Directional selection shifts trait distributions of planted species in dryland restoration

The match between species trait values and local abiotic filters can restrict community membership. An often‐implicit assumption of this relationship is that abiotic filters select for a single locally optimal strategy, though difficulty in isolating effects of the abiotic environment from those of dispersal limitation and biotic interactions has resulted in few empirical tests of this assumption. Similar constraints have made it difficult to assess whether the type and intensity of abiotic filters shift along gradients of environmental harshness, as predicted by the stress‐dominance hypothesis. We planted 9,216 plants of 29 perennial grass and forb species that had a range of functional trait values and were assigned to a warm, intermediate or cool temperature tolerance pool across eight sites on the Colorado Plateau. We compared the distributions of traits of surviving individuals to null distributions to evaluate whether there were shifts in trait means and variation. Borrowing from phenotypic selection concepts in evolutionary biology, we assessed support for stabilizing, directional and disruptive abiotic filtering of trait distributions and whether these types of filtering varied with initial species pool. Functional composition was significantly different from null distributions for nearly all traits at all sites, with trait variation more restricted in harsher abiotic conditions, supporting the stress‐dominance hypothesis. Contrary to expectations, we primarily found evidence for directional selection, which increased in frequency in warm species pools while disruptive selection was found more often in cool and intermediate species pools. Synthesis. This study provides a controlled experimental approach to test the effect of the abiotic environment on plant trait filtering. We found that opportunistic strategies allowing for rapid water acquisition during favourable periods improved survival at warmer sites. Species with these strategies may be expected to benefit from increasing aridity and may be selected for active management efforts. More generally, the prevalence of directional selection may have important implications for dynamic vegetation models that rely on trait distributions for translating environmental variation into ecosystem processes.


