A complex performance landscape for suction-feeding reveals constraints and adaptations in a population of reef damselfish

The ability to predict how multiple traits interact in determining performance is key to understanding the evolution of complex functional systems. Similar to Simpson’s adaptive landscape, which describes the fitness consequences of varying morphological traits, performance landscapes depict the performance consequences of varying morphological traits. Mapping the population’s location with respect to the topographic features of the landscape could inform us on the selective forces operating on the traits that underlie performance. Here, we used a mechanistic model derived from first principles of hydrodynamics to construct a hypothetical performance landscape for zooplankton prey capture using suction feeding. We then used the landscape to test whether a population of Chromis viridis, a coral reef zooplanktivore, is located on a performance peak or ridge based on measurements of kinematic variables recorded in-situ during undisturbed foraging. Observed trait combinations in the wild population closely matched regions of high feeding performance in the landscape, however the population was not located on a local performance peak. This sub-optimal performance was not due to constraints stemming from the observed trait correlations. The predominant directions of variation of the phenotypic traits was tangent to the ‘path of steepest ascent’ that points towards the local peak, indicating that the population does not reside on a “performance ridge”. Rather, our analysis suggests that feeding performance is constrained by stabilizing selection, possibly reflecting a balance between selection on feeding performance and mechanical or genetic constraints.

Introduction 1 analysis of the 110 observed strikes. We then calculated the angle in the 6-1 dimentional space, between the two vectors as: is the vector of coefficients (eigenvectors) for PC1 and ‫ݒ‬ Ԧ 1 9 The density of the contours was different for the different trait combinations, 1 indicating variable performance gradients for each set of traits. Of the six traits and 2 their paired interactions, all but two interactions were significant in the GAM model 3 (Table S2), indicating the importance of trait combinations in determining 4 performance. The range of strain rate thresholds estimated across the landscape is 5 ecologically plausible, ranging from 1-10 s -1 (Green et al., 2003;Kiorboe and Visser, 6 1999;Viitasalo et al., 1998;Visser, 2001), reinforcing the validity of this model in 7 predicting feeding performance based on phenotypic data. The median strain rate threshold of the observed population (2.16 s -1 ) was 1 2 significantly lower by ~25% than the mean median of the 1000 "uncorrelated BM" 3.09 s -1 ), and was outside the distribution of median performance in the simulated 1 5 populations (p<0.001; Figure 5A), refuting the hypothesis that the distribution of 1 6 phenotypic traits is unrelated to the performance landscape. While the performance of the observed population was higher than that of the 2 0 simulated population, the observed population was located off (local and global) 2 1 performance peaks. This was evident by the individual performance surface 2 2 calculated for the six traits (Fig 6). Five of the selection surfaces were significantly 2 3 larger than zero, indicating that a change equivalent to 0.1 sd of the trait value would 2 4 significantly increase mean performance (decrease threshold strain rate). TTPJP was 2 5 the only trait for which individual performance surface was not significantly different 1 than 0 ( Fig 6D). The RSM analysis identified a local performance peak at peak gape 2 diameter of 5.1 mm, TTPG of 0.019 s, jaw protrusion distance of 6.6 mm, TTPJP of 3 0.007s (TTPG -0.012), ram speed of 61.8 mm/s and a timing differential of 0.012s.

