Statistical analysis
All analyses were performed using R version 4.0.2 (R Core Team 2022).
Germination analysis
Seed germination across treatments was compared using generalised linear mixed effects models with a binomial link function with the R package lme4 (Bates et al. 2015). Soil source, sterilisation and seed mass were used as fixed effects and pot ID was included as a random effect. Seed mass was included as a fixed effect to account for its potential effect on germination.
Plant functional trait analysis
To compare the differences in total biomass, aboveground biomass, belowground biomass, and root-mass fractions of T. triandra , we used randomised linear mixed-effects models. Across our models, we included soil source, sterilisation and water stress as fixed effects, with interaction terms in different combinations, and random effects to account for within-group variation (for model details, see Table S2). Significance was determined by permuting each model 10,000 times and comparing observed test-statistics with those of the simulated random distributions.
We assessed plant-soil feedback (PSF) ratios for each plant trait across the different aridity soils and water stress treatment groups. For each treatment group, we calculated the average plant response under live and sterilised conditions, using the following formula, wherex˝̄ represents average plant biomass from the live or sterile treatment groups:
\begin{equation} PSF\ ratio=\frac{\left(x\ Live-\text{x\ Sterile}\right)}{\text{X\ Sterile}}\nonumber \\ \end{equation}
Using the R package boot, we generated distributions of plant-soil feedback ratios by calculating 95% bias-corrected and accelerated (BCa) bootstrapped confidence intervals from 10,000 repetitions. Significant differences were found when there was no overlap between the 95% confidence intervals with the mean PSF ratios of other treatments.
Bacterial diversity analysis
Samples were rarefied to 18,738 reads to normalise variation in library sizes across samples of the soil, rhizosphere and endosphere samples (Cameron et al. 2021) (Figure S4). We also visualised the relative abundance of major phyla, and used differential abundance analysis to evaluate differences across each treatment using the ancombc2 () function in the R package ANCOMBC using non-rarified data (Lin & Peddada 2020).
To calculate alpha diversity across plant compartments and treatments, we estimated the effective number of ASVs by taking the exponential transformation of Shannon’s diversity (Jost 2006). Comparisons in alpha diversity levels across treatments were conducted using permuted linear mixed effects models, and permuted analysis of variance (ANOVAs).
Bacterial communities were visualised using non-metric multidimensional scaling (NMDS) and principal coordinates analysis (PCoA) ordinations with Bray-Curtis distances. The effect of treatments on the bacterial communities were estimated via permutational multivariate analysis of variance (PERMANOVA) using the adonis2 () function in vegan (Oksanen J. et al. 2019).
RESULTS