplementary material, table S5). We identified significant quadratic selection on body size (disruptive) and GTs (stabilizing) in the control treatment and on SDs (disruptive) in the predation treatment. In the control treatment, all male traits apart from sexual behaviour were involved in significant correlational selection (electronic supplementary material, table S5), whereas a single negative correlational gradient, between sperm number and boldness, was significant in the predation treatment (electronic supplementary material, table S5). When we compared the multiple regression coefficients between treatments, we identified significant differences in two linear coefficients (GTs and boldness), two quadratic coefficients (sperm velocity and SDs) and four correlational coefficients associated with sperm number in combination with body size, iridescence, sperm velocity and SDs, respectively. Among these, SDs showed the most pronounced difference between treatments (electronic supplementary material, table S6). In summary, we identified different predictors of fitness in the two treatments whose nonlinear coefficients, in turn, are associated with significant between-treatment differences. Thus, in the predation treatment, high and low frequencies of SDs predict high fitness (i.e. disruptive selection), whereas reproductive success in the control treatment is associated with a positive correlation between body size and sperm number and a negative correlation between area of iridescence and sperm number.

Canonical rotations produced ten new axes of selection in each treatment (table 1). A curvature different from 0 (given by the lambda value) indicates significant nonlinear selection along the respective axis. This was the case for five axes in each treatment (control: M1_C, M2_C, M8_C, M9_C, M10_C; predation: M1_P, M2_P, M8_P, M9_P, M10_P). For simplicity, we restrict our discussion to the strongest (highest absolute lambda value) and most significant (lowest p-value) two axes in each treatment [68]. Thus, the highest lambda values in the control treatment corresponded to M1_C and M10_C, which described disruptive and stabilizing selection, respectively. Axis M1_C was primarily loaded by body size (positive) and sperm number (negative), while axis M10_C was mainly loaded by area of iridescence (positive) and sperm viability (negative). The fitness surface defined by these two axes (figure 2) reveals peaks at extreme values of M1_C and average values of M10_C. The highest peak is associated with small body size and high sperm count in combination with intermediate values for area of iridescence and sperm viability, whereas the lower peak corresponds to males with large body size, low sperm count and again intermediate values of sperm viability and iridescent coloration (figure 2). The most significant axes in the predation treatment (with the highest associated lambda values) were M1_P and M10_P, describing disruptive and stabilizing selection, respectively (figure 3). Axis M1_P was mainly loaded by SDs (negative) and body size (positive) and axis M10_P was primarily described by sperm number (positive) followed by GTs (positive) and boldness (positive). The surface built by these vectors indicates that most successful phenotypes concentrate around extreme negative values of M1_P and intermediate values of M10_P. These males are small, perform SDs at high frequencies and have intermediate values for sperm count, GTs and boldness. A secondary area of high relative fitness, at the positive end of M1_P and average values of M10_P, is associated with intermediate to low sperm count, GT frequency and boldness and also large body size and low SD frequency (figure 3).

Figure 2.

Figure 2. Fitness surface (a) and two-dimensional contour plot (b) illustrating the relationships between relative fitness and major axes of selection in the control treatment. Axis M1_C represents disruptive selection and M10_C stabilizing selection.

Figure 3.

Figure 3. Fitness surface (a) and two-dimensional contour plot (b) illustrating the relationships between relative fitness and major axes of selection in the predation treatment. Axis M1_P represents disruptive selection and M10_P stabilizing selection.

Table 1. Eigenvectors obtained by canonical rotations of the gamma matrices and estimates of linear (theta) and nonlinear (lambda) selection gradients along each axis (M1-M10) in each predation treatment. Trait loadings on each eigenvector can be interpreted similarly to those obtained by a principal component analysis. The strength of selection (curvature of the surface) is given by eigenvalues and the shape by their signs (positive = disruptive; negative = stabilizing). Significant lambda values (p < 0.05) are indicated in bold. p-values obtained with permutation tests (5000 iterations) following [64].

