Selection and the direction of phenotypic evolution

  1. François Mallard
  2. Bruno Afonso
  3. Henrique Teotónio

  • Curated by eLife

    eLife logo

    eLife assessment

    This is a potentially important paper that takes advantage of an unusually comprehensive evolutionary genetic dataset to tease apart the relationship between genetic variation and phenotypic divergence over the ~medium term (50 generations). The questions addressed have broad relevance across evolution, conservation, and agricultural fields, and this paper will particularly appeal to evolutionary biologists. Nonetheless, the strength of evidence is incomplete for the major results and conclusions reported.

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

Predicting adaptive phenotypic evolution depends on invariable selection gradients and on the stability of the genetic covariances between the component traits of the multivariate phenotype. We describe the evolution of six traits of locomotion behavior and body size in the nematode Caenorhabditis elegans for 50 generations of adaptation to a novel environment. We show that the direction of adaptive multivariate phenotypic evolution can be predicted from the ancestral selection differentials, particularly when the traits were measured in the new environment. Interestingly, the evolution of individual traits does not always occur in the direction of selection, nor are trait responses to selection always homogeneous among replicate populations. These observations are explained because the phenotypic dimension with most of the ancestral standing genetic variation only partially aligns with the phenotypic dimension under directional selection. These findings validate selection theory and suggest that the direction of multivariate adaptive phenotypic evolution is predictable for tens of generations.

Article activity feed

  1. Version published to 10.7554/elife.80993 on eLife
  2. eLife

    Author Response

    Reviewer #1 (Public Review):

    The introduction does not clearly set up the background for the key questions that the manuscript addresses. One of the key parts of the manuscript is to attempt to determine whether locomotory behaviour evolves because of direct or indirect selection of the traits. However, the authors don't provide an argument for why a salty environment would select for locomotory traits. Indeed, in the discussion, the authors point out that it is likely an unmeasured trait (body size) correlated with locomotory traits that are under selection. They present arguments for why this might be the case and point to un-included data that show body size significantly genetically covaries with all of the traits studied. Since the authors appear to have these data, and one of their key questions is comparing direct vs. indirect responses to selection, it would be more powerful to include the body size data and estimate selection on all traits together.

    We now include body size in all of our phenotypic and genetic analyses. We also include estimates of selection gradients from the ancestral selection differentials and the Gmatrix. We detail in the Introduction the biological significance of locomotion traits and their potential relationship with body size, in low and high salt environments. The experimental results show that divergence in locomotion traits (Figure 6) correlates with adaptation (Figure 5), because of direct and indirect selection (Figure 9).

    Phenotypic plasticity was estimated from a series of univariate models, with estimates arranged in a vector. As the authors point out in the manuscript, traits that are not included in a model but covary with traits that are can largely bias estimates of the traits that are included. For this reason, it would make sense to estimate phenotypic plasticity using a multivariate model, as has been done for G matrices.

    We analyze the ancestral phenotypic plasticity and the phenotypic divergence during evolution using a multivariate approach (MANOVA). This approach simplifies the text as from the eigen decomposition of the SSCP matrices we can estimate canonical traits of ancestral phenotypic plasticity (pmax; see Table 1 with notation definitions) and phenotypic divergence in the new target high salt environment (dmax). We continue to do the univariate analysis as it allows us to estimate BLUPs for each inbred line (used for visual representation), as well as the significance of phenotypic divergence at each replicate population relative to the ancestral population (delta_q). Both multivariate and univariate approaches led to similar results (shown as supplementary figures).

    The estimation and interpretation of G matrices are a critical part of the manuscript. The authors state that broad sense estimates of G are a good proxy for additive genetic variation in this system, but in the Discussion they also state that overdominance was likely important during evolution to the salt environment, leading to some lack of clarity on whether dominance is important or not.

    We are sorry for the lack of clarity. We have eliminated the discussion on overdominance as it was peripheral to our results. Broad-sense genetic variances should be a good proxy for additive genetic variances when there is no inbreeding depression and no directional dominance or dominance epistasis; cf. Lynch and Walsh 1998. We previously showed that there is no inbreeding depression for the trait we use as surrogate for relative fitness (self-fertility) and also that there is no directional dominance for locomotion behavior traits. We now explain our use of broad-sense genetic (co)variances as a proxy for additive genetic (co)variances in the Introduction and Methods.

