Hybridization alters the shape of the genotypic fitness landscape, increasing access to novel fitness peaks during adaptive radiation
Curation statements for this article:-
Curated by eLife
Evaluation Summary:
This study reports on the inference of the evolutionary trajectory of two specialist species that evolved from one generalist species. The process of speciation is explained as an adaptive process and the changing genetic architecture of the process is analyzed in great detail. The genomic dataset is big and the inference from it solid. The authors reach the conclusion that introgression and de novo mutations, but not standing genetic variation, are the main players in this adaptive process.
(This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. The reviewers remained anonymous to the authors.)
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
- Evolutionary Biology (eLife)
Abstract
Estimating the complex relationship between fitness and genotype or phenotype (i.e. the adaptive landscape) is one of the central goals of evolutionary biology. However, adaptive walks connecting genotypes to organismal fitness, speciation, and novel ecological niches are still poorly understood and processes for surmounting fitness valleys remain controversial. One outstanding system for addressing these connections is a recent adaptive radiation of ecologically and morphologically novel pupfishes (a generalist, molluscivore, and scale-eater) endemic to San Salvador Island, Bahamas. We leveraged whole-genome sequencing of 139 hybrids from two independent field fitness experiments to identify the genomic basis of fitness, estimate genotypic fitness networks, and measure the accessibility of adaptive walks on the fitness landscape. We identified 132 single nucleotide polymorphisms (SNPs) that were significantly associated with fitness in field enclosures. Six out of the 13 regions most strongly associated with fitness contained differentially expressed genes and fixed SNPs between trophic specialists; one gene ( mettl21e ) was also misexpressed in lab-reared hybrids, suggesting a potential intrinsic genetic incompatibility. We then constructed genotypic fitness networks from adaptive alleles and show that scale-eating specialists are the most isolated of the three species on these networks. Intriguingly, introgressed and de novo variants reduced fitness landscape ruggedness as compared to standing variation, increasing the accessibility of genotypic fitness paths from generalist to specialists. Our results suggest that adaptive introgression and de novo mutations alter the shape of the fitness landscape, providing key connections in adaptive walks circumventing fitness valleys and triggering the evolution of novelty during adaptive radiation.
Article activity feed
-
-
-
Author Response:
Reviewer #1:
This study reports on the inference of the evolutionary trajectory of two specialist species that evolved from one generalist species. The process of speciation is explained as an adaptive process and the changing genetic architecture of the process is analyzed in great detail. The genomic dataset is big and the inference from it solid. The authors reach the conclusion that introgression and de novo mutations, but not standing genetic variation, are the main players in this adaptive process.
I would avoid the term adaptive radiation for the group of fish studied here. It is misleading. It is generally accepted to use the term adaptive radiation when a fairly large number of new species originate from a common ancestor (cichlids in big African lakes, gammarids in Lake Baikal, etc). Here are only 2 new …
Author Response:
Reviewer #1:
This study reports on the inference of the evolutionary trajectory of two specialist species that evolved from one generalist species. The process of speciation is explained as an adaptive process and the changing genetic architecture of the process is analyzed in great detail. The genomic dataset is big and the inference from it solid. The authors reach the conclusion that introgression and de novo mutations, but not standing genetic variation, are the main players in this adaptive process.
I would avoid the term adaptive radiation for the group of fish studied here. It is misleading. It is generally accepted to use the term adaptive radiation when a fairly large number of new species originate from a common ancestor (cichlids in big African lakes, gammarids in Lake Baikal, etc). Here are only 2 new lines that evolved from a common ancestor. Furthermore, I do not see much parallel between the ideas and concepts used when people study real adaptive radiations and one studied here. I actually believe that the term adaptive radiation even distracts from the beauty of the current study.
We would like to acknowledge that the usage of the term “adaptive radiation” has a long, rich history of debate in the literature over how it should be applied to empirical systems. Some example definitions of adaptive radiation are listed below:
“The evolution of ecological and phenotypic diversity within a rapidly multiplying lineage” - Schluter, 2001 (The ecology of adaptive radiation). This definition implies that abundant ecological and morphological diversity that arose in a single lineage over a short time are the hallmarks of adaptive radiation and has been frequently applied to stickleback species pairs. The pupfishes of San Salvador Island meet these criteria (two trophic specialists arose from a generalist ancestor within 10,000 years). Importantly, please note that in this foundational textbook on adaptive radiation, no statement is made about the number of species necessary to be considered an adaptive radiation.
