Integration between swim speed and mouth size evolves repeatedly in Trinidadian guppies and aligns with suction-feeding fishes

Many animals, including fishes, rely on integrated locomotor and feeding performance during prey capture to ensure the correct positioning relative to prey (Higham, 2007b; Kane and Higham, 2015). The need for integrated parts to work together successfully (Olson and Miller, 1958) can have consequences for the ability of traits to evolve (Goswami et al., 2014). Fishes are a useful taxa for understanding performance integration during prey capture because many exploit suction to capture prey (Wainwright et al., 2015), a behavior that relies on the locomotor system to maximize its effectiveness (Kane and Higham, 2014; Longo et al., 2015; Wainwright et al., 2001). Suction is generated as rapid mouth expansion pulls water and prey into the mouth (Day et al., 2015; Muller et al., 1982). However, this mechanism is temporally and spatially limited (Day et al., 2005, 2007; Ferry-Graham et al., 2001; Jacobs and Holzman, 2018) such that prey must be located close to the mouth at the time suction is generated and integration with swimming movements is critical.

Swim speed and mouth size have been established as primary drivers of the relationship between locomotor and feeding performance. As suction is generated prior to opercular opening, the faster a fish swims with this valve closed the more likely it will be to push the water away from its mouth rather than pull it in (Muller and Osse, 1984; Van Leeuwen, 1984) and swim speed is likely sub-maximal. Additionally, because water has to move faster for the same volume to travel through a smaller opening, fishes with smaller mouth size relative to body size create a steeper pressure gradient and increased pulling force (suction) on the prey (Wainwright et al., 2007; Wainwright and Day, 2007). Therefore, fish with smaller mouth apertures should rely on slower swim speeds and vice versa. This prediction of integration has been supported using correlations between swim speed at peak gape and peak gape in several fishes (Table 1).

Despite this broad pattern of integration, considerable residual variation in correlations remains. This variation suggests that integration may be flexible and may not represent an ultimate constraint for fishes. For example, suction velocity was not compromised with increased swim speed in bluegill sunfish (Higham et al., 2005). But mouth size was not reported and these fish used comparatively slow approach speeds where integration may be less necessary. Whether this relationship is a constraint for fishes that primarily use other feeding modes is also unknown. For example, in eels and labrid fishes which are specialized for biting, correlations break down or potentially change when fish rely on biting to capture prey (Collar et al., 2014; Rice, 2008). Therefore, the relationship between swim speed and mouth size may not generally apply to fishes specialized for alternative feeding modes, where suction forces may be less important.

Understanding the effects of specialization on the integration of locomotion and feeding has been confounded by additional differences among experimental taxa in previous research. Our current understanding is based on interspecific comparisons of divergent but primarily suction-feeding species (Table 1) or by comparing these fishes with different species that use biting (Rice, 2008; Rice and Westneat, 2005). However, comparisons of phylogenetically distinct fish exhibiting large differences in physiology or morphology can confound interpretations of adaptation and specialization (Harvey and Pagel, 1991). To control for this, we utilized divergent populations of Trinidadian guppies (Poecilia reticulata) to ask: (1) is the correlation between swim speed and mouth size generalizable to a morphological biting specialist utilizing suction to capture prey?; and (2) is this integration maintained between populations that are locally adapted to their environments?