    It is also unclear how uncertainty in estimated G matrices was assessed. Showing that G differs from noise is critical to the majority of the results presented. The authors cite Morrissey and Bonnet (2019) as providing the method for generating the null distribution of G, however, this paper does not appear to propose or describe a method to do this.

    Thanks for this comment. Morrissey and Bonnet (J Heredity, 2019) was incorrectly cited and the explanation for finding the expected noise distributions was misleading. In brief, we produced a set of 1000 G-matrices each computed after shuffling the line ID and the block ID from the phenotypic dataset. This was done to produce random expectations of the genetic variances as the MCMC estimates are positive-definite. We computed the posterior mode for each of these 1000 G-matrices to obtain a null distribution (shown in orange). To infer significance, we compared the posterior mode of the empirical estimate with the 95% CI of the posterior mode distribution obtained from the randomized G-matrices. When determining which eigenvectors explain standing genetic variation we also used the distribution of posterior modes of the randomized G-matrices. However, as pointed out by Sztepanacz and Blows (Genetics, 2017), the eigenvalues of the eigenvectors do not follow a uniform distribution, as would be expected by chance. Because of this we asked the question of whether the amount of variance in the eigenvectors of the empirical G-matrix (gmax, g2, etc.) was expected, by projecting the random G-matrices onto these eigenvectors. This is a null that is conditional on the observed data. We show these results in Figure 2 - supplement figure 3. Both approaches are similar, particularly for the first 2 eigenvectors. There is now a paragraph in the Discussion about finding potential consequences for adaptation of traits with little genetic variance.

    Although the figure captions state that they are showing estimates of genetic variances, it appears to be heritability (bounded between 0 and 1). Whether the authors are studying heritability or genetic variance is an important difference, particularly in the context of a changing environment and phenotypic plasticity, where environmental variation is important and expected to change. For example, the result that G is smaller in evolved populations could simply be due to their being larger environmental variance in the salt environment (as you would expect). This is unrelated to an evolutionary response.

    There might have been some confusion because transition rates are positive and not normally distributed. To achieve normality they were log transformed. We have not reported estimates of heritability, all estimates presented are of genetic variances, unscaled. The only exception is body size where the raw data was multiplied by 50 in order to have a similar phenotypic scale as the transition rates when estimating genetic (co)variances, not heritability. We agree that the evolution of environmental stochastic variance is interesting but not immediately relevant to the questions we address.

    It seems that comparisons to the ancestral population were done for A160, not the founding population for each evolved line at G0. It is not clear whether the founder effects of each replicate are important and if this is the most appropriate comparison (the Discussion suggests that founder effects are important).

    We have better detailed in the Methods, and also with an introductory section in the Results section, the derivation of the experimental populations. The population acronyms might have been misleading. The A6140 is a population that was domesticated to the lab conditions for 140 generations (replicate #6 of the domestication process). We report the evolution of 3 GA populations, which were all derived from A6140 with minimal sampling problems for the estimated effective population sizes (sampled 10^4 individuals from A6140 for each GA, for Ne of 1000 during domestication - Chelo and Teotónio Evolution 2013 -). Therefore, GA populations after 50 generations of evolution are appropriately compared with their (unique) ancestor population. We no longer discuss potential founder effects.

    Overall, there is much interesting data collected and analysed in this manuscript, addressing a valuable question. However, it is not obvious whether the estimates of G matrices are different from noise, and heritability may not be the most appropriate scale to ask questions about phenotypic plasticity and evolution in a novel stressful environment that may affect levels of environmental variation.