“The evolutionary divergence of members of a clade to adapt to the environment in a variety of different ways.” – Losos, 2009 (Lizards in an evolutionary tree: Ecology and adaptive radiation of Anoles). Here again, the pupfish system described meets the definition. Unlike the previous definition, no statement about the rate of diversification (species or morphological/ecological) is made.
“The rise of a diversity of ecological roles and attendant adaptations in different species within a lineage” – Givnish, 1997 (Adaptive plant evolution on islands: classical patterns, molecular data, new insights. Evolution on islands). As with the previous definition, no qualification is made with respect to rates of diversification. The pupfishes again meet the definition.
As discussed by Givnish in 2015 (“Adaptive radiation versus ‘radiation’ and ‘explosive diversification’: why conceptual distinctions are fundamental to understanding evolution” – New Phytologist), few of the early definitions of adaptive radiations contained any reference to the rapidity of speciation – Simpson (1953) perhaps being the only notable exception. However, despite this, no definition states that the application of “adaptive radiation” to a given system is contingent upon a given number of species having arisen by the present day.
The pupfishes of Salvador island meet all definitions of adaptive radiation – exceptional rates of morphological diversification and ecological diversification, as well as truly exceptional rates of speciation – focusing just on the three species here, two species have arisen within the last 10,000 years – this roughly translates to a speciation rate of 200 species per million years. While this pace is highly unlikely to be maintained, we feel that every line of evidence points towards the pupfishes of San Salvador Island as an adaptive radiation at the earliest stages of the process. We disagree that an adaptive radiation must be ‘complete’ or nearly so, for it to be deemed as such.
Finally, we have also discovered a fourth pupfish species on the island (Richards and Martin 2016; Richards et al. 2021), and even more undiscovered species may exist there. Thus, this is an adaptive radiation of four sympatric species, not two as suggested.
The "Result and discussion" section has rather little discussion. There is not much about other systems or studies, neither in concepts nor in biology. The results are not linked to the bigger questions and the larger field. The same is true for the conclusion, which is very strongly centered on the here reported study. What can we learn from this study for other systems? Is there a generalizable take-home message? How do the findings relate to commonly held ideas/theory on how adaptive speciation works? Without this, it reads like a report of a case study, disconnected from the larger field. To achieve this aim, it may be good to split the main section into a result and a discussion section, but this is only a suggestion.
We followed this helpful suggestion and have split the results and discussion section and significantly expanded and revised our discussion section. We now relate our findings to the broader fitness landscape theory literature and emphasize how our findings inform the process of speciation. We conclude by emphasizing that our findings point to a process in which adaptive introgression and de novo mutation not only provide diversity that is useful in reaching novel fitness peaks on a static landscape but alter the shape of the landscape itself.
Reviewer #2:
This is a really interesting and challenging question the authors are addressing here. I enjoyed reading the manuscript and a few comments below:
One major concern I have concerns the analysis of the two treatments (low and high density, l411). I believe that the two treatments should analyzed separately as the authors are estimating two different fitness landscapes. When conducting their analysis, experiment is treated as a single factor. Yet, in Martin and Wainwrigth (2013), it was established that the fitness landscapes were quite different between the two treatments (Figure S7 of said paper), meaning that different phenotypes (and therefore genotypes) were affected differently. I do not think that the complex effect described there can be capture by a single factor as done here.
We examined this concern further and now include new analyses of only data from the second field experiment to address these concerns (described in more detail below), resulting in qualitatively similar conclusions to those conducted using all samples.
Please also note that only the high-density treatments from the 2013 study were included in the current study due to the low sample sizes of the original low-density treatments. In the 2020 fitness landscape study, we found no evidence of a treatment effect (frequency-manipulation) on the curvature of the fitness landscape. In all our analyses, we do include the effect of lake accounting for environmental differences between lake replicates.
While the two high-density treatments in Martin and Wainwright 2013 were analyzed and visualized in some cases as distinct adaptive landscapes as pointed out by the reviewer, many aspects of stabilizing and disruptive selection were comparable between the lake environments and detected in similar regions of morphospace as described in Table 1 in that paper. All statistical analyses of the second field experiment (e.g. Figure 5A of Martin & Gould 2020 Evol. Letters) indicated no effect of the frequency treatment between the two field enclosures in each lake; accounting for treatment did not improve model fit to the data. In the second field experiment, the authors found that the two frequency treatments in each lake could in fact be summarized by a single fitness landscape accounting for lake-specific effects which was as the best fitting GAM model. This surface bore remarkable similarities to the high-density fitness surfaces of the 2013 in the placement of fitness peaks and valleys on the morphospace (Martin and Gould 2020). Thus, we tend to view the fitness landscape of interest to us as a single landscape connecting the fitness of different species phenotypes while treating lake-specific environmental effects on this landscape as background noise.
