Diverse Genotype-by-Weather Interactions in Switchgrass

Reviewer #1 (Public review):

Summary:

The authors present results and analysis of an experiment studying the genetic architecture of phenology in two geographically and genetically distinct populations of switchgrass when grown in 8 common gardens spanning a wide range of latitudes. They focused primarily on two measures of phenology - the green-up date in the spring, and the date of flowering. They observed generally positive correlations of flowering date across the latitudinal gradient, but negative correlations between northern and southern (i.e. Texas) green-up dates. They use GWAS and multivariate meta-analysis methods to identify and study candidate genetic loci controlling these traits and how their effect sizes vary across these gardens. They conclude that much of the genetic architecture is garden-specific, but find some evidence for photoperiod and rainfall effects on the locus effect sizes.

Strengths:

The strengths of the study are in the large scale and quality of the field trials, the observation of negative correlations among genotypes across the latitudinal gradient, and the importance of the central questions: Can we predict how genetic architecture will change when populations are moved to new environments? Can we breed for more/less sensitivity to environmental cues?

Weaknesses:

I have tried hard to understand the concept of the GxWeather analysis presented here, but still do not see how it tests for interactions between weather and genetic effects on phenology. I may just not understand it correctly, but if so, then I think more clarity in the logical model would help - maybe a figure explaining how this approach can detect genotype-weather interactions. Also, since this is a proposal for a new approach to detecting gene-environment effects, simulations would be useful to show power and false positive rates, or other ways of validating the results. The QTL validation provided is not very convincing because the same trials and the same ways of calculating weather values are used again, so it's not really independent validation, plus the QTL intervals are so large overlap between QTL and GWAS is not very strong evidence.

The term "GxWeather" is never directly defined, but based on its pairing with "GxE" on page 5, I assumed it means an interaction between genotypes (either plant lines or genotypes at SNPs) and weather variables, such that different genotypes alter phenology differently as a response to a specific change in weather. For example, some genotypes might initiate green-up once daylengths reach 12 hours, but others require 14 hours. Alternatively (equivalently), an SNP might have an effect on greenup at 12 hours (among plants that are otherwise physiologically ready to trigger greenup on March 21, only those with a genotype trigger), while no effect on greenup with daylengths of 14 hours (e.g., if plants aren't physiologically ready to greenup until June when daylengths are beyond 14 hours, both aa and AA genotypes will greenup at the same time, assuming this locus doesn't affect physiological maturity).

Either way, GxE and (I assume) GxWeather are typically tested in one of two ways. Either genotype effects are compared among environments (which differ in their mean value for weather variables) and GxWeather would be inferred if environments with similar weather have similar genotype effects. Or a model is fit with an environmental (maybe weather?) variable as a covariate and the genotype:environment interaction is measured as a change of slope between genotypes. Basically, the former uses effect size estimates across environments that differ in mean for weather, while the latter uses variation in weather within an experiment to find GxWeather effects.

However, the analytical approach here seems to combine these in a non-intuitive way and I don't think it can discover the desired patterns. As I understand from the methods, weather-related variables are first extracted for each genotype in each trial based on their green-up or flowering date, so within each trial each genotype "sees" a different value for this weather variable. For example, "daylength 14 days before green-up" is used as a weather variable. The correlation between these extracted genotype-specific weather variables across the 8 trials is then measured and used as a candidate mixture component for the among-trial covariance in mash. The weight assigned to these weather-related covariance matrices is then interpreted as evidence of genotype-by-weather interactions. However, the correlation among genotypes between these weather variables does not measure the similarity in the weather itself across trials. Daylengths at green-up are very different in MO than SD, but the correlation in this variable among genotypes is high. Basically, the correlation/covariance statistic is mean-centered in each trial, so it loses information about the mean differences among trials. Instead, the covariance statistic focuses on the within-trial variation in weather. But the SNP effects are not estimated using this within-trial variation, they're main effects of the SNP averaged over the within-trial weather variation. Thus it is not clear to me that the interpretation of these mash weights is valid. I could see mash used to compare GxWeather effects modeled in each trial (using the 2nd GxE approach above), but that would be a different analysis. As is, mash is used to compare SNP main effects across trials, so it seems to me this comparison should be based on the average weather differences among trials.

A further issue with this analysis is that the weather variables don't take into account the sequence of weather events. If one genotype flowers after the 1st rain event and the second flowers after the 2nd rain event, they can get the same value for the cumulative rainfall 7d variable, but the lack of response after the 1st rain event is the key diagnostic for GxWeather. There's also the issue of circularity. Since weather values are defined based on observed phenology dates, they're effectively caused by the phenology dates. So then asking if they are associated with phenology is a bit circular. Also, it takes a couple of weeks after flowering is triggered developmentally before flowers open, so the < 2-week lags don't really make developmental sense.

Thus, I don't think this sentence in the abstract is a valid interpretation of the analysis: "in the Gulf subpopulation, 65% of genetic effects on the timing of vegetative growth covary with day length 14 days prior to green-up date, and 33% of genetic effects on the timing of flowering covary with cumulative rainfall in the week prior to flowering". There's nothing in this analysis that compares the genetic effects under 12h days to genetic effects under 14h days (as an example), or genetic effects with no rainfall prior to flowering to genetic effects with high rainfall prior to flowering. I think the only valid conclusion is: "65% of SNPs for green-up have a GxE pattern that mirrors the similarity in relationships between green-up and day length among trials." However I don't know how to interpret that statement in terms of the overall goals of the paper.