Trinidadian guppies belong to a clade of fishes that is specialized for scraping and picking prey (Ferry-Graham et al., 2008; Gibb et al., 2008; Hernandez et al., 2008, 2009). However, guppies retain the ability to feed on suspended prey using suction, facilitating direct comparisons of behaviors with other suction-feeding fishes. Guppies can be found in contrasting habitats that vary in their exposure to predation pressure, prey resources and competition, leading to local adaptation in the ability to escape from predators using a locomotor fast-start response (Dial et al., 2015; Ghalambor et al., 2004; O'Steen et al., 2002; Walker et al., 2005), feeding morphology and prey capture behavior (Bassar et al., 2010; Dussault and Kramer, 1981; Palkovacs et al., 2009, 2011). Therefore, divergent populations may differ in their ability to employ suction and/or integrate this behavior with swimming. Wild guppies from both high-predation and low-predation environments consume similar prey types ranging from periphyton to invertebrates (although in different proportions and with different selectivity depending on the season) (Zandonà et al., 2011, 2017), suggesting that divergent populations may alternatively retain a generalized prey capture strategy and show similar levels of integration. Because low-predation populations have repeatedly evolved from adjacent high-predation populations (Willing et al., 2010), such comparisons also allow testing of how patterns of integration have changed between ancestral and derived populations.

Guppies (Poecilia reticulata W. Peters 1859) were collected in the Northern Range Mountains of Trinidad, West Indies, from replicate high-predation (ancestral) and low-predation (derived) populations found in independent river drainages on the north and south slopes (Table S1). These sites were chosen by considering accessibility and our confidence in their undisturbed or natural state. Only adult females were collected as they gestate young internally, suggesting that prey acquisition behaviors could have fitness consequences. Up to 15 fish from each population were used to facilitate detection of linear relationships (Garland and Adolph, 1994) but unforeseen circumstances limited sample size in one population (see discussion below). No individuals were intentionally excluded from analysis. Fish were shipped to the lab in Fort Collins, CO, USA and housed individually in 1.5 l tanks (Pentair Aquatic Eco-systems, Atlanta, GA, USA) on a recirculating system, similar to other studies (Broder and Angeloni, 2014; Handelsman et al., 2013). Fish were maintained on 40 μl of flake paste (1:1 ratio of ground flake and water; API tropical greens, Mars Inc.) fed once daily. All methods were approved by the Colorado State University IACUC (protocol 14-4998A).

Several factors affected the length of stay in the lab prior to filming trials (described below), resulting in fish being filmed at time points that varied between 43 and 385 days since arrival. As stage of pregnancy affects locomotor performance (Ghalambor et al., 2004, 2003) and wild-caught adults could not be considered virgin, fish were allowed to acclimate to the lab environment and give birth to previously fertilized developing offspring. The birth schedules of individual females were tracked daily and guppies were not used unless they had just given birth or birth rates slowed to fewer than 3 babies per month. In all cases, females were visibly lacking a distended abdomen, suggesting female performance was unlikely to be impacted by pregnancy. In addition to this, several unavoidable circumstances arose during this project that resulted in the need to re-sample populations or postpone filming trials (e.g. failure of equipment, unexpected death of fish and/or an inability to acquire local plankton prey in winter). However, we argue that this variation in lab acclimation time (Table S1) had little consequence for our results and conclusions. The most significant result of this variation was an increase in body size over time, as female guppies experience indeterminate growth. However, as differences in body size were anticipated a priori between high- and low-predation populations and this factor was included in statistical analyses, an acclimation effect was not modeled separately. The strength of divergence between populations (see below) despite these unfortunate circumstances further supports our conclusion that an acclimation effect was minimal and suggests either that differences in prey capture behaviors may be genetically fixed or that plastic responses may be relatively fixed in adult fish such that responses to the lab environment appear negligible.