Unfortunately, we do not have sufficient resequenced samples to analyze only data from the first experiment alone (Martin and Wainwright 2013); fewer than half of our samples come from the 2013 study – the remainder come from the second field experiment. Therefore, we now include a second set of analyses focused on just the subset of resequenced fish from the second field experiment (Figure 5—figure supplement 1-2, Appendix 1—table 18-19). Our primary goal was to assess whether our major findings held within a single field experiment by focusing on the latter, more data-rich experiment.
Because we believe the most significant analyses from our paper are those pertaining to genotypic fitness landscapes and accessibility, using the subset of data from the second field experiment we performed 1) analyses of models fit between ancestry proportion and fitness (i.e. Figure 1—figure supplement 3), and 2) analyses estimating accessibility between generalists and either trophic specialist (reported in Appendix 1—table 19).
Overall, we found qualitatively similar results between analyses conducted using either all samples or only those in the second experiment. As a result, we report results for all samples in the main text while referencing the analyses of the second field experiment alone which are presented in the supplementary material.
A second major concern I have is in the use of the Admixture software (Figure 1 and l152.) The generalist type is assumed to be the ancestral type. Yet, a unique group was not assigned to it. This is a known problem for Admixture (Lawson et al. 2018). Groups that are under-sampled are far more likely to be consider a mixture of different ancestry groups even when this is impossible (Rasmussen et al 2010, Skolung et al 2012). While this in itself is not problematic, I am concerned about the use the authors are making of these ancestry proportions (l 156-165). The authors analyzed how ancestry of scale eater or molluscivore affect survival probability, growth, or the hybrid composite fitness. However, the ancestries values are partly generated due to an artefact, so I wonder how modelling the ancestral type as a group, and therefore acknowledging some amount of share ancestry between the three species may further affect this analysis.
We agree that the ancestries estimated for the generalists by our unsupervised admixture analyses appear to be confounded and we briefly allude to this in the text. In our original submission, we focused exclusively on molluscivore and scale-eater ancestry, which appear less biased by this artifact. To address this concern, we ran new admixture analyses using a supervised analysis, a priori assigning generalists, molluscivores, and scale-eaters to one of three populations. Ancestry proportions of hybrids were then inferred for each of three clusters. We now include new analyses of fitness by ancestry associations using these admixture proportions and found qualitatively similar results. We report these new analyses in the results and supplemental material.
We also conducted analyses using only samples from the second field experiment (related to the first concern raised by the reviewer). In all, we now include the following analyses of the extent to which the three fitness measures are associated with each of the three ancestry proportions using:
- an unsupervised admixture analysis (Appendix 1—table 2),
- all samples using a supervised admixture analysis (i.e. model is informed a priori which samples are known to belong to either of the three assumed populations/parental species: Appendix 1—table 3),
- only samples from the second field experiment (Martin & Gould 2020) in which lake was not found to significantly affect fitness using an unsupervised analysis (Appendix 1—table 4).
Importantly, results are qualitatively the same; ancestry proportions do not strongly influence fitness in this system. There is one exception – generalist ancestry appears to positively predict growth when modeled using all samples and the supervised admixture analysis (Appendix 1—table 3). However, the inconsistency of this result across the three analyses leads us to cautiously interpret this exception
I understand the need to use subsets of a network, due to impossibly large dimension size of the network in the first place. However, subsetting said network may give the wrong impression of the whole network (Fragata et al 2019). I wish this point was further discussed here.
We have followed this suggestion. In our now-expanded and significantly revised discussion, we include discussion of this limitation, citing Fragata et al (2019) as well as related works. We also discuss how estimation of combinatorially complete fitness landscapes may be misleading, as their topography is determined in part by epistasis that occurs among loci that are not segregating in natural populations. We also suggest that the ‘realized epistasis’ that occurs among only those loci that are naturally segregating in a population may be why the shape of the fitness landscape, and thus accessibility of fitness peaks, changes upon the appearance of adaptive introgression and de novo mutations.