4
None of the individuals in the observed population had such a combination, however 5 they are within the range observed for similar-sized fish. Scenario III: the population evolves toward the 'adaptive peak', but has not 8 reached there 9 We calculated the angle in the 6-dimentional space between the eigenvector of 1 0 PC1 of the observed data (explaining ~35% of the total variance), and the vector of 1 1 the steepest descents, which depicts the direction of change, on the landscape, that 1 2 will result in maximal performance increase. Contrary to the "performance ridge" Are trait correlations responsible for the off-peak location of the population? 1 9 Trait correlations can limit the distribution of traits in the trait space and therefor 2 0 constrain performance, providing a possible mechanism for the off-peak location of 2 1 our population. However, in contrast to this expectation, observed trait correlations correlations resulted in a higher performance (lower strain rate threshold) compared to 2 4 the "uncorrelated BM" populations. However, the median strain rate threshold of the 2 5 1 "correlated BM" populations was still significantly higher by ~15% than the observed 1 population (mean 2.55 s -1 ; 95% confidence intervals for = 2.30 -2.81 s -1 ; P=0.005; 2 Figure 5B). The positive effect of trait correlations was consistent for all 6 traits; Thus, the observed trait correlations are associated with higher feeding performance 7 than the non-correlated case, but it is unlikely that they result from trait distribution 8 along a performance ridge. In this study, we refine the "performance landscape" framework suggested by performance landscape beyond the ranges observed in the sampled population, and 1 6 curtails "edge effects" that originally resulted in higher inaccuracy of the landscape 1 7 near the population edge. Mapping the observed population on the landscape 1 8 indicated that the population was located on the "upper slopes" of an adaptive peak, 1 9 but not on the peak itself. Nevertheless, trait correlations did not seem to limit 2 0 performance, as breaking trait correlations reduced performance. The major axes of 2 1 phenotypic variation were tangent to the direction to the local performance peak, 2 2 inconsistent with the scenario of trait distribution along a performance ridge (Fig 1).

3
We conclude that our approach can help uncover the relationships between 2 4 performance and trait evolution in complex functional systems. The Chromis viridis population is located off a performance peak 2 The performance landscape for suction feeding reported here was complex, 3 rugged, and featured performance troughs and ridges ( Figure 4). The landscape 4 clearly showed that different phenotypic traits have a different effect on performance 5 as indicated by the different slopes of the performance surfaces (Fig 4), and the 6 different selection gradients for the different traits (Fig 6). Moreover, these slopes 7 often change in relation to the values of the other traits. For example, for low values 8 of time to peak jaw protrusion (TTPJP < 0.02 s), the effect of time to peak gape 9 (TTPG) on performance is much stronger than at high values (TTPJP > 0.06 s; Fig 4).

0
This is probably responsible for some of the variation in individual performance 1 1 gradients (Fig 6). These trends demonstrate the complexity of the relationship 1 2 between form and function in this functional system, and the importance of an 1 3 integrative analysis to accurately model it.

4
Functional systems are often considered to be 'optimized' for performance, i.e.
1 5 the distribution of phenotypic traits within a species or a population is assumed to 1 6 correspond to performance peak (Bishop et al., 2008;van Leeuwen and Muller, 1984).

7
This notion is often supported by using the measured performance data to construct a 1 8 putative landscape (Arnegard et al., 2014;Arnold and Bennett, 1988). More often, 1 9 and especially in an inter-specific context, it is the distribution of phenotypic traits 2 0 themselves that is used to infer the existence and location of a "peak" (Collar et al., correspondence (or lack thereof) between the performance landscape and the 2 3 distribution of species traits are poorly demonstrated.

4
As pointed out by Arnold, (Arnold, 2003) it is problematic to infer the location of 1 performance peaks from such data. Therefore, the location of the population with 2 respect to performance peaks and ridges is difficult to resolve and is largely unclear.

3
Here we show that suction feeding performance of Chromis viridis is not optimal, 4 although it is significantly higher than that expected under a model of random motion  However C. viridis is a zooplankton specialist, feeding mainly on copepods (Allen 9 and Randall, 1980), and it is highly likely that the performance landscape can be 1 0 reduced to one prey escape strategy (i.e. strain-sensitive prey). Unlike in mouth- functions impose strong direct selection on the traits we measured.

5
Instead, we speculate that suction feeding performance is constrained by   Genetic linkages are also expected to constrain the distribution of traits in the 7 observed population (Lande and Arnold, 1983;Schluter, 1996). In our simulated 8 populations, we accounted for such correlations, which had a strong effect on the 9 phenotypic distribution of our populations and on performance.

0
Alternatively, it could be that that the location of the population off the 1 1 performance peak could result from a misalignment of the performance and fitness 1 2 peaks. For example, it could be that the performance peak allows the capture of highly 1 3 strain-sensitive prey. However, the energetic cost of feeding with a morphology that 1 4 sits upon that peak is high. If the abundance of such highly evasive copepods is very 1 5 low compared to other (less evasive) copepod species, the fitness peak (or peak 1 6 energetic gain) will be shifted towards the current location of the population.

7
However, we are unaware of community-wide data on the distribution of 1 8 hydrodynamic performance of copepods that would enable testing this idea.