To quantify prey capture behavior, fish were recorded as they captured live, locally collected, plankton prey that included cladocerans and copepods. Although these specific prey species have not been represented in natural diets (Zandonà et al., 2011), they are an ideal choice for these experiments because they (1) elicit a mid-water suction response in guppies comparable to that of other fishes and (2) perform evasive escape responses that challenge guppies and can help determine differences between populations in prey capture ability. Fish were starved for up to 3 days before filming and were removed from the recirculating system in their housing tanks and placed on a table so the long axis of the tank was perpendicular to the camera. Additional baffles were added to the tank to restrict the fish to an approximately 5 cm width×12 cm length×12 cm water height filming arena. Fish were allowed at least 15 min to re-acclimate before filming. All fish were provided the opportunity to practice with novel plankton prey prior to recording video. During trials, a 600 W tungsten light (Smith-Victor Corporation, Bartlett, IL, USA) was used to illuminate the filming arena. Fish were recorded capturing prey in the lateral perspective at 500 frames s⁻¹ using a single high-speed camera (Edgertronic SC1, Sanstreak Corp., San Jose, CA, USA) with a fixed lens (50 mm, Nikon, Nikon Corporation, Tokyo, Japan). At least 3 trials were saved per fish, and the best trial was used for analysis. These were trials where the fish was in focus and perpendicular to the camera, the fish's mouth was at least 1 cm away from a surface, and prey were verified as plankton. Video data were managed in accordance with best practices for video data management in organismal biology at level 2 or above (Brainerd et al., 2017).

Trials were digitized and analyzed using Matlab (version 7.14, MATLAB, MathWorks Inc., Natick, MA, USA) to determine swimming and feeding kinematics (Hedrick, 2008). All videos were digitized (Fig. 1A) starting several frames before mouth opening and ending several frames after mouth closing. Kinematic traces were smoothed using a quintic spline (Korff and McHenry, 2011) and visually inspected for accuracy of the smoothed trace. Gape was calculated as the distance between the upper and lower jaws and peak gape (hereafter PG) was the maximum value of this measurement. Swim speed was determined by calculating the speed of the approximate center of mass as the distance traveled between frames divided by the elapsed time between frames, or 2 ms. Swim speed was taken at the time of peak gape (hereafter SPG) as this measure of locomotor performance has shown the strongest correlation with mouth size in previous studies (Kane and Higham, 2011; 2015). Standard length (SL) was determined by calculating the distance between the lower jaw and the tail using a single frame prior to mouth opening when the fish was in a resting position.

Three general linear models (GLMs) were used to assess the relationship between mouth size and swim speed in guppies. These models differed in how they accounted for fish body size and the degrees of freedom available to detect significant differences. Qualitative comparisons among models were used to assess the effect of different data transformations on the results. In each case, predictors were sequentially removed from the model using the criteria that (1) effects with the lowest sum of squares at each iteration were removed and (2) effects could not be removed if they were contained in interactions with a larger sum of squares value. In this case, the predictor with the next lowest sum of squares was then evaluated. Predictors were removed until the predicted residual error sum of squares (PRESS, a leave-one-out cross-validation statistic of model fit) no longer decreased. This point indicates that additional predictors could result in the model overfitting the data. Once the model was selected, a partial eta-squared (η²) statistic was then calculated for each predictor to determine effect size relative to other predictors (Maher et al., 2013).

For the first model, raw data were used to determine the effect of swim speed on mouth size without accounting for differences in body size. This model provided a baseline to determine the effects of further transformations on the results. In this case, PG was the response and drainage (north versus south slope), predation (high versus low), SPG and all interactions in a full factorial design were included as predictors (Model 1). Second, all kinematics were log₁₀-transformed and the role of body size was added as a covariate to retain degrees of freedom and account for variation in body size among populations. In this model (Model 2), log(PG) was the response and drainage, predation, log(SPG), log(SL) and all interactions were included as predictors. Finally, residuals from regressions of log₁₀-transformed kinematics were used to account for the effect of body size while minimizing the degrees of freedom in the model. These residuals represent variation in mouth size and swim speed that is not attributable to the scaling effect of body size. Allometric relationships were determined for log-transformed data by first using a GLM to determine whether mouth size and swim speed scale similarly with size across populations, and second by using ordinary least squares regression to determine the scaling relationships. For the third model, PG residuals were the response and drainage, predation, SPG residuals and all interactions were included as predictors (Model 3). Models were compared qualitatively as well as using Akaike's information criterion (corrected) (AICc) scores (Burnham and Anderson, 2002). All statistical analyses were performed in JMP 12.1.0 (SAS Institute Inc., Cary, NC, USA).