L 294-295: I wonder whether the results here could be used to discuss the geometry of the different fitness peaks. The small number of steps within molluscivores suggest a rather narrow peak, while the rather large ones within the generalist suggest a rather flat fitness peak. The shape of the peak can be linked to the amount of genetic variation that can be maintained within populations, as well as the mutational load of said populations.
This is an excellent suggestion and led us to consider the ruggedness of our fitness landscapes as an additional factor affecting evolutionary accessibility. We now interrogate the geometry of the fitness landscape further, asking for each specialist, how many local peaks exist on their respective landscapes (i.e. the ruggedness), how far specialists are from these peaks, and how accessible these peaks are to specialists. We elaborate on these findings in the discussion as recommended.
These expanded analyses further led us to similarly investigate the influence of each source of genetic variation on the ruggedness of the fitness landscape. Consequently, we now discuss in more detail the interplay between fitness landscape ruggedness and accessibility of interspecific genotypic paths, in the context of what sources of genetic variation are available. We show that the presence of adaptive introgression and de novo mutations both increase the accessibility of interspecific genotypic paths, while decreasing fitness landscape ruggedness. We now discuss how this finding makes sense in light of epistasis; changes to the pool of segregating genetic variation alters the ‘realized epistasis’ in natural populations, thus altering the shape of the fitness landscapes and ultimately the evolutionary outcomes favored by natural selection.
L74-75 I would suggest to more cautious in the phrasing here. While this is true within Fisher geometric model, where population are assumed monomorphic and infinite, this is not true in general. Deleterious mutations can fix within populations, especially when drift is non negligible. Crossing fitness valleys has been quite widely investigated (see Weissman et al 2010 for example). Even the authors themselves mention it later (l 108).
We tempered these statements as recommended and expand our references to include Weissman et al. 2010 and additional references describing these caveats.
Lastly, I would be more cautious about the conclusion. Line 373-374, the authors mentioned that "de novo mutations may enable the crossing of a large fitness valley". Given that the authors focus only on adaptive walk (fitness always has to increase between each mutational step), there is no crossing of fitness valleys. Switching from one fitness peak to another is simply a matter of walking along a (very) narrow ridge.
We revised our language as recommended, emphasizing that our results support an interpretation in which apparent phenotypic fitness valleys are crossed along narrow fitness ridges, which are not observed in a three dimensional morphospace, to reach new fitness optima.
Reviewer #3:
This paper uses sophisticated regression methods and numerical experiments to produce a genotype-fitness relationship for three closely related sympatric pupfish species, forming an adaptive radiation. In addition to providing insights into the genetic targets of selection, this paper goes further in attempting to tease out what types of genetic variation were most likely to have played key roles in this radiation.
Strengths:
The idea behind this study is excellent, and clearly a large amount of thought and effort went into collecting the underlying data. The attention paid to linking evolutionary dynamics with the fitness results is laudable. The system is extremely exciting and I think an experiment and analysis of this sort could potentially be interesting to a broad audience within evolutionary biology.
Weaknesses:
The claim that this is the first genotypic fitness network in a vertebrate needs additional qualifiers: as far as I can tell, the claim to novelty is based on the inclusion of multiple species, the number of alleles, and measuring fitness in the field. I can't fully assess this claim but I would urge the authors to avoid staking a stronger claim to priority than is really needed, as it might be a lightening rod for criticism and hair-splitting that would distract from the contents of the paper.
We tempered this claim as suggested, removing it from the title, and de-emphasizing or removing this claim elsewhere throughout the manuscript.
One of my major questions while reading this was whether these three species were better or worse adapted to subenvironments within the lakes. This is partially answered in a few places in the manuscript, but I think that resolving this point more precisely would help interpret if positioning all three species on the same fitness landscape is fair.
We have included more description/discussion of the ecological differences between species to the manuscript, particularly their habitats within the lake. We now point out that all three species coexist within the benthic littoral zone of each lake. No habitat segregation among these species has been observed in 13 years of field studies, suggesting that it is reasonable to position all three species within the same fitness landscape. Their foraging also occurs within the same benthic microhabitat throughout each lake; indeed, the scale-eaters target their generalist neighbors for scale attacks. This thinking also underlies much of the theory of speciation and adaptive radiation. We now include these qualifiers in the text as well.
I find it a little hard to follow the construction of the landscapes in Fig. 2 B and C. I am not clear why the landscapes don't cover the location of the molluscivore population.
