FACE facts hold for multiple generations; Evidence from natural CO2 springs (Global Change Biol)
<![CDATA[By Jasmine M. Saban, Mark A. Chapman and Gail Taylor
Rising atmospheric CO2 concentration is a key driver of enhanced global greening, thought to account for up to 70% of increased global vegetation in recent decades. CO2 fertilization effects have further profound implications for ecosystems, food security and biosphere‐atmosphere feedbacks. However, it is also possible that current trends will not continue, due to ecosystem level constraints and as plants acclimate to future CO2 concentrations. Future predictions of plant response to rising [CO2] are often validated using single‐generation short‐term FACE (Free Air CO2 Enrichment) experiments but whether this accurately represents vegetation response over decades is unclear. The role of transgenerational plasticity and adaptation in the multigenerational response has yet to be elucidated. Here, we propose that naturally occurring high CO2 springs provide a proxy to quantify the multigenerational and long‐term impacts of rising [CO2] in herbaceous and woody species respectively, such that plasticity, transgenerational effects and genetic adaptation can be quantified together in these systems. In this first meta‐analysis of responses to elevated [CO2] at natural CO2 springs, we show that the magnitude and direction of change in eight of nine functional plant traits are consistent between spring and FACE experiments. We found increased photosynthesis (49.8% in spring experiments, comparable to 32.1% in FACE experiments) and leaf starch (58.6% spring, 84.3% FACE), decreased stomatal conductance (gs, 27.2% spring, 21.1% FACE), leaf nitrogen content (6.3% spring, 13.3% FACE) and Specific Leaf Area (SLA, 9.7% spring, 6.0% FACE). These findings not only validate the use of these sites for studying multigenerational plant response to elevated [CO2], but additionally suggest that long‐term positive photosynthetic response to rising [CO2] are likely to continue as predicted by single‐generation exposure FACE experiments.
1 INTRODUCTIONAverage atmospheric global [CO2] is now consistently above 400 ppm for the first time in around 23 million years of evolutionary time (Pearson & Palmer, 2000). Increased atmospheric [CO2] will be a key feature of future climates, and although there is clear resolve to cap atmospheric [CO2] to below 530 ppm in order to avoid catastrophic ecosystem change under global warming, it remains unclear whether these [CO2] targets will be met (Stocker, 2013). Despite the profound impact of [CO2] on plant functioning, future predictions of plant responses to elevated [CO2] are predominantly validated using experimental data derived from single‐generation experiments, which model only plant phenotypic plasticity. These plastic responses have been extensively quantified in experimental systems ranging from small controlled environment studies to large ecosystem experiments using FACE, and generalized through meta‐analyses that are used to inform or validate models and predictions (Ainsworth & Long, 2005; Dybzinski, Farrior, & Pacala, 2015; Vanuytrecht & Thorburn, 2017). While these experiments have played a pivotal role in informing short‐term projections of, for example, food security (Myers et al., 2014; Wheeler & Von Braun, 2013) and the likely distribution of plant ecotones in a changing climate (Barnaby & Ziska, 2012; Forkel et al., 2016; Smith et al., 2016), extrapolating to predict consequences of climate change for the end of the century may be precarious. Beyond single‐generation plastic plant responses to elevated [CO2] there is some evidence for adaptation (the inheritance of derived characteristics that enhance fitness in a given environment) but a lack of conclusive evidence that elevated [CO2] could act as a selective agent on either genetic or epigenetic variation under climate change in the natural environment (Frenck, Linden, Mikkelsen, Brix, & Jørgensen, 2013; Leakey & Lau, 2012; Ward, Antonovics, Thomas, & Strain, 2000). Regardless, there is a wealth of evidence to suggest that transgenerational effects can and do contribute to plant response to elevated [CO2] over multiple generations (Jablonski, Wang, & Curtis, 2002; Johnston & Reekie, 2008; Springer & Ward, 2007). Multigenerational experiments are a key challenge for the study of plant adaptation, owing to the time, energy and expense of growing plants under such conditions long‐term, especially for long‐lived and large plant species. Facilities are expensive and labour intensive to build and maintain, and cannot provide information on population responses to elevated [CO2] over generations in the timeframe needed to prepare for climate change. To this end, plants surrounding natural CO2 springs are a precious resource to further elucidate evolutionary adaptation and long‐term response to elevated [CO2]. Plants growing at natural CO2 springs have previously been utilized to study physiological response to rising [CO2] but have largely been abandoned due to concerns about CO2 emission variability over time and contamination by other exhaust gases. Here, we propose that as with other systems, provided these limitations are appropriately managed, spring sites represent a valuable resource that can contribute to our understanding of multigenerational plant response to elevated [CO2] in combination with other systems. In this first meta‐analysis of natural CO2 spring plant response to elevated [CO2], we highlight sites at which research has been conducted and synthesize available data, comparing responses to those in FACE experiments.