| INTRODUC TI ON
Numerous processes shape the distribution of biodiversity.
One of the most frequently invoked is environmental filtering (Butterfield, 2015;Cingolani et al., 2007;Cornwell et al., 2006;Craven et al., 2018;de Bello et al., 2013a;Le Bagousse-Pinguet et al., 2017;Lebrija-trejos et al., 2010;Li et al., 2018;Maire et al., 2012;Swenson et al., 2012), which is the effect of abiotic factors on the sorting of species into local communities based on their functional trait values (Keddy, 1992). Isolating filtering effects is complicated, however, due to potentially similar signatures of abiotic, biotic and dispersal processes on the distribution of functional traits within and among communities (Butterfield & Callaway, 2013;Cadotte & Tucker, 2017;Kraft et al., 2015;Mayfield & Levine, 2010;Nash et al., 2003). Distinctions between these filtering mechanisms are well-defined in theory (Funk et al., 2008), but in practice the ability to isolate the effect of a single filter in existing plant communities is very difficult (but see Fraaije et al., 2015). Experiments that manipulate filters are crucial for understanding their individual effects, and for predicting how communities may respond to shifts in abiotic conditions in the future.
One consequence of the challenge in differentiating among filtering mechanisms is that we can rarely quantify the degree to which the abiotic environment produces stabilizing, directional or disruptive selection on the local species pool. These terms, borrowed from evolutionary biology, have been used recently in the field of community ecology (Ackerly, 2003;Loranger et al., 2018;Muscarella & Uriarte, 2016;Rolhauser et al., 2019;Rolhauser & Pucheta, 2017) to represent shifts in functional trait distributions in the context of environmental filtering. These different forms of selection can have implications for the long-term growth, survival and reproduction of members of the community as they experience environmental changes. Stabilizing selection, also referred to as optimizing selection (sensu Travis, 1989), results in a unimodal trait distribution centring on 'optimal' trait values for a given set of conditions ( Figure 1). It is often assumed that unimodal patterns of trait distribution dominate in natural communities, though it has been shown that functional diversity is better accounted for when trait distributions are explicitly tested for rather than assumed (Laughlin et al., 2015). Trait distributions shaped by directional and disruptive selection move distributions from central towards extreme trait values. Disruptive selection can occur when there are highly distinct trait strategies that are viable in a single locale-this phenomenon is often discussed in the context of niche partitioning where coexistence is maintained by temporal or spatial separation in resource uptake (Rolhauser et al., 2019). Directional selection is not commonly discussed in the context of plant communities, though some evidence suggests it can be caused by successional changes in a community (Houssard & Escarré, 1991;Mori et al., 2017) or the pressures of environmental 'mismatch' (Ackerly, 2003;Enquist et al., 2017;Kooyers, 2015;Tielbörger et al., 2014). The structure of existing trait distributions is likely maintained by biotic interactions in a community; however, the extent to which abiotic factors alone can produce these different patterns of trait abundance is unknown. Identifying the influence of the abiotic environment on trait distributions, and determining whether such effects are consistent or highly variable across multiple traits and environments is an important step towards more predictable trait-based models of community assembly (Funk et al., 2016).
Stabilizing, directional and disruptive selection can be detected by investigating mean trait values and trait variation found within a plant community (Loranger et al., 2018). While inspecting the shape of trait distributions is often approached with the four moments of distribution (Enquist et al., 2015;Gross et al., 2017), additional investigation of skewness and kurtosis is often redundant with these measures of mean and variation (Loranger et al., 2018).
While other elements of functional diversity can be estimated using indices such as functional richness, evenness and divergence (Carmona et al., 2016), mean trait values and trait variation are good predictors of environmental selection (Denelle et al., 2019). Using theoretical trait distributions, we show that significant restriction in variation but no change in mean values reflects stabilizing selection; restriction in variation and significant shift in mean values reflect directional selection; greater variation but no change in trait mean reflects disruptive selection; and greater variation along with a significant shift in mean values reflect a combination of directional and disruptive selection (Figure 1).
Since most observational studies cannot separate the effects of biotic and abiotic factors, it is a challenge to address hypotheses about the relative contribution of each filtering process. The stressdominance hypothesis (Coyle et al., 2014;Swenson & Enquist, 2007) predicts that environmental (abiotic) filtering will play a greater role in structuring communities in more stressful environments, while competition (biotic filtering) is more important when there is lower abiotic stress. According to the stress-dominance hypothesis, we might expect increasing stabilizing selection in more severe environments (Rolhauser & Pucheta, 2017;Weiher & Keddy, 1995a) where trait variation is restricted due to abiotic filtering on traits critical for stress tolerance resulting in convergence on optimal trait values (Coyle et al., 2014). The 'favourability hypothesis' (Swenson et al., 2012;Wieczynski et al., 2019) also predicts a similar phenomenon where environmental filtering is stronger when conditions for growth are less favourable. Most studies within arid regions focus on biotic interactions such as facilitation (Berdugo et al., 2019); however, there is little investigation of how individual plants respond to abiotic conditions. Aridity gradients related to soil moisture and/ or atmospheric demand represent strong environmental variation across space and time. Within arid systems, the abiotic environment represents a particularly strong filter due to the overwhelming role of physical processes on plant-available soil moisture in water-limited environments (Goldberg & Novoplansky, 1997). While there is variation by region, the hotter, drier conditions forecasted for many global drylands (Bradford et al., 2020;Koutroulis, 2019) make it increasingly important to develop predictive models of traitbased community assembly in dryland regions. Drylands occupy 47% of the globe and often experience a disproportionate amount of anthropogenic disturbance (Koutroulis, 2019) which further increases abiotic stress. These dryland areas are therefore an excellent and important study system to test the stress-dominance hypothesis and to investigate how abiotic factors shape trait distributions.
Ecological restoration experiments provide an opportunity to isolate the signature of environmental filters on community assembly, and reciprocally a knowledge of trait-based environmental filtering can better inform restoration practice . The need for restoration is brought on by fire, overgrazing, invasive species and other anthropogenic disturbances (e.g. mining, recreation) that often erase the signature of the previous community, both in terms of seed sources (dispersal filters) and biotic feedbacks (biotic filters). The early stages of restoration experiments, in which the unfiltered species pool is known and initial plant densities are low, therefore provide an opportunity to identify a nearly pure abiotic filter in determining the subsequent composition of new plant communities. Quantifying the impacts of these environmental filters on trait distributions is not only of theoretical interest, but may also reflect an increasingly common context for assisted migration into disturbed environments .
In this study, we conducted a dryland restoration experiment replicated across a climatic gradient ('RestoreNet'; Havrilla et al., 2020) to identify the outcomes of abiotic filtering on community assembly, with an experimentally controlled dispersal filter and minimal biotic filter. In contrast to most observational studies of environmental filtering that compare species and traits along an abiotic gradient such as elevation (Alexander et al., 2011;Read et al., 2014), aridity (Dwyer & Laughlin, 2017;Nunes et al., 2017), light (Lusk & Laughlin, 2017) or multiple gradients (de Bello et al., 2013b;Le Bagousse-Pinguet et al., 2017;Menezes et al., 2020), to a regional species pool, our study compares traits of surviving species to those that were planted in an early stage of restoration with minimal biotic interaction. The unfiltered species pools in our study are defined by 16 species that were planted at each site based on temperature tolerances, thus the initial pool of available species influences the outcome of filtering.
Our objectives were to first test the hypothesis that non-random patterns of functional composition are present in the experimental communities, meaning that environmental filtering is indeed occurring. Second, we investigated the hypothesis that these non-random patterns vary as a function of species pool, trait or their interaction, meaning that shifts in trait means and variances could be predicted by trait or by the initial unfiltered pool of planted species and its suitability for local conditions. We explored the alternative outcomes of stabilizing, directional and disruptive selection found within restored communities.
Previous observational studies in similar environments have found overall more restrictive assembly filters in both very cold and very hot dryland environments (Butterfield & Munson, 2016), and drylands with primarily cool-season rather than warm-season precipitation regimes (Butterfield, 2015). Thus, we predicted that stabilizing selection would be more common and intense at sites with the most stressful conditions (within the warm species pool), and directional or disruptive selection would occur more frequently at sites with favourable conditions (within cool and/or intermediate species pools).