We now include a brief statement that estimated values of fitness are only plotted for samples within the observed morphospace in the hybrids. That is, because none of the hybrid phenotypes were morphologically similar to the most divergent molluscivore phenotypes, we could not measure fitness values for this region of morphospace. However, there were hybrid phenotypes that fell within the 95% confidence ellipse of the lab-reared molluscivore population, suggesting that we have good power to detect adaptive walks to this region of the morphospace.
I think the fitnesses predicted for the main bulk of the generalists and scale-eaters are the same across the two landscapes (as I expect they would be), but this is obscured by the differing fitness ranges of the two landscapes. I would suggest using a single color-fitness relationship for the two panels to aid cross-comparison.
We re-plotted these landscapes using a uniform color scheme across panels as recommended.
Also, two salient features of the landscape-the major peak at the top center and the deep pit at the bottom center-seem to be supported by few fish in each case. I would imagine that something like boot-strapping could be done for fitness landscapes, where the support for each feature of the landscape could be judged by how often it appears in subsets of the data (or in inferred models with nearly as high support as the best model), but I acknowledge that might be very hard to do. Still, I think some statement of uncertainty should be prominently included.
We followed this suggestion and now more explicity address uncertainty in our estimation of three-dimensional fitness landscapes, with particular focus on the landscape we devote the most attention to (Fig. 2c-d – composite fitness + genotypes).
To quantify uncertainty, we conducted a bootstrap procedure as suggested in which we resampled hybrids with replacement, re-estimated the fitness landscape, and compared the topology of the predicted fitness landscapes to that of the observed fitness landscape (Figure 2—figure supplement 7). Even across the bootstrap replicates, we still recovered the same general features – a peak localized near generalists, a fitness valley near scale-eaters, and a fitness ridge/modest peak near molluscivores.
Furthermore, we emphasize more strongly in the revised manuscript our point that three-dimensional representations of the fitness landscape may in fact mislead interpretations of how evolution proceeds. In that respect, even though we recover the same features of the landscape when accounting for uncertainty, we articulate that these inferred peaks and valleys separating populations may be bridged in multidimensional genotype space.
More generally, the landscapes reconstructed in Fig. 2 do not show very clear evidence that the M or S types are separated by valleys from the G type. Close inspection of the figure suggests a very shallow valley might be present between G and M, but the overall trend is declining fitness; between G and S, fitness appears to simply decline. While peaks may occur within the landscapes composed of limited sets of loci, the overall pattern seen in Fig. 2 doesn't seem conducive to analyzing how adaptive evolution in generalists crossed valleys to reach the putatively higher peaks of the two specialists. As such, I find the connection between these phenotypic-fitness landscapes and the later genotypic fitness landscapes quite confusing.
We thank the reviewer for this comment. The apparent disconnect noted by the reviewer is in fact a point that we would like to draw more attention to. Thus, we have revised much of the discussion of these results to address this.
As discussed in our response to the reviewer’s previous comment, the three dimensional landscape contrasts with our inferences from genotypic fitness landscapes. This incongruence demonstrates, through example, how three-dimensional fitness landscapes may in fact mislead our intuition about how evolution proceeds.
As has been discussed extensively in the fitness landscape literature (e.g. Kaplan et al. 2008; Gavrilets 2010; Fragata et al. 2019), reduction of the fitness landscape, which is inherently highly multidimensional (as originally recognized by Wright), to only three dimensions can mask viable evolutionary trajectories, underestimate the number of peaks, and oversimplify our understanding of how populations evolve. We now attempt to better clarify and discuss this in the revised manuscript.
I also had trouble understanding the role of fitness in the analysis of mutational distances in a subset of loci between the three species (lines 282-296). While the illustration in Fig. 3C uses directed edges to capture fitness data, this framework doesn't seem to be applied in Fig. 3d or the resulting analyses in 3e. As such, I don't see how this section is about genotypic fitness landscapes at all.
We followed this suggestion and have rearranged our figures and their constituent panels to provide a more coherent illustration of our results and analyses. Figure 3 now serves to describe 1) the focal loci used to construct genotypic networks and 2) the general structure of genotypic networks constructed using loci sampled across all three species. What is now figure 4 is dedicated explicitly towards investigation of genotypic fitness landscapes, describing how we incorporated fitness measures into these networks to identify accessible path. This figure also serves to describe the fitness landscapes for each specialist, quantifying accessibility of interspecific genotypic trajectories, and landscape ruggedness. Our discussion of these sections similarly attempts to distinguish their respective focus, emphasizing that investigation of the general isolation of each species on genotypic networks will help provide context for our later focused investigation of fitness landscapes.