    Please see previous replies. Our ancestral G-matrix estimates indicate that at least 3 eigentraits are different from random expectation in both environments (Figure 2, supplement figure 3), and in high salt evolved populations continue to have more than expected genetic variance at 3-5 eigentraits (Figure 7, supplement 2). We are conservative in these estimates as depending on the null we could consider more eigentraits. In the previous version of the manuscript we concluded that only 2 ancestral eigentraits were orthogonal due to an error in the code (we did not divide by 2 the null expectations). But even presuming that only one eigentrait (gmax) has genetic variance in the ancestral population, we previously reported that mutational variance is not in the same trait (see Mallard et al., G3, 2023; and mmax in Table 3), and further that the trait under selection is neither gmax or mmax (compared in Table 3 the selection gradients with gmax or mmax). At a minimum there are 3 genetically or environmentally independent traits. As noted in previous replies, we estimate and present genetic variances throughout. We do not present estimates of environmental variances and feel that doing so would make the manuscript overly complicated.

    Reviewer #2 (Public Review):

    Response to selection: It was not clear to me that it was appropriate to interpret locomotor behavior as having evolved in response to the salinity environment. Specifically, where is the evidence that any change in trait means is a (direct or indirect) response to selection imposed by increased salinity rather than the neutral drift of a trait due to the reduction in population size caused by the salinity? Strong evidence of adaptive evolution would be provided by all 3 replicates significantly diverging from the ancestor in the same direction. Model 2 seems to aim to test the null hypothesis that the three replicates diverged from one another via a random effects model - but with only three replicates, there is very low power, and variance is likely to be estimated as zero. I'm not sure what is shown in Tables 3 & 4, or how these results relate to models 2 & 3, so my interpretation of the information may be incorrect. Nonetheless, and noting that the errors around estimates are not presented, there seems to be considerable heterogeneity in size and direction of divergence between replicates for most of the traits. Is this study really dissecting responses to directional selection, or is it dissecting drift?

    We have modified the statistical modelling of the phenotypic data. Model 2 is no longer presented. We provide a MANOVA multivariate analysis equivalent to model 2 (with replicate populations as fixed effects) but now including both environments, together with the univariate models. MANOVA results show that all traits are significantly different across populations (i.e., at least two populations differ from one another). The fitted estimates from the MANOVA are not reported with errors in R but it is obvious that not all traits evolved in each replicate GA population (Figure 6). We therefore tested the difference between each of the evolved populations and the ancestral population using a univariate approach (Figure 6, supplemental source data table 2). In this univariate analysis, block was modeled as having random effects (which we could not model with MANOVA). In the high salt environment, the replicates GA 1,2,4 differed significantly for respectively 4, 6 and 4 transitions rates (out of 6). The traits are all evolving in the same direction, and this even when the trait difference between evolved and ancestral populations is not significant. We provide compelling evidence of parallel evolution and thus selection (see review about how to infer selection in evolution experiments in Teotónio et al. Genetics 2017). We tried to be exhaustive in our statistical reporting but would happily provide additional details if requested.

    What are the traits, and what is the confidence in G? My outsider's interpretation of these results is that defining 6 transition states is a way of getting at a single behavioral trait, and I was not convinced that these data were suitable for addressing questions about multivariate evolution. Genetic parameters were estimated using MCMCglmm, which imposes boundaries on estimates. The authors state that they followed Morrissey and Bonnet 2019, but I was unable to infer what this means with respect to accounting for the contribution of sampling error to covariances (or how they accounted for the positive variance constraint). Because I was unsure how sampling error was being assessed for G, I was not confident about the interpretation of statistical support for individual parameters, or for eigenvalues of G. Following this forward, if the measured characteristics constitute a single trait, with an entirely shared genetic basis, then the results of strong alignment of everything with gmax makes complete sense - there is a single trait, that is heritable and plastic, and for which the mean evolved.

    Our initial draft was misleading and we now provide more detailed description (see also replies #5 and #12 above). We computed 1000 randomized G matrices to account for the constraints imposed by the MCMCglmm algorithm. This should account for the bias inherent with variance estimation and the eigen decomposition we did given our sample sizes. You will find that all 6 transition rates show genetic variance (Figure 2, supplement figure 2) and that up to three eigentraits have more genetic variance than the randomized G-matrices (Figure 2, supplement figure 3).