All three statistical approaches supported the same result, showing a correlation between swim speed and mouth size, as in previous studies (Table 1), but that this relationship only occurs in low-predation guppy populations (Table 2). This pattern was more prominent when body size effects were accounted for (Fig. 1C). All models indicated that swim speed and the predation level×swim speed interaction are strong predictors of mouth size (Table 2). Despite a lack of significance for the interaction in Model 3, an effect size of 0.071 indicated that this factor is likely relevant. Therefore, we conclude that the relationship between swim speed and mouth size is present, but differs between the high- and low-predation populations in a repeatable way. Specifically, a lack of response is observed in high-predation populations (regression, north slope: F_1,12=0.8788, P=0.3670; south slope: F_1,11=0.5406, P=0.4776) compared with a strong response in low-predation populations (north slope: F_1,4=16.0461, P=0.0161; south slope: F_1,13=27.0239, P=0.0002) (Fig. 1B). The strong fit in the north slope low-predation population is an artifact of the small sample size and the true correlation of these traits is likely more similar to the south slope low-predation population. Despite this, all three models lead to the same general conclusion across drainages and suggest the presence of an effect in this population is likely accurate. The consistency of swim speed and population interaction effects across populations and statistical techniques suggests that the population-level differences in the correlation are potentially ecologically significant as guppies invade and adapt to low-predation environments.

The effect of body size was stronger for mouth size than for swim speed, suggesting that functional systems are not performing at similar levels. Body size was the only significant effect on mouth size [retained effects: predation, log(SL), and predation×log(SL) interaction; adjusted r²=0.62, ANOVA F_3,44=026.559, P<0.0001, effect test F₁=45.9098, P<0.0001, other effects P>0.13] and a single regression of log-transformed traits was performed prior to Model 3. However, the only significant effect on swim speed was the drainage×predation interaction [retained effects: drainage, predation, drainage×predation interaction, and log(SL); adjusted r²=0.325, ANOVA F_4,43=6.6475, P=0.0003, effect test F₁=7.8788, P=0.0075, other effects P>0.09] and log-transformed regressions for swim speed were performed for each population separately. Individuals can utilize a range of performance values depending on the stimulus, and kinematic traits will more likely relate to body size if they represent a maximum ability that is dependent on size. Therefore, guppies likely relied on near-maximum mouth opening to capture suspended prey, but sub-maximal swim speeds. In fact, guppies from all populations often slowed to a near-stop prior to mouth opening and forward velocity in many trials was due to suction generation (E.A.K., unpublished observation).

A covariate model including body size (Model 2) was the best model for understanding the relationship between swim speed and mouth size (Table 2). When body size was not accounted for, a drainage effect was recovered but this was due to the larger size of south slope fish (Table S1). Drainage effects were small in the additional models, suggesting that patterns recovered are repeated across drainages when size is accounted for (Table 2, Fig. 1C). Model 2 explained 67% of the variation in mouth size, which was a 12% increase compared with when size was excluded (Model 1), and body size was the largest effect on mouth size. This model also had the largest AICc value, corroborating this increase in explanatory power. Although Model 3 had fewer predictors, potentially more power to detect effects and a similarly high AICc value, it was constrained to explaining only residual variation. This is likely the reason the explanatory power of Model 2 was 4.5 times greater than that of Model 3 and its AICc values were also greater (Table 2). The poor fit of Model 3, the large effect of SL and the highest AICc values indicate that Model 2 should be given the most weight in drawing conclusions.