The final part of the conclusion sketches a story in which de novo and introgressed alleles reduce the accessibility of reverse evolution, back to a generalist. I think this is conceptually confusing because we don't expect evolution to favor paths toward lower fitness, even if those paths do not pass through a valley. Again, the framing here-that generalists are less fit than either specialist-is hard to square with the facts that generalists seem to be coexisting with the specialists, and much closer to the hypothesized fitness peak than is either specialist.
We agree and have completely rewritten this section and removed this framing. We omitted this part of the conclusion entirely, as we felt it too speculative, and as noted by the reviewer, difficult to square with some of the rest of our findings. Instead, we now devote more focus on other aspects and implications of our findings in a new discussion section as requested by reviewer 1.
This is a complicated and ambitious paper, on an exciting system and aiming at important questions. I think the main results about genotypic-fitness networks are hard to relate back to the other major analyses in the paper due to the points raised above. Moreover, using fitness measurements of three coexisting species to infer how they evolved faces a major obstacle: if fitnesses are frequency-dependent, then the actual trajectory of an initially rare variant will be completely obscured post-invasion. This possibility, as well as the potential issue that data on reproductive success might change these findings, need to be discussed, especially in light of the puzzling fact that the specialists appear less fit than their ancestor in at least one of the paper's major analyses.
We now emphasize the apparent disconnect between three-dimensional fitness landscapes and the highly dimensional genotypic fitness landscapes as noted by the reviewer (see above). We hope to demonstrate through example how highly dimensional genotypic fitness landscapes may harbor numerous viable evolutionary trajectories (e.g. fitness ridges) on rugged fitness landscapes that are unobservable on low-dimensional representations. Additionally, we expand our discussion of the caveats in our analyses pertaining to the use of data on contemporary species to infer historical dynamics on the fitness landscape as recommended by the reviewer.
We also now note that no evidence for frequency-dependent selection has been found in this system (Martin and Gould 2020; Martin 2016). We previously explicitly manipulated the frequency of rare phenotypes between treatments and found no effect of treatment across lake populations. Rather, these fitness peaks and valleys appear surprisingly stable across lakes, treatments, and years.
Regardless, we now include in the discussion that we necessarily have taken a ‘birds-eye view’ of evolution here, describing the influences of different sources of genetic variation on the fitness landscape, after these have already undergone selective sweeps. Likewise, we acknowledge that it is impossible to quantify reproductive success in this system using field enclosures due to the very small size of newly hatched fry and continuous egg-laying life history of pupfishes. This is a limitation of our system. We take this opportunity to emphasize that other experimental or simulation studies would be invaluable to quantify the changing influence of these different sources of genetic variation on the fitness landscape as a function of time, during the process of selective sweeps.
-
Evaluation Summary:
This study reports on the inference of the evolutionary trajectory of two specialist species that evolved from one generalist species. The process of speciation is explained as an adaptive process and the changing genetic architecture of the process is analyzed in great detail. The genomic dataset is big and the inference from it solid. The authors reach the conclusion that introgression and de novo mutations, but not standing genetic variation, are the main players in this adaptive process.
(This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. The reviewers remained anonymous to the authors.)
-
Reviewer #1 (Public Review):
This study reports on the inference of the evolutionary trajectory of two specialist species that evolved from one generalist species. The process of speciation is explained as an adaptive process and the changing genetic architecture of the process is analyzed in great detail. The genomic dataset is big and the inference from it solid. The authors reach the conclusion that introgression and de novo mutations, but not standing genetic variation, are the main players in this adaptive process.
I would avoid the term adaptive radiation for the group of fish studied here. It is misleading. It is generally accepted to use the term adaptive radiation when a fairly large number of new species originate from a common ancestor (cichlids in big African lakes, gammarids in Lake Baikal, etc). Here are only 2 new lines …
Reviewer #1 (Public Review):
This study reports on the inference of the evolutionary trajectory of two specialist species that evolved from one generalist species. The process of speciation is explained as an adaptive process and the changing genetic architecture of the process is analyzed in great detail. The genomic dataset is big and the inference from it solid. The authors reach the conclusion that introgression and de novo mutations, but not standing genetic variation, are the main players in this adaptive process.