theta (p value) lambda (p value) body size gonopodium orange iridescent sperm number sperm velocity sperm viability SD GT boldness
M1_C −0.038 (0.739) 2.000 (<0.010) 0.694 −0.020 −0.171 −0.277 −0.536 −0.157 −0.284 −0.071 0.079 0.093
M2_C 0.068 (0.414) 0.854 (0.020) 0.040 −0.071 −0.416 −0.281 0.212 0.412 −0.382 0.116 0.129 −0.593
M3_C 0.102 (0.171) 0.342 (0.086) 0.578 −0.230 0.377 0.140 0.504 −0.174 0.009 0.186 −0.208 −0.298
M4_C −0.046 (0.533) 0.204 (0.296) −0.186 −0.670 0.286 −0.429 −0.260 0.212 0.110 0.253 −0.235 0.078
M5_C 0.002 (0.992) −0.077 (0.690) −0.360 −0.196 0.052 −0.133 0.062 −0.712 −0.496 −0.193 −0.045 −0.126
M6_C −0.019 (0.788) −0.183 (0.216) 0.089 −0.286 −0.291 0.083 0.108 0.160 0.102 −0.738 −0.474 0.030
M7_C 0.046 (0.583) −0.338 (0.128) −0.006 0.210 −0.243 −0.028 0.193 0.055 −0.369 0.400 −0.595 0.453
M8_C 0.028 (0.744) 0.585 (<0.010) 0.040 −0.428 −0.640 0.190 0.098 −0.323 0.329 0.314 0.193 0.110
M9_C 0.094 (0.353) 0.996 (<0.010) 0.095 −0.011 0.033 −0.544 0.533 0.031 0.030 −0.199 0.387 0.468
M10_C 0.069 (0.445) 1.096 (<0.010) 0.017 −0.381 0.131 0.531 0.009 0.300 −0.509 −0.048 0.333 0.304
M1_P 0.159 (0.246) 1.501 (0.010) 0.405 −0.124 0.070 −0.100 0.222 −0.386 0.040 −0.774 0.076 −0.026
M2_P 0.15 (0.253) 1.412 (0.014) 0.405 −0.254 −0.519 −0.248 0.041 −0.226 −0.184 0.307 −0.339 0.380
M3_P 0.065 (0.584) 0.679 (0.206) 0.033 −0.112 0.427 0.598 0.213 −0.326 −0.446 0.171 −0.246 0.073
M4_P −0.111 (0.309) 0.220 (0.597) 0.021 0.245 0.231 0.076 −0.556 −0.341 0.287 −0.002 0.105 0.601
M5_P −0.041 (0.66) 0.036 (0.910) −0.457 −0.396 −0.167 0.005 −0.286 0.164 −0.505 −0.380 0.114 0.291
M6_P −0.009 (0.946) −0.034 (0.898) 0.495 −0.466 0.169 0.298 −0.085 0.547 0.205 0.038 0.203 0.170
M7_P 0.023 (0.819) −0.247 (0.202) −0.048 0.323 0.092 0.006 0.193 0.446 0.077 −0.308 −0.656 0.343
M8_P 0.048 (0.629) 0.524 (0.040) 0.238 −0.534 0.503 −0.476 0.009 −0.116 0.250 0.114 −0.296 −0.031
M9_P 0.061 (0.631) 0.913 (0.032) 0.389 −0.256 −0.365 0.411 0.350 −0.179 0.546 −0.036 0.005 0.167
M10_P 0.173 (0.127) 1.715 (<0.010) 0.074 0.132 0.196 −0.283 0.588 0.077 −0.139 0.149 0.486 0.479

The fitness surface associated to the control treatment is saddle-shaped, characteristic of nonlinear selection, in accordance with previous results [19,52], indicating that the alternative phenotypes at extreme ends of M1_C benefit from similar reproductive success (figure 2). In the predation treatment, the fitness surface is asymmetrical, with the peak at the negative extreme of M1_P higher than the one at the positive end, suggesting that relatively small males performing a high frequency of SDs are better favoured compared to males showing the reverse combination of traits (figure 3).