| Species pool and trait screening
We tested the filtering of plant functional traits using a grassland restoration experiment on the Colorado Plateau, a high-elevation semi-arid region in the western United States. Twenty-nine species of perennial grasses and forbs native to the region were selected F I G U R E 1 Conceptual diagram of theoretical selection types. To investigate the patterns of trait abundance, the standard effect size (SES) of functional composition indices, coefficient of variation (CoV) and community-weighted mean (CWM), was evalulated on communities to determine if they displayed stabilizing (n-shaped), directional or disruptive (u-shaped) selection. Stabilizing selection within a community reflects a lower CoV than expected (less than −1.96) and CWM is unchanged, directional selection reflects lower CoV than expected and CWM is significantly different from null expectation, while disruptive selection reflects greater sesCoV than expected and sesCWM is unchanged. The instance of greater sesCoV than expected in addition to a significant shift of CWM in either direction reflects a combination of directional and disruptive selection. Each hypothetical distribution contains the same area under the curve representing the number of surviving individuals in a community based on their prevalence in seed mixes used in restoration actions carried out by the Bureau of Land Management and US Forest Service on the Colorado Plateau and adjacent ecoregions (https:// wri.utah.gov/wri).
Ten replicates of each species were grown for a destructive trait screening at the Research Greenhouse at Northern Arizona University in the summer of 2017. Plants were grown to vegetative maturity in 3.79-L pots with a fired clay growing medium (Turface ® Pro League ® ) that facilitated oxygenation of the root systems of these arid-adapted plants while also permitting complete removal of the medium for root system analysis. Each plant was harvested prior to flowering to collect plant trait data (Table 1). Height data were collected prior to harvesting plants. Five fully developed leaves were then taken from each plant and immediately weighed, scanned on a flatbed scanner at 300 dpi and dried to calculate SLA and leaf dry matter content (LDMC). Roots were cleaned and samples of tap roots (forbs only), lignified coarse roots and absorptive fine roots (non-lignified with root hairs) were collected. Three coarse root samples from each plant were weighed, then scanned with a flatbed scanner to calculate root length and then dried to obtain specific root length (SRL) and root dry matter content (RDMC). The fine root samples were weighed, then scanned submerged in water using the WinRHIZO TM scanner bed software for analysis of root length and finally dried to obtain SRL and RDMC. Above-ground biomass was separated from below-ground biomass and dried separately to calculate root-to-shoot ratio.