I would avoid the term adaptive radiation for the group of fish studied here. It is misleading. It is generally accepted to use the term adaptive radiation when a fairly large number of new species originate from a common ancestor (cichlids in big African lakes, gammarids in Lake Baikal, etc). Here are only 2 new lines that evolved from a common ancestor. Furthermore, I do not see much parallel between the ideas and concepts used when people study real adaptive radiations and one studied here. I actually believe that the term adaptive radiation even distracts from the beauty of the current study.
The "Result and discussion" section has rather little discussion. There is not much about other systems or studies, neither in concepts nor in biology. The results are not linked to the bigger questions and the larger field. The same is true for the conclusion, which is very strongly centered on the here reported study. What can we learn from this study for other systems? Is there a generalizable take-home message? How do the findings relate to commonly held ideas/theory on how adaptive speciation works? Without this, it reads like a report of a case study, disconnected from the larger field. To achieve this aim, it may be good to split the main section into a result and a discussion section, but this is only a suggestion.
-
Reviewer #2 (Public Review):
This is a really interesting and challenging question the authors are addressing here. I enjoyed reading the manuscript and a few comments below:
One major concern I have concerns the analysis of the two treatments (low and high density, l411). I believe that the two treatments should analyzed separately as the authors are estimating two different fitness landscapes. When conducting their analysis, experiment is treated as a single factor. Yet, in Martin and Wainwrigth (2013), it was established that the fitness landscapes were quite different between the two treatments (Figure S7 of said paper), meaning that different phenotypes (and therefore genotypes) were affected differently. I do not think that the complex effect described there can be capture by a single factor as done here.
A second major concern I …
Reviewer #2 (Public Review):
This is a really interesting and challenging question the authors are addressing here. I enjoyed reading the manuscript and a few comments below:
One major concern I have concerns the analysis of the two treatments (low and high density, l411). I believe that the two treatments should analyzed separately as the authors are estimating two different fitness landscapes. When conducting their analysis, experiment is treated as a single factor. Yet, in Martin and Wainwrigth (2013), it was established that the fitness landscapes were quite different between the two treatments (Figure S7 of said paper), meaning that different phenotypes (and therefore genotypes) were affected differently. I do not think that the complex effect described there can be capture by a single factor as done here.
A second major concern I have is in the use of the Admixture software (Figure 1 and l152.) The generalist type is assumed to be the ancestral type. Yet, a unique group was not assigned to it. This is a known problem for Admixture (Lawson et al. 2018). Groups that are under-sampled are far more likely to be consider a mixture of different ancestry groups even when this is impossible (Rasmussen et al 2010, Skolung et al 2012). While this in itself is not problematic, I am concerned about the use the authors are making of these ancestry proportions (l 156-165). The authors analyzed how ancestry of scale eater or molluscivore affect survival probability, growth, or the hybrid composite fitness. However, the ancestries values are partly generated due to an artefact, so I wonder how modelling the ancestral type as a group, and therefore acknowledging some amount of share ancestry between the three species may further affect this analysis.
I understand the need to use subsets of a network, due to impossibly large dimension size of the network in the first place. However, subsetting said network may give the wrong impression of the whole network (Fragata et al 2019). I wish this point was further discussed here.L 294-295: I wonder whether the results here could be used to discuss the geometry of the different fitness peaks. The small number of steps within molluscivores suggest a rather narrow peak, while the rather large ones within the generalist suggest a rather flat fitness peak. The shape of the peak can be linked to the amount of genetic variation that can be maintained within populations, as well as the mutational load of said populations.
L74-75 I would suggest to more cautious in the phrasing here. While this is true within Fisher geometric model, where population are assumed monomorphic and infinite, this is not true in general. Deleterious mutations can fix within populations, especially when drift is non negligible. Crossing fitness valleys has been quite widely investigated (see Weissman et al 2010 for example). Even the authors themselves mention it later (l 108).
Lastly, I would be more cautious about the conclusion. Line 373-374, the authors mentioned that "de novo mutations may enable the crossing of a large fitness valley". Given that the authors focus only on adaptive walk (fitness always has to increase between each mutational step), there is no crossing of fitness valleys. Switching from one fitness peak to another is simply a matter of walking along a (very) narrow ridge.
-
Reviewer #3 (Public Review):
This paper uses sophisticated regression methods and numerical experiments to produce a genotype-fitness relationship for three closely related sympatric pupfish species, forming an adaptive radiation. In addition to providing insights into the genetic targets of selection, this paper goes further in attempting to tease out what types of genetic variation were most likely to have played key roles in this radiation.