The presence of a relationship between swim speed and mouth size in low-predation guppies supports the hypothesis that a correlated relationship is common among fishes relying on suction to capture prey (Higham, 2007b; Kane and Higham, 2015), despite morphological specialization for other behaviors. When present and with the effect of body size removed (Fig. 1C, pooled size-residual r²=0.40), this relationship is stronger than a similar univariate measure in bluegill sunfish (Lepomis macrochirus), a species that is a model for suction-feeding mechanics. It was also stronger than in young silver-spotted sculpins (Blepsias cirrhosus), which were more similar in body size to guppies (Table 1). However, the magnitude of the observed relationship is consistent with that of other interspecific comparisons (Table 1), suggesting that the level of integration in low-predation guppies is reasonable.

The most surprising result is the lack of any observed integration in guppies from high-predation environments (pooled size-residual r²=0.05). This may be the first example of a lack of performance integration during a suction-feeding behavior. It is possible that univariate analyses may not be sufficient to capture integration. For example, tidepool sculpins (Oligocottus maculosus) were the only previously documented species lacking integration (r²=0.18), but this changed when a re-analysis included additional locomotor and feeding variables (Table 1). A similar scenario is unlikely in guppies as the correlation in high-predation guppies is an order of magnitude lower than that of tidepool sculpins. Alternatively, decreased performance may be associated with decreased integration such that at low swim speeds integration may not be necessary (Higham et al., 2005). In guppies, swim speeds are similarly low across populations (Fig. 1C) suggesting that integration should be possible in high-predation guppies as it is present in contrasting populations. Therefore, the kinematics of each functional system and their integration are likely decoupled, such that integration is difficult to predict based on underlying components. In this way, performance integration is an emergent organismal property of system-level kinematic performance (Korn, 2005) that may be responding to differences in selection pressures that manifest at the level of the whole organism.

We suggest that direct selection on integration or indirect selection on component functional systems could result in the absence of integration in some populations. Intraspecific competition for rare high-value prey items is strong in low-predation populations (Zandonà et al., 2017) and may drive selection for specialized solutions for prey capture in these habitats. Low-predation guppies may be prioritizing stability and accuracy during prey encounters that require precision (Kane and Higham, 2015), leading to an integrated phenotype that helps individuals compete for limited resources. When competition is reduced and prey availability changes seasonally (Zandonà et al., 2017), such as in high-predation environments, a generalist strategy including biting may be more prominent. A lack of integration may allow for flexibility across prey types or feeding modes (Kane and Higham, 2015) and provide a selective advantage. An alternative hypothesis is that selection may act on component systems, affecting the ability to integrate across systems. It is well documented that the superior ability to escape from predators in high-predation guppies is due to differences in both morphology and performance of the locomotor system (Dial et al., 2015; Ghalambor et al., 2004; O'Steen et al., 2002; Walker et al., 2005). If the whole-organism response is tuned for high speed or acceleration during interactions with predators (Walker et al., 2005), guppies may find it difficult to coordinate functional systems, even at slow speed. This may be similar to the difficulty we have in coordinating walking while manipulating a cell phone, where the ability to avoid collisions (integration) is compromised despite decreased walking speed, especially for those of us who do this integrated behavior less frequently. We suggest that specialization of one functional system may come at a cost to integrated behaviors at the level of the whole organism. Additional comparative studies of integration could provide a better understanding of when such trade-offs might occur.

Guppies have repeatedly and independently invaded and evolved in low-predation habitats from ancestors in high-predation habitats (Willing et al., 2010). Therefore, similarities in integration among replicate drainages suggest the repeated evolution of an integrated phenotype from a non-integrated ancestor. The strong presence of integration across fishes using suction coupled with the divergence in integration between guppy populations suggests that integration may provide an advantage in certain contexts and could readily evolve. However, these selection pressures might only be imposed above a critical threshold of specialization for prey capture behaviors. Below this threshold, regardless of kinematic similarity, integration may be absent because of the need for flexibility or in response to demands imposed on component systems. Nevertheless, this work provides key insights into the potential for integration to evolve, as well as how these patterns may govern diversification patterns in fishes.

We thank Porsche Robison for assistance with animal husbandry.