    The 6 transition rates are the mathematical description of changing movement states in 1-dimensional space (under memoryless assumptions). A priori we do not know how many relevant traits there are, if they are genetically or environmentally independent. To help the reader, we provide a Table 3 with the trait loadings for the several canonical traits of phenotypic plasticity, divergence and selection. The first canonical trait of standing genetic variation, gmax, is indeed aligned with phenotypic divergence (dmax; Figure 8, panels A and B) and with the axis of genetic variance reduction during evolution (emax; Figure 8, panels C and D), but not with ancestral plasticity (pmax; Figure 3) or mutational variance (mmax, from Mallard et al. G3 2023). pmax, for example, is aligned with g3, the third eigenvector of the ancestral G matrix. Note, however, that we do not have any power to detect the influence of g2 or g3 on phenotypic divergence or genetic divergence (Figure 8), though they together explain about 15% of the genetic variance. This is because performing such a test would require an alignment of the deviations in divergence not explained by gmax with g2 or g3. We now mention this issue in the Discussion. Overall, however, there are clearly several behavioral traits.

  3. eLife

    eLife assessment

    This is a potentially important paper that takes advantage of an unusually comprehensive evolutionary genetic dataset to tease apart the relationship between genetic variation and phenotypic divergence over the ~medium term (50 generations). The questions addressed have broad relevance across evolution, conservation, and agricultural fields, and this paper will particularly appeal to evolutionary biologists. Nonetheless, the strength of evidence is incomplete for the major results and conclusions reported.

  4. eLife

    Reviewer #1 (Public Review):

    In this manuscript, the authors use experimental evolution in C. elegans to ask whether evolution in locomotor traits in a high salt environment can be predicted using the Roberston-Price identity and whether the evolutionary response to the salt environment is due to direct selection on locomotor phenotypes or indirect selection via an unmeasured trait. The authors also examine the alignment between phenotypic plasticity in the ancestral environment aligns with G.

    The experimental evolution system in C. elegans is a powerful model system to test these types of questions, in particular, because it is possible to resurrect ancestral populations and compare them contemporaneously to evolved lines. It is also possible to estimate broad sense genetic covariance matrices from inbred lines, as the authors have done here.

    The introduction does not clearly set up the background for the key questions that the manuscript addresses. One of the key parts of the manuscript is to attempt to determine whether locomotory behaviour evolves because of direct or indirect selection of the traits. However, the authors don't provide an argument for why a salty environment would select for locomotory traits. Indeed, in the discussion, the authors point out that it is likely an unmeasured trait (body size) correlated with locomotory traits that are under selection. They present arguments for why this might be the case and point to un-included data that show body size significantly genetically covaries with all of the traits studied. Since the authors appear to have these data, and one of their key questions is comparing direct vs. indirect responses to selection, it would be more powerful to include the body size data and estimate selection on all traits together.

    Phenotypic plasticity was estimated from a series of univariate models, with estimates arranged in a vector. As the authors point out in the manuscript, traits that are not included in a model but covary with traits that are can largely bias estimates of the traits that are included. For this reason, it would make sense to estimate phenotypic plasticity using a multivariate model, as has been done for G matrices.

    The estimation and interpretation of G matrices are a critical part of the manuscript. The authors state that broad sense estimates of G are a good proxy for additive genetic variation in this system, but in the Discussion they also state that overdominance was likely important during evolution to the salt environment, leading to some lack of clarity on whether dominance is important or not. It is also unclear how uncertainty in estimated G matrices was assessed. Showing that G differs from noise is critical to the majority of the results presented. The authors cite Morrissey and Bonnet (2019) as providing the method for generating the null distribution of G, however, this paper does not appear to propose or describe a method to do this.

    Although the figure captions state that they are showing estimates of genetic variances, it appears to be heritability (bounded between 0 and 1). Whether the authors are studying heritability or genetic variance is an important difference, particularly in the context of a changing environment and phenotypic plasticity, where environmental variation is important and expected to change. For example, the result that G is smaller in evolved populations could simply be due to their being larger environmental variance in the salt environment (as you would expect). This is unrelated to an evolutionary response.