Strengths:
The idea behind this study is excellent, and clearly a large amount of thought and effort went into collecting the underlying data. The attention paid to linking evolutionary dynamics with the fitness results is laudable. The system is extremely exciting and I think an experiment and analysis of this sort could potentially be interesting to a broad audience within evolutionary biology.
W…
Reviewer #3 (Public Review):
This paper uses sophisticated regression methods and numerical experiments to produce a genotype-fitness relationship for three closely related sympatric pupfish species, forming an adaptive radiation. In addition to providing insights into the genetic targets of selection, this paper goes further in attempting to tease out what types of genetic variation were most likely to have played key roles in this radiation.
Strengths:
The idea behind this study is excellent, and clearly a large amount of thought and effort went into collecting the underlying data. The attention paid to linking evolutionary dynamics with the fitness results is laudable. The system is extremely exciting and I think an experiment and analysis of this sort could potentially be interesting to a broad audience within evolutionary biology.
Weaknesses:
The claim that this is the first genotypic fitness network in a vertebrate needs additional qualifiers: as far as I can tell, the claim to novelty is based on the inclusion of multiple species, the number of alleles, and measuring fitness in the field. I can't fully assess this claim but I would urge the authors to avoid staking a stronger claim to priority than is really needed, as it might be a lightening rod for criticism and hair-splitting that would distract from the contents of the paper.
One of my major questions while reading this was whether these three species were better or worse adapted to subenvironments within the lakes. This is partially answered in a few places in the manuscript, but I think that resolving this point more precisely would help interpret if positioning all three species on the same fitness landscape is fair.
I find it a little hard to follow the construction of the landscapes in Fig. 2 B and C. I am not clear why the landscapes don't cover the location of the molluscivore population. I think the fitnesses predicted for the main bulk of the generalists and scale-eaters are the same across the two landscapes (as I expect they would be), but this is obscured by the differing fitness ranges of the two landscapes. I would suggest using a single color-fitness relationship for the two panels to aid cross-comparison. Also, two salient features of the landscape-the major peak at the top center and the deep pit at the bottom center-seem to be supported by few fish in each case. I would imagine that something like boot-strapping could be done for fitness landscapes, where the support for each feature of the landscape could be judged by how often it appears in subsets of the data (or in inferred models with nearly as high support as the best model), but I acknowledge that might be very hard to do. Still, I think some statement of uncertainty should be prominently included.
More generally, the landscapes reconstructed in Fig. 2 do not show very clear evidence that the M or S types are separated by valleys from the G type. Close inspection of the figure suggests a very shallow valley might be present between G and M, but the overall trend is declining fitness; between G and S, fitness appears to simply decline. While peaks may occur within the landscapes composed of limited sets of loci, the overall pattern seen in Fig. 2 doesn't seem conducive to analyzing how adaptive evolution in generalists crossed valleys to reach the putatively higher peaks of the two specialists. As such, I find the connection between these phenotypic-fitness landscapes and the later genotypic fitness landscapes quite confusing.
I also had trouble understanding the role of fitness in the analysis of mutational distances in a subset of loci between the three species (lines 282-296). While the illustration in Fig. 3C uses directed edges to capture fitness data, this framework doesn't seem to be applied in Fig. 3d or the resulting analyses in 3e. As such, I don't see how this section is about genotypic fitness landscapes at all.
The final part of the conclusion sketches a story in which de novo and introgressed alleles reduce the accessibility of reverse evolution, back to a generalist. I think this is conceptually confusing because we don't expect evolution to favor paths toward lower fitness, even if those paths do not pass through a valley. Again, the framing here-that generalists are less fit than either specialist-is hard to square with the facts that generalists seem to be coexisting with the specialists, and much closer to the hypothesized fitness peak than is either specialist.
This is a complicated and ambitious paper, on an exciting system and aiming at important questions. I think the main results about genotypic-fitness networks are hard to relate back to the other major analyses in the paper due to the points raised above. Moreover, using fitness measurements of three coexisting species to infer how they evolved faces a major obstacle: if fitnesses are frequency-dependent, then the actual trajectory of an initially rare variant will be completely obscured post-invasion. This possibility, as well as the potential issue that data on reproductive success might change these findings, need to be discussed, especially in light of the puzzling fact that the specialists appear less fit than their ancestor in at least one of the paper's major analyses.
-
-