Including genetic and environmental variation improves model
performance
The inclusion of temperature predictors improved model performance
relative to genotype-only models for both traits. Genetic variation
alone explained only a small proportion of phenotypic variation and
could be captured in a GSM. This supports our hypothesis that GxE alters
the genetic architecture of traits across environments and renders
individual markers less informative (Brachi et al., 2010; El-Soda et
al., 2014; Fournier-Level et al., 2011; Linde et al., 2006) which is
consistent with results from provenance testing in different tree
species (Benito Garzón et al., 2019). The method used to compute
pairwise genetic similarity did not affect model performance (Table S2
& S3, Appendix S2), suggesting they produced functionally identical
descriptions of genetic similarity. Eu-ahsunthornwattana and colleagues
(2014) previously reported high correlation between different genetic
similarity estimates in humans, suggesting that our framework is likely
to be broadly applicable.
Phenotypic
variation was mainly explained by temperature differences between the
plantings. This was expected: climate-responsive traits are by
definition affected by environmental conditions, and the influence of
temperature on plant phenotypes is well-established (Anderson et al.,
2012; Arft et al., 1999; Foden et al., 2007; Schwartz & Hanes, 2010;
Sun et al., 2020; Zhao et al., 2017). The better performance of models
predicting DTB than SP is likely caused by the higher heritability of
phenological traits over seed traits (Fournier-Level et al. 2013; Gnan
et al., 2014). Moreover, the photothermal model of A. thalianaflowering time (Chew et al., 2012) suggests a mostly-linear relationship
between flowering phenology and temperature that was effectively
captured by our model. Still, internal validation results for SP are
comparable to those for yield traits reported in non-model species
(Deomano et al., 2020; Velazco et al., 2019).
Within the existing literature, our framework positions itself alongside
similar models designed to predict climate response. This includes
frameworks like ΔTraitSDMs
(Benito Garzón et al., 2019) that link functional traits to species
distributions and landscape genomic models measuring genomic offset
based on associations between genetic and environmental variation
(Capblancq et al., 2020; Gougherty et al., 2021; Supple et al., 2018).
In comparison to these methods, our approach distinguishes itself on
three main points.
Firstly, we focus on predicting quantitative traits. This provides a
more straightforward measure of prediction accuracy and interpretation
of model results. In the absence of trait data, mismatch between present
and future conditions has been quantified using metrics like
FST (Gougherty et al., 2021) and habitat suitability
(Benito Garzón et al., 2019). These metrics cannot be estimated for
individuals and do not provide actionable targets that land managers can
directly manipulate or select for, contrasting the simplicity of using
functional traits.
Secondly, we use fine-scale time series data instead of low-resolution
environmental predictors such as the Bioclim variables (c.f. Gougherty
et al., 2021 and Supple et al., 2018; Fick & Hijmans, 2017). This
functional approach defines conditions as experienced by plants
throughout their growing period, rather than through summary climate
variables that condense years of weather data into a single statistic.
This is necessary because A. thaliana plants can occupy the same
geographical site but experience very different environments due to
variation in germination time (Donohue et al., 2005). Predictors based
on monthly, quarterly, or yearly averages cannot account for the
multiple seasonal cohorts germinating in a single location. Moreover,
long-term averages cannot account for the effects of climate change on
temperature variability (Bathiany et al., 2018; Schär et al., 2004;
Screen, 2014) and the distinct responses of plants to changes in mean
temperature and temperature variability (Burghardt et al., 2016;
Scheepens et al., 2018; Wheeler et al., 2000). Experimental studies have
typically used a consistent increase in temperature to simulate climate
change (Fournier-Level et al., 2016; Li et al., 2014; E. S. Post et al.,
2008; Sherry et al., 2007; Springate & Kover, 2014) while maintaining
current patterns of variability (Springate & Kover, 2014), but this may
not reflect actual patterns of climate change. By considering both the
daily temperature range and temperature variation between days, our
predictions may better match trait values seen in natural populations.
This is particularly relevant because revegetation will introduce plants
to uncontrolled conditions.
Thirdly, we account for GxE (analogous to phenotypic plasticity;
Ghalambor et al., 2007) in a way that facilitates out-of-sample
predictions. GxE is the immediate and potentially adaptive response of
organisms to environmental change (Ghalambor et al., 2007). Trait models
typically consider GxE on the basis of genotypic and/or environmental
identity (Montesinos-López et al., 2018;
de Oliveira et al.,
2020). In contrast, our ancestry-based approach allows for continuous
GxE, allowing for estimation of phenotypic plasticity even in novel
genotypes. We rely on the assumption that shared neutral ancestry
results in similar climate responses, which may not be true if GxE
depends on only a few important variants. However, environmental
blocking results indicate our approach can be applied to traits with
high heritability in the absence of knowledge regarding specific trait
architecture, which lends itself well to application in poorly
characerised species.
The consistent focus on transferability to novel conditions results in a
model devoid of identity-based descriptors. Both genetic and
environmental variation are described continuously, using predictors
that can be straightforwardly estimated for any novel
environment-genotype combination. This approach does have a tradeoff in
terms of multicollinearity: we relied on multiple non-independent
predictors to describe both genetic and environmental variation. Thus,
the final model contain multiple correlated predictors and does not lend
itself to biological interpretation. However, our validation results
show this does not impact its ability to predict the climate response of
multiple genotypes and identify those suitable for revegetation.