2 MATERIALS AND METHODS
2.1 Systematic searchTo evaluate research at CO2 springs, we captured available data through a systematic search of the literature on 3rd July 2017. Using a structured string search and standard systematic review methodology, 3,294 studies were collated from Web of Science and screened according to strict inclusion criteria to provide a database of studies measuring traits in plants at natural CO2 springs compared to an ecologically similar control site in close proximity. These inclusion criteria are outlined in Supporting Information Appendix S1 and include (among others) that there must be a difference in [CO2] of at least 100 ppm between spring and control sites, and that sites are only included where contamination by [H2S] < 0.02 ppm and [SO2] < 0.015 ppm, as detailed in Supporting Information Table S1. To avoid non‐independence as a result of multiple measurements of a trait being reported in a single publication, only one data point was taken for a trait for each species in each study. The data point extracted was decided on a trait by trait basis, for example photosynthetic measurements were taken at midday and during summer months if they were measured multiple times. In order to calculate effect sizes, mean, sample size and standard deviation were obtained from the text, tables or extracted from figures using DATATHIEF (Tummers, 2006). Authors were contacted if there was insufficient data reported for inclusion in the meta‐analysis and many authors kindly provided additional data. Ultimately, we analysed data from 16 sufficiently replicated traits across 39 species in 25 papers (Supporting Information Appendix S2 and Table S1). This represents a subset of studies that have ever been used to study plant response at natural CO2 springs because we were unable to include traits (and therefore studies) where fewer than five species or studies measured the trait across the database.
2.2 Statistical analysis
2.2.1 Effect size calculationTo compare trait differences between spring (elevated [CO2]) and control (ambient [CO2]) groups, we calculated the log response ratio (lnR) for each trait under elevated [CO2] as a metric for analysis. Log response ratio quantifies the proportional difference in population mean for a trait under elevated [CO2] at the spring site relative to ambient [CO2] at the control site. The log transformation is used to linearize the relationship between the two variables and to obtain residuals that are approximately symmetrically distributed where the sampling distribution may otherwise be skewed (particularly in small samples) (Hedges, Gurevitch, & Curtis, 1999). Log response ratio was calculated as:
where is the mean trait value for plants growing under elevated [CO2] at the spring site and is the mean trait value for plants growing in ambient [CO2] at the control site. For more intuitive presentation, the log response ratio is converted to percentage difference using the formula [(R‐1) × 100]. All statistical analyses were performed in R version 3.2.2 (R‐CORE‐TEAM, 2015).
2.2.2 Meta‐analysisA random effects model was applied to calculate overall effect of elevated [CO2] on populations at the spring site relative to the control populations. Random effects models were used to account for environmental variation by assuming that true effect size varies between studies forming a distribution of effect sizes. The studies within the analysis are assumed to be a random sample of this distribution and the overall summary effect of a random effects model estimates the mean of the distribution of true effect sizes. The null hypothesis is that the mean of the distribution of effects is zero. The effect size of each species from each study was weighted using the inverse of its variance. All models used restricted maximum likelihood estimation. If a 95% confidence interval for a trait did not overlap zero then a significant response was considered in plants exposed to elevated [CO2] relative to their ambient counterparts at control sites.