| Experimental design
The field experiment was conducted at eight restoration sites located on the Colorado Plateau ( Figure S1; Table 2) that are a part of a broader RestoreNet study (Havrilla et al., 2020). RestoreNet systematically tests multiple restoration techniques using standardized protocols across drylands of the southwestern United States and is coordinated by the Restoration Assessment and Monitoring Program for the Southwest (https://usgs.gov/sbsc/ramps). Restoration sites were selected within common potential vegetation zones across the region where overgrazing or other disturbances have reduced perennial vegetation. The sites have a range in mean annual temperature of 8.8-16.5°C, mean annual precipitation of 206-491 mm and 32%-45% of yearly precipitation comes in the summer (July-September) monsoon season (Table 2). Initial perennial vegetation cover was low at all sites, though non-native weeds were common.   Multivariate analysis of variance revealed that the species pools did not differ in their average trait values (see Appendix S1); however, we retain the initial unfiltered 'species pool' as a factor that may influence the outcome of our study to account for any differences in the variance or range.  (6%-24%), while the range for an average year is 32%-45%.

| Data analysis
To test for environmental filtering, null models were employed to de-  To test the hypothesis that filtering produces non-random patterns of functional trait composition, we calculated the proportion of SES values greater than |1.96| for CoV and CWM.
Then, after determination that non-random patterns were present, we investigated the hypothesis that these non-random patterns vary as a function of species pool, trait or their interaction by employing linear mixed-effects models using the 'lmer' function in the lme4 package with site as a random factor. We used the lmerTesT package to obtain Type III Analysis of Variance  (Tables S1 and S2).
Finally, we assigned a type of selection (stabilizing, directional or disruptive) for each trait and site using the sign and magnitude of standardized effect sizes for CoV and CWM. Stabilizing selection occurred in communities with significant restriction in variation but no change in CWM; directional selection occurred in communities with restriction in variation and significant shift in CWM; and disruptive selection occurred when there was greater variation but no change in CWM. The applicability of CoV and CWM to testing the alternative assembly outcomes was confirmed with application to several idealized trait distributions (Figure 1). We acknowledge that a range of distributions that are intermediate between these idealized distributions can be produced within a similar selection category, but these categories based on CWM and CoV help to simplify the interpretation of a complex set of processes.

| RE SULTS
The average total survival of restoration plantings ranged from 41% to 89%, with cool and intermediate temperature sites generally performing better than warm sites (

| H2. Environmental filtering varies as a function of species pool, trait or their interaction
We expected that non-random patterns in trait distributions would depend on the initial unfiltered species pools that were planted as well as individual functional traits of those plants. A large absolute value for sesCoV would illustrate that variation within the surviving plants was highly different from that of the planted community, helping to identify where trait variation can be targeted to improve restoration outcomes. A large absolute value for sesCWM would identify where mean trait values of the species used in restoration can be adjusted in the future restoration efforts. We found that trait and species pool did significantly interact to influence both sesCoV and sesCWM (Table 4) (Table S4; Figure 2A).
We see in Figure 2A that variation in other traits also appears to be moderately influenced by species pool with a similar trend for lower variation found within the warmer species pools for height and SRLc. Only one instance of stabilizing selection was found (for SRL within one site that received a cool species pool). There were 15 instances (23%) of a selection pattern that is a combination of disruptive and directional where sesCoV > 1.96 and sesCWM > |1.96|. Finally, there were 19 instances of no selection occurring (30%).