4. Discussion

In a previous study [32] we demonstrated that the perception of an imminent predation risk increases the total opportunity for sexual selection (IRS). This was due to an increased variance in male mating success and lower polyandry as indicated by the number of sires per brood. In the present study, we used the data on male reproductive success to test whether realised sexual selection differed in strength and shape in response to predation risk. Our first aim was to explore whether the greater opportunity for selection in the presence of predation risk resulted in stronger selection on male traits. Despite the larger IRS under imminent predation risk, the overall strength of sexual selection on male traits (expressed as the proportion of the variance explained by traits multiplied by the total variance) did not differ significantly between treatments. This was because in the predation treatment, male traits explained a lower proportion of the total variance in reproductive success compared to the control treatment (figure 1). A widespread limitation of studies aiming to quantify selection in small experimental populations (as in the current work) consists of the noise generated by random variation in trait values. Here we mitigated this issue by using a repeated measures design [69], therefore our quantification of sexual selection indices under different levels of predation is particularly informative. Our results provide reliable experimental evidence for the theoretical notion that the opportunity for sexual selection (and selection in general) does not necessarily equal realised sexual selection [33,37,38].

There are multiple non-mutually exclusive explanations for the observed relationships between the sexual selection metrics we computed in the two treatments. Here we discuss three. First, our result may be explained by the reduced female mating rate under predation risk and the consequent reduction in the contribution of traits under postcopulatory selection towards the variance in male reproductive success [32]. Second, female mate assessment could be less accurate under predation risk. Guppies are an extreme example of multiple male ornaments under simultaneous selection by female choice [19,52]. Evaluating complex phenotypes requires time and cognitive effort [70,71] that may be limited under an imminent predation threat. Assuming a theoretically preferred male phenotype, ‘errors’ in mate choice could occur more frequently in these conditions. However, stochasticity in female choice should reduce the variance in male reproductive success (if female mate choice was purely stochastic the variance in male reproductive success should tend to zero), which contrasts with our observation that predation risk was associated with an increased variance in both male mating and reproductive success. A higher variance in male reproductive success without a corresponding effect on overall selection on traits may arise if mate choice copying, which has been documented in female guppies both in benign and risky environments [29,72], becomes more important in the presence of predator cues. In this scenario, the initial choice of the first mating female may benefit the first male to mate, irrespective of his phenotype. Third, the higher portion of unexplained variance in the presence of predation could be a by-product of traits we did not quantify becoming more important for male fitness under these circumstances. Since it is difficult to know whether our analyses captured all components of male reproductive phenotype, this explanation cannot be ruled out.

Our second aim was to test whether predation risk influences which traits contribute towards male fitness and/or the shape of selection on these traits. Previous analyses [32] indicate that in our experimental populations under predation risk, selection on postcopulatory traits such as sperm number, velocity and viability should be weaker and selection on precopulatory traits should be stronger. We did find significant differences between selection gradients under the two conditions (electronic supplementary material, table S6), but our results partly deviated from this prediction. Two considerations are important here: (1) we assume that our boldness estimate in standard conditions reflects male propensity to take risks in other contexts, including mating trials; and (2) values for morphological traits are the same in both treatments, as they are unlikely to change over the duration of the experiment (electronic supplementary material, figure S1), while male sexual behaviour was recorded during mating trials in both treatments and is significantly affected by predation risk (electronic supplementary material). It is therefore not surprising that the largest difference in selection gradients involved sexual behaviour, although not in the expected direction (electronic supplementary material, table S6).

In agreement with previous work on the same population of guppies in conditions similar to our control treatment [19,52], we found that sexual selection was largely nonlinear. Such pattern of disruptive selection, with multiple phenotypes being similarly advantaged, has been described in multiple taxa including insects [73], birds [74] and fish [75]. While we found no significant linear β regression coefficients or θ coefficients on the M vectors in either treatment (table 1; electronic supplementary material, table S5), comparisons between treatments revealed that selection on GTs and boldness was more strongly linear in the predation treatment, and opposite in direction, compared with control (electronic supplementary material, table S6). In agreement with our expectation that postcopulatory traits should be more important for male reproductive success in the absence of predation, all ejaculate traits contributed towards male fitness in the control treatment, in combinations with morphological traits or boldness (electronic supplementary material, table S5). Canonical rotations confirmed these patterns: phenotypes under strongest selection in the control treatment were characterized by intermediate values for area of iridescence and sperm viability (M10_C) and either large body size and low sperm number, or small body size and large sperm number (M1_C, figure 2).

