Connecting genetic incompatibilities with natural selection on additive genetic variation during adaptive radiation

Evolutionary biologists have long sought to identify the links between micro and macroevolution to better understand how biodiversity is created. Despite this pursuit, it remains a challenge to understand how allele frequency changes correlate with the evolution of morphological diversity, and the build-up of reproductive isolation amongst taxa. To connect mechanisms of microevolution with patterns of diversification, we tested the adaptive importance of alleles underlying genetic incompatibilities, and the consequences for predicting evolutionary trajectories of multiple ecotypes of an Australian wildflower. Using a quantitative genetics crossing design, we produced an F4 generation Advanced Recombinant Form (ARF) between four contrasting ecotypes, which we phenotyped in the glasshouse (N=770) and transplanted into the four natural habitats (N=14,265 seeds), alongside the parental ecotypes. F2 hybrid breakdown was associated with the loss of extreme phenotypes and habitat-specific genetic variation in field performance. Genetic trade-offs existed among habitats, but only in axes describing smaller amounts 27 of genetic variance for fitness. Habitats that showed stronger patterns of adaptive divergence for native 28 versus foreign ecotypes, also showed lower genetic variance in fitness of the ARF. Integrating data from the field and glasshouse predicted patterns of selection on morphological traits in a similar direction to the 30 parental ecotypes. Overall, our results provide strong empirical evidence linking ecotype specific alleles with phenotypic divergence, fitness trade-offs, rapid adaptation and the accumulation of genetic incompatibilities among recently derived ecotypes. Our data connects microevolutionary change with macroevolution through adaptive radiation, where selection for environment specific alleles creates rapid adaptive divergence leading to speciation.


Implementation of Bayesian models
In the subsequent analyses we implemented Bayesian models to 1) compare field performance 1 8 1 (experiment 2) of the ARF with the parental ecotypes, 2) identify whether genotype-by-environment between morphological traits (experiment 1) and field performance (experiment 2), to identify differences 1 8 5 in natural selection on morphological traits, among habitats. implemented Markov chains of different lengths (listed in Table S2), while ensuring that we included a 1 9 0 sufficient burn-in period and thinning interval to sample the parameters with autocorrelation values of 1 9 1 less than 0.05 and effective sample sizes exceeding 85% of the total number of samples, for all 1 9 2 parameters. We used uninformative parameter expanded priors and checked their sensitivity by re-1 9 3 All rights reserved. No reuse allowed without permission.
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint . http://dx.doi.org/10.1101/520809 doi: bioRxiv preprint first posted online Jan. 16, 2019; implementing all models while adjusting the parameters and ensuring the posterior distribution did not For the analyses estimating genetic variance, comparing estimates of genetic variance with zero provides 1 9 6 an uninformative test of significance because estimates are restricted to be greater than zero (positive- re-implemented the same model on 1,000 randomizations of the data, and extracted the posterior mean for 2 0 0 each randomization. We then compared the distribution of means from models conducted on the implemented on each randomization of the data, we could reduce computing time by reducing the total 2 0 6 number of sampling iterations. To do so, we maintained the same burn-in period and sampling interval to number required to obtain a stable estimate of the mean. We calculated the number of sampling iterations 2 0 9 required using the models implemented on the observed data, which was different for each of the analyses 2 1 0 outlined below (Table S2). All rights reserved. No reuse allowed without permission.
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint . http://dx.doi.org/10.1101/520809 doi: bioRxiv preprint first posted online Jan. 16, 2019; 1 0 and the ARF we estimate D, the variance-covariance matrix representing multivariate phenotypic 2 2 0 divergence. To do so, we first conducted another MANOVA that included all ecotypes (but not the ARF).

1
From this, we extracted the sums of squares and cross-product matrices for the ecotypes (SSCP H ) and 2 2 2 error terms (SSCP E ) to calculate their mean-square matrices by dividing by the appropriate degrees of variation. To visualize the phenotypic space occupied by the ARF relative to the parental ecotypes, we 2 2 8 decomposed D into orthogonal axes (eigenvectors) and calculated the phenotype scores for the first two eigenvectors for all ecotypes, and the ARF. Comparing ARF and ecotype field performance We estimated fitness at early life history stages for the ARF and parental ecotypes transplanted into all 2 3 2 four habitats. To do so, we created a dummy variable that represented the ARF and native versus foreign 2 3 3 ecotypes in each habitat. We then used MCMCglmm to implement the model, All rights reserved. No reuse allowed without permission.
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint . http://dx.doi.org/10.1101/520809 doi: bioRxiv preprint first posted online Jan. 16, 2019; multivariate response to selection. Estimating the response to selection in this way includes both direct To estimate the predicted response to selection (s g ) in the ARF we estimated the (co)variance between the 3 0 7 four morphology traits and field performance by implementing for field performance as the ability to reach maturity in each of the four transplant habitats. We calculated 3 2 0 the additive genetic (co)variance matrix as four times the sire variance component and extracted s g as the 3 2 1 vector of covariances between morphological traits and field performance (rows one to four of the fifth 3 2 2 column).