| DISCUSS ION
Our experiment sought to better understand community assembly by isolating the abiotic filter acting on trait distributions of planted perennial grass and forb species. Two growing seasons after planting TA B L E 4 Predictive models for SES coefficient of variation (CoV) and community-weighted mean (CWM variance and stronger shifts in CWMs for sites that had the warmest and driest conditions, meaning that filtering was more intense in these stressful environments. Contrary to expectations, however, we found that the type of environmental filtering in these arid conditions could predominantly be associated with directional, not stabilizing, selection as we had hypothesized. The greater intensity of environmental filtering found in the warmest environments supported the stress-dominance hypothesis, but also emphasized the importance of gradient length and environmental context. For many years, there has been a discussion on the relative impact of environmental filtering in stressful environments versus competition in productive environments (Cornwell & Ackerly, 2009;Grime, 1977;May et al., 2013;Weiher & Keddy, 1995); however, the gradient upon which this idea is tested can influence the outcome.
For example, Swenson & Enquist (2007) (Grime, 2006;MacArthur & Levins, 1967), or environmental heterogeneity in space or time (Cavender-Bares et al., 2009;Lhotsky et al., 2016;Weiher & Keddy, 1995b), or other frequencydependent interactions. In this study, environmental heterogeneity is restricted to temporal or soil moisture availability at depth; however, an alternative explanation is that there is lower restriction of trait variability in favourable conditions. One instance that aligns with this explanation is the pattern of higher CoV found in root-to-shoot ratio at one of the intermediate sites, Flying M Ranch (Table S4). This site received the highest precipitation during the study period, so it could be that species with deeper roots are accessing soil moisture at depth and other species are taking advantage of surface soil moisture. This conclusion tends to agree with the literature that more soil resources allow for higher trait divergence (Bernard-Verdier et al., 2012;Wang et al., 2018 We applied the rules outlined in Figure 1 to determine whether sesCoV and sesCWM predicted stabilizing, direction or disruptive selection. The combination of directional and disruptive selection was determined when both sesCoV and |sesCWM| were higher than expected stabilizing selection, we found that these warmer sites most often produced patterns of directional selection. The intensity of heat and aridity at all sites was likely exacerbated by the high heat absorption of bare ground, often found at this early stage of restoration. The effect of aridity is often buffered by existing vegetation (Berdugo et al., 2018;Michalet, 2006), therefore an experiment conducted during early plant establishment can reveal the true abiotic filter in these arid environments, demonstrating that restoration in highly disturbed environments has an even greater barrier to overcome than can be demonstrated in observational studies of existing communities. A further examination of the individual traits driving these alternative selection outcomes helps solidify the argument that sup-  , though additional understanding of the traits that influence biotic interactions should also be included (Funk et al., 2008). Continuing to follow the current experiment as plants get larger and begin to interact with one another may shed light on this issue. Regardless, the trait-based approach presented here can help managers to select seed sources based on their trait values without conducting extensive species-by-species trials. A greater understanding of species trait responses, rather than using existing species occurrence, can improve restoration efforts, especially in topographically diverse regions such as the southwestern United States that can be difficult to model (Butler et al., 2017). These traitenvironment models may become increasingly relevant as climate conditions may preclude restoration of sites to historical conditions. As extreme climate events become more common, we are seeing that these extremes can additionally shift functional composition away from historical values (Griffin-Nolan et al., 2019).
Restoration practitioners may need to consider assisted migration and other 'prestoration' strategies  to match species to climate projections for their sites in the future and using species with traits that may benefit survival in disturbed environments (Ferguson & Nowak, 2012). Our study suggests certain trait values such as high LDMC and low RDMC may be beneficial broadly for restoration treatments on the Colorado Plateau; however, there are many opportunities to select species with better suited traits to match site conditions (e.g. higher SRL at our warm, dry sites). The ability to predict the outcome of seeding and transplanting efforts is highly valuable to the field of restoration ecology, especially as land managers and other restoration practitioners plan for a warmer future.

ACK N OWLED G EM ENTS
This research was supported by the BLM Colorado Plateau Native

Plant Program and the USGS Restoration Assessment and Monitoring
Program for the Southwest within the Ecosystems Mission Area. The authors thank Katherine Laushman for site installation and monitoring support. Any use of trade, firm or product names is for descriptive purposes only and does not imply endorsement by the US Government.

CO N FLI C T O F I NTE R E S T
None of the authors of this article have a conflict of interest.

PEER R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1111/1365-2745.13816.

DATA AVA I L A B I L I T Y S TAT E M E N T
Associated data are available on the TRY Plant Trait Database at