Extreme (high and low) frequencies of SDs were advantaged under predation risk. We also identified negative correlational selection between boldness and sperm number in the same treatment (electronic supplementary material, table S5), indicating that bolder males with low sperm count or shy males with high sperm reserves had a higher reproductive success. Canonical rotations confirmed disruptive selection on SDs under predation risk: axis M1_P was loaded positively by body size and negatively by SDs, with the highest relative fitness concentrated around the negative extreme (figure 3). The most advantageous phenotype in the presence of predation risk consisted of small body size, high SD frequency and intermediate sperm number, boldness and GT frequency (figure 3).

These patterns only partly reflect our expectations. We identified no contribution from orange coloration or gonopodium length in the control treatment. Also, we did not find a linear relationship between GTs and male fitness in either treatment, despite a significant difference (yet in the unexpected direction) in linear gradients (electronic supplementary material, table S6). We did not observe any successful coercive mating, but note that observations only covered 50% of the duration of the trials, thus we cannot exclude that forced copulations occurred. Even so, their contribution to male reproductive success was most likely limited, given the low insemination success of this mating tactic [76] and considering that virgin females are expected to be sexually receptive (i.e. to mate cooperatively more often). The use of virgin females was necessary to avoid the production of offspring from previously stored sperm that would have biased our measures of male reproductive success, but virgin females are a minority in the wild and their mating behaviour may not be representative to that of the population at large [26]. In addition, the stronger positive correlation between SDs and reproductive success in the predation treatment is surprising but also instructive, in our opinion, because it highlights a situation in which males respond to the presence of predator cues by reducing, on average, the frequency of the behaviour (electronic supplementary material), yet its importance for male reproductive success increases. This observation, coupled with the lower polyandry, suggests that cooperative female mating rate is a key determinant of the strength of sexual selection on male traits while sexual conflict plays a minimal role, at least in our population under these experimental conditions.

To conclude, our results demonstrate that sexual selection trajectories are affected by adjustments of mating behaviours in response to the perception of an imminent predation risk but in counterintuitive ways that cannot be easily predicted from lethal effects (e.g. selection against more conspicuous male phenotypes). Finally, we confirm that, on its own, the variance in male reproductive success is not a sufficiently informative predictor of the strength of sexual selection, at least in polyandrous species [33,37,38].

In our view, our study opens two main promising avenues for future research. First, our predation treatment is equivalent to a natural situation in which fish resume activity after a predator attack, but it captures a single ‘snapshot’ of male fitness, which may not be representative to lifetime reproductive success [77]. Thus, future studies could aim to extend data collection beyond a single reproductive event to account for the variable nature of imminent predation risk in the wild. Second, the best available evidence suggests that changes in guppy male behaviour under predation risk are mediated by females [28], but we cannot fully confirm from the current experimental design if this is also true for the observed consequences for sexual selection dynamics. Therefore, future work could aim to explore if the operation of sexual selection under predation risk is mostly male or female-driven, or a combination of both.


Our data collection protocol was approved by the University of Padova Institutional Ethical Committee (permit no. 256 /2018).

Data accessibility

The data are provided as electronic supplementary material [78].

Authors' contributions

A.G.: conceptualization, formal analysis, investigation, methodology, writing—original draft, writing—review and editing; S.C.: conceptualization, formal analysis, methodology, writing—review and editing; A.D.: formal analysis, software, validation, visualization, writing—review and editing; A.P.: conceptualization, formal analysis, funding acquisition, methodology, supervision, writing—review and editing.

All authors gave final approval for publication and agreed to be held accountable for the work performed therein.

Conflict of interest declaration

We declare we have no competing interests.


A.G. was supported by a CARIPARO scholarship for non-Italian PhD students and by a MIUR PRIN Grant (grant no. 20178T2PSW). S.C. was funded by a post-doc fellowship from the University of Padova. A.D. was supported by a grant from the University of Padova (grant no. STARS-CoG-2019). A.P. was supported by grants from University of Padova (grant nos. PRAT-CPDA120105-2012 and BIRD-175144-2017).


We are grateful to graduate students Martina Bonaldi and Marta Guerra for their valuable help with male traits quantification and paternity analyses, respectively. We thank Sasha Dall and two anonymous reviewers for their comments and suggestions on a previous version of this manuscript.


Electronic supplementary material is available online at

Published by the Royal Society. All rights reserved.