2.2.3 Assessing heterogeneity between studiesWe examined variation between studies, partitioning it from within study error using the heterogeneity statistic Q and subsequently I2 using the formula I2 = 100% × (Q‐df)/Q (Higgins & Thompson, 2002). The I2 statistic describes the percentage of variation across studies, that is due to heterogeneity rather than chance. Of the sixteen traits that were measured, the Q and I2 statistics indicated that thirteen traits showed a significant degree of between‐study heterogeneity and effect sizes were calculated using a random effects model to account for this (Supporting Information Table S2). For three traits (Vcmax, Jmax and leaf carbon: nitrogen ratio), we found Q with p > 0.05 and/or an I2 statistic <50% suggesting the variation in findings is compatible with chance alone (homogeneity) and therefore a fixed effect model was used to calculate these effect sizes. Significant heterogeneity between studies existed for all traits analysed, suggesting that almost all of the variability in estimates was due to variation between samples rather than sampling error. This is common among ecological studies where an average I2 of 83%–92% were reported in an analysis of ecological meta‐analyses (Senior et al., 2016). Given that individual samples come from a diverse array of global sites and from multiple functional groups, this heterogeneity is to be expected, but it is also useful to explore the basis of this heterogeneity by modelling potential moderator variables. Subgroup analysis was performed to examine trait changes in functional groups where sample size permitted (as trees, including both deciduous and evergreen trees, and herbs, including grasses, with forbs also analysed separately for stomatal conductance for comparison to FACE analyses), and a random effects meta‐regression model with defined moderator variables was fitted to the data to examine the effect of these moderator variables in the R package glmulti (Calcagno & De Mazancourt, 2010). Plant functional group and climate zone were used as moderator variables for meta‐regression analysis. For categorical variables, the category was considered an important predictor if the 95% confidence intervals of the category estimate did not overlap those of the overall effect size. Photosynthetic rate at growth [CO2] was the only trait where either of these categorical predictors were considered significant in predicting the estimate under meta‐regression. For this trait, we further decomposed the categorical variable “climate zone” to two continuous variables; average maximum daily temperature and annual precipitation for meta‐regression. Variance explained by a predictor variable was calculated through ANOVA of the model containing only this predictor variable versus the null model.
2.2.4 Publication biasIn ecological studies, there may be a bias towards publishing positive and significant results, and studies with larger sample size have more power to detect significant differences, indeed Haworth, Hoshika, and Killi (2016) have suggested that publication bias has resulted in a significant over‐estimation of the impacts of elevated [CO2] on plants in FACE study meta‐analyses. Publication bias was quantified using weighted regression with multiplicative dispersion using standard error as the predictor to detect funnel plot asymmetry (the classical Egger’s test), using the regtest function in the METAFOR package (Viechtbauer, 2010), by examining plots of the data and by estimating the fail‐safe number (Supporting Information Table S3; Rosenberg, 2005). From analyses of these tests and examination of the normal Q‐Q and funnel plots, we acknowledge that publication bias and the presence of outliers reduce confidence in the model estimates of summary effect for adaxial stomatal density, leaf chlorophyll content and leaf carbon content. Our interpretation of these results is duly cautious. We additionally performed sensitivity analysis by applying weight functions to the effect sizes of studies to determine the impact of moderate publication bias. Assuming moderate selection of publication bias on the gathered dataset, we estimate that effect sizes in this study may be inflated by 6%–13%. This is similar in magnitude to the estimated inflation of FACE study effect sizes by 5%–15% due to moderate reporting bias (Haworth et al., 2016).
3 RESULTSA systematic search of the literature revealed CO2 springs that have previously been utilized for this research occur extensively across the globe and range in latitude, temperature and rainfall (Figure 1). Significant differences in vegetation types and species present at each site are apparent, including many long‐lived tree species that are difficult to work with experimentally. The most comprehensively studied and characterized springs are located in Italy and Japan (Figure 1, Supporting Information Table S1)
|Meta‐analysis||Experimental designs analysed||Average [CO2] of elevated treatments (ppm)|
|J. Saban, M.A. Chapman, and G. Taylor, (unpublished data)||Natural CO2 springs||791|
|Ainsworth and Long (2005||FACE||~560|
|Ainsworth and Rogers (2007||FACE||567|
|Wang et al. (2012||Semi‐open and closed systems||702|
|Semi‐open and closed systems||732|