2 3
To identify whether we captured biologically meaningful differences in selection among habitats, we from the random distribution, we took this as evidence we detected biologically meaningful estimates of differences in s g among the four transplant habitats we estimated variance is selection gradients using 3 3 0 where Z then represents the among-habitat variance in s g for the nth trait along the diagonal. The off-

A) Seedling establishment B) Maturity
All rights reserved. No reuse allowed without permission.
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint . http://dx.doi.org/10.1101/520809 doi: bioRxiv preprint first posted online Jan. 16, 2019; Table S4. genetic variance in the randomized matrices described by the observed eigenvectors. Only the first three eigenvectors for 4 0 4 maturity described more genetic variance than expected by random sampling. Credible intervals represent 95% HPD intervals.  Overall, our results showed strong patterns of adaptive divergence (Figure 2), and although there appears 4 0 9 to be a common genetic basis to fitness in all environments (e 1 ; Table 1) we also detected genetic trade-4 1 0 offs for fitness among certain habitats (Table 1). Despite strong adaptive divergence in Figure 2, the weaker genetic trade-offs with other environments (Table 1), when compared to the dune and woodland. (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint . http://dx.doi.org/10.1101/520809 doi: bioRxiv preprint first posted online Jan. 16, 2019; 0 adaptive genetic variation that was lost in the ARF, reducing genetic variance for field performance in 4 1 5 certain environments and producing weaker genetic trade-offs than expected. To test this, for each habitat 4 1 6 we compared the strength of adaptive divergence (Figure 2; native ecotype performance -foreign ecotype 4 1 7 performance) against the level of genetic variance exhibited by the ARF. As predicted, we found a strong 4 1 8 negative association for seedling establishment and a weaker negative association for maturity ( Figure 5), 4 1 9 suggesting alleles associated with strong adaptive divergence were also responsible for genetic each MCMC iteration showed a significant negative association at 88% HPD for seedling establishment, but a non-significant 4 2 7 relationship for maturity.

2 8
Natural selection on morphological traits 4 2 9 To quantify selection in each habitat we calculated s g as the genetic covariance between morphological 4 3 0 traits measured in the glasshouse, and field performance measured in each of the four transplant habitats.

3 1
We then isolated direct selection by calculating ࢼ , the genetic selection gradient for each habitat. (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint . http://dx.doi.org/10.1101/520809 doi: bioRxiv preprint first posted online Jan. 16, 2019; ecotype morphology. representing the orientation of B and Z in relation to D, and d max . Here, we have used ecotype-specific genetic variation to connect adaptation and speciation during 4 5 7 adaptive radiation. We found that an ARF exhibited a multivariate phenotype intermediate to the four 4 5 8 parental ecotypes, but was lacking in much of the phenotypic variation of the parental ecotypes. Genetic The strength of divergence (and consequently, reproductive isolation) among coadapted gene complexes 5 0 1 will be population and environment specific, and depend on the interaction between mutation, migration, same species will give rise to reproductive isolation remains unexplored, but could provide important 5 0 9 insights into the relationship between adaptation and divergence across a heterogeneous landscape. All rights reserved. No reuse allowed without permission.
(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint . http://dx.doi.org/10.1101/520809 doi: bioRxiv preprint first posted online Jan. 16, 2019; of adaptive radiation, adaptation will be constrained to follow g max (Lande and Arnold 1983; Arnold 5 3 9 1992; Schluter 1996). As environment specific adaptive alleles increase in frequency, g max alters to align 5 4 0 with the phenotypic optimum and evolution is determined by the long-term correlated response to 5 4 1 selection (Zeng 1988). Thus, adaptive radiation occurs when environment-specific alleles increase in 5 4 2 frequency, causing changes in the distribution of genetic variance and ameliorates genetic constraints as 5 4 3 adaptive divergence proceeds.

4 4
In conclusion, we identified patterns of phenotypic and adaptive divergence among recently derived 5 4 5 ecotypes, created by the accumulation of environment-specific alleles in response to natural selection. We show that these alleles likely created ecotype-specific adaptive phenotypes and fitness trade-offs between 5 4 7 habitats that also lead to genetic incompatibilities between divergent ecotypes and reduced genetic identify the connection between microevolutionary genetic changes and macroevolutionary