    It seems that comparisons to the ancestral population were done for A160, not the founding population for each evolved line at G0. It is not clear whether the founder effects of each replicate are important and if this is the most appropriate comparison (the Discussion suggests that founder effects are important).

    Overall, there is much interesting data collected and analysed in this manuscript, addressing a valuable question. However, it is not obvious whether the estimates of G matrices are different from noise, and heritability may not be the most appropriate scale to ask questions about phenotypic plasticity and evolution in a novel stressful environment that may affect levels of environmental variation.

  5. eLife

    Reviewer #2 (Public Review):

    This manuscript represents the drawing together of a series of published results with new information, aiming to comprehensively address questions about the predictability of phenotypic evolution. The authors have data that may allow them to simultaneously investigate how genetic variation influences phenotypic evolution, and how that evolution, in turn, shapes genetic variation.

    I had several concerns about the suitability of the data for the questions posed, and about the robustness of the evidence presented. These concerns may be resolved via more detailed information about what was done and/or further analyses to provide robust support for conclusions. However, if not resolved, the data may be fundamentally unsuited to the purpose.

    Response to selection:

    It was not clear to me that it was appropriate to interpret locomotor behavior as having evolved in response to the salinity environment. Specifically, where is the evidence that any change in trait means is a (direct or indirect) response to selection imposed by increased salinity rather than the neutral drift of a trait due to the reduction in population size caused by the salinity? Strong evidence of adaptive evolution would be provided by all 3 replicates significantly diverging from the ancestor in the same direction. Model 2 seems to aim to test the null hypothesis that the three replicates diverged from one another via a random effects model - but with only three replicates, there is very low power, and variance is likely to be estimated as zero. I'm not sure what is shown in Tables 3 & 4, or how these results relate to models 2 & 3, so my interpretation of the information may be incorrect. Nonetheless, and noting that the errors around estimates are not presented, there seems to be considerable heterogeneity in size and direction of divergence between replicates for most of the traits. Is this study really dissecting responses to directional selection, or is it dissecting drift?

    What are the traits, and what is the confidence in G?
    My outsider's interpretation of these results is that defining 6 transition states is a way of getting at a single behavioral trait, and I was not convinced that these data were suitable for addressing questions about multivariate evolution. Genetic parameters were estimated using MCMCglmm, which imposes boundaries on estimates. The authors state that they followed Morrissey and Bonnet 2019, but I was unable to infer what this means with respect to accounting for the contribution of sampling error to covariances (or how they accounted for the positive variance constraint). Because I was unsure how sampling error was being assessed for G, I was not confident about the interpretation of statistical support for individual parameters, or for eigenvalues of G. Following this forward, if the measured characteristics constitute a single trait, with an entirely shared genetic basis, then the results of strong alignment of everything with gmax makes complete sense - there is a single trait, that is heritable and plastic, and for which the mean evolved.

  6. eLife

    Reviewer #3 (Public Review):

    Using experimental evolution with a nematode model system in a novel salt environment, Mallard et al. present a very nice experiment testing whether plasticity aligns with genetic variance and whether phenotypic divergence can be predicted from patterns of genetic variance. They find that although plasticity is not in the direction of genetic variance, estimates of selection that predict divergence are concordant with observed divergence. However, direct selection on a trait not included in the analysis is expected to be the underlying cause of phenotypic evolution. I commend the authors on their experiment and the framing of such a conceptually difficult topic.

    Strengths:

    Comparing a common ancestor to evolved populations to predict evolution has rarely been achieved, and the authors provide a strong test for predicting evolution in a novel environment.

    Weaknesses:

    Although a valuable dataset, the framing of the paper needs to be more focused on the question of predicting phenotypic evolution and comparing direct versus indirect selection. There are many details that are missing in the methods, which include the biological importance of the traits that are being studied, and how adaptation to the novel environment has occurred. In the discussion, they reveal that a measured trait that hasn't been included in the analyses is likely responsible for the observed patterns and it is unclear why this trait wasn't included in the analyses. In addition, direct versus indirect selection is not formally compared, which makes it difficult to interpret their results.

  7. Version published to 10.1101/2022.05.28.493855v1 on bioRxiv