Next, I am confused about the framing in the abstract and the introduction of the GxE within and between subpopulations. The statement: "the key expectation that different genetic subpopulations, and even different genomic regions, have likely evolved distinct patterns of GxE" needs justification or clarification. The response to an environmental factor (ie plasticity) is a trait that can evolve between populations. This happens through the changing frequencies of alleles that cause different responses. But this doesn't necessarily mean that patterns of GxE are changing. GxE is the variance in plasticity. When traits are polygenic, population means can change a lot with little change in variance within each population. Most local adaptation literature is focused on changes in mean trait values or mean plasticities between populations, not changes in the variance of trait values or plasticities within populations. Focusing on the goal of this paper, differences in environmental or weather responses between the populations are interesting (Figure 1). However the comparisons of GxE between populations and with the combined population are hard to interpret. GxE within a population means that that population is not fixed for this component of plasticity, meaning that it likely hasn't been strongly locally selected. Doesn't this mean that in the context of comparing the two populations, loci with GxE within populations are less interesting than loci fixed for different values between populations? Also, if there is GxE in the Gulf population, by definition it is also present in the "Both" population. Not finding it there is just a power issue. If individuals in the two subpopulations never cross, the variance across the "Both" population isn't relevant in nature, it's an artificial construct of this experimental design. I wonder if there is confusion about the term "genetic" in GxE and as used in the first paragraph of the intro ("Genetic responses" and "Genetic sensitivity"). These sentences would be most clear if the "genetic" term referred to the mechanistic actions of gene products. But the rest of the paper is about genetic variation, ie the different effects of different alleles at a locus. I don't think this latter definition is what these first uses intend, which is confusing.

Note that the cited paper (26) is not relevant to this discussion about GxE patterns. This paper discusses the precision of estimating sub-group-specific genetic effects. With respect to the current paper, reference 26 shows that you might get more accurate measures of the SNP effects in the Gulf population using the full "Both" population dataset because i) the sample size is larger, and ii) as long as the true effects are not that different between populations. That paper is not focused on whether effect size variation is caused by evolution but on the technical question of whether GxG or GxE impacts the precision of within-group effect size estimates. The implication of paper 26 is that comparing SNP effects estimated in the "Both" population among gardens might be more powerful for detecting GxE than using only Gulf samples, even if there is some difference in SNP effects among populations. But if there magnitudes (or directions) of SNP effects change a lot among populations (ie not just changes in allele frequency), then modeling the populations separately will be more accurate.

Reviewer #2 (Public review):

The provided evidence in the study by MacQueen and colleagues is convincing, albeit some methodological challenges still exist. The authors rightly state that different subpopulations are likely to have evolved distinct patterns of GxE. It has been recently shown that the genetic architecture for adaptive traits differs across subpopulations (Lopez-Arboleda et al. 2021), hence this effect should be even more pronounced for GxE. How to best account for this in a statistical framework is not utterly clear. Here the authors describe their efforts to asses these interactions and to estimate the magnitude of the respective effects. Building on the statistical framework described, it could be possible to translate their findings from switchgrass to other species. A plus of the study is the effort to use an independent pseudo-F2 population to confirm the found associations.
The manuscript is written coherently and all data and code used is freely available and explained in detail in the supplementary information.

Nevertheless, I feel that there are some points in the data analysis that could be clarified some more.

(1) Dividing GxE interactions into discrete, measurable GxWeather terms is a nice idea to gain a reliable measurement of E. I also appreciate the effort to create date-related values as a summary function of a weather variable across a specified date range. Using cumulative data the week prior to flowering seems like a good choice to associate weather patterns to this phenotype, but there are many - including non-linear ways - to accumulate these data. Additionally, weather parameters like temperature and precipitation can show interaction effects. I wonder if there is a way to consider these.

(2) As pointed out in Section S1, a trait measured in eight common gardens could be modeled at eight genetically correlated traits. To assess the genetic correlation one would need to estimate the genetic variance within each trait and 28 genetic covariance structures. Here model convergence would be painful given the sample sizes. There are different statistical solutions for this including the mash algorithm the authors choose. I highly appreciate the effort in how the rationale is described in the supplementary information, but to me, it is still not completely clear how 'strong' and random effects have been selected from GWAS. How sensitive is the model to a selection of different effects? Could one run permutations to assess this? Why is the number of total markers different for different phenotypes and subsets and does this affect statistical power?

(3) The mash model chooses different covariance matrices for the different analyses. Although I do understand the rationale for this, I am not sure how this will impact the respective analysis and how comparable the results are. Would one not like to have the same covariance matrices selected for all analyses?

(4) Although the observed pattern of different GxE in different subpopulations is intriguing, it remains a little unclear what we actually learn apart from the fact that GxE in adaptive traits is complex. Figure 3 divides GxE into sign and magnitude effects. Interestingly the partition differs significantly between Greenup date and Flowering Date. Still, the respective QTLs in Figure 4 do - at least partially - overlap (e.g. on CHR05N). What is the interpretation of these? Here, I would appreciate a more detailed discussion and hearing the thoughts of the authors.

(5) Figure 4 states that Stars indicate QTLs with significant enrichment for SNPs in the 1% mash tail. The shown Rug plots indicate this, but unfortunately, I am missing the respective stars. Is there a way to identify what is underlying these QTLs?

To summarize, the manuscript nicely shows the complex nature of GxE in different switchgrass subpopulations. The goal now would be to identify the causative alleles for these phenomena and understand how these have evolved. Here the provided study paves the way for further analyses in this perspective.