Statistical analyses
To assess the effects of fish kairomones and subpopulation on the metabolome of D. magna , we applied two parallel multivariate analyses: a principal component analysis (PCA) coupled with two-way ANOVA and an ANOVA-simultaneous component analysis (ASCA). The results of both analyses were very similar and here we only show the PCA - ANOVA results (see details for the ASCA results in the Supplementary Information). PCA was conducted on the processed data matrices to assess the broad-scale variation between the two treatments using the PLS Toolbox (version 5.5.1, Eigenvector Research, Manson, WA, USA) within Matlab (version 7.8; The MathsWorks, Natick, MA, USA) following mean centring of the processed DIMS data. We extracted the first two PC axes for both ion modes; these explained 43.1 % (positive ion mode) and 42.9 % (negative ion mode) of the total variation. In order to test the effects of fish kairomones and subpopulation on the metabolome ofD. magna , we then applied two-way ANOVAs on the generated PC scores in Statistica v12.0. In each analysis, the fish kairomone treatment, subpopulation and their interaction were included as fixed factors and clone was nested in subpopulation as a random factor. A significant effect of the fish kairomone treatment indicates plasticity, while a subpopulation effect indicates rapid evolution of the trait means, and a fish kairomone × subpopulation interaction indicates rapid evolution of plasticity.
As we found strong fish kairomone × subpopulation interactions on the metabolome, we applied partial least squares discriminant analysis (PLS-DA) to each subpopulation separately to identify the specific metabolic responses to fish kairomones for each subpopulation. PLS-DA uses prior knowledge of the sample classes (here the fish kairomone treatments) to maximize separation of the metabolic profiles of the different classes and to derive predictive models (Nicholson et al. 2002). Internal cross-validation and permutation testing (see details in Supplementary Information) were employed to prevent over-fitting of the data (Westerhuis et al. 2008). Putative marker metabolites in response to fish kairomones for each subpopulation were screened using as criterion a Variable Importance in Projection (VIP) threshold greater than 1 (Xuan et al. 2011). All putative marker metabolites for each subpopulation were compared to screen for the general and subpopulation-specific metabolites responsive to fish kairomones. PLS-DA was conducted using in-house scripts with the PLS-Toolbox in Matlab.
In addition, changes in the intensities of individual m/z peaks were also assessed using t-tests for each subpopulation separately. All t-tests were corrected using a false discovery rate (FDR, Benjamini & Hochberg, 1995) of 5% to account for multiple testing and adjusted p-values are reported. Differences in the number of significantly changed peaks among subpopulations were tested using a chi-square test.
Pathway analyses
We used MI-Pack and KEGG to annotate the metabolites (see details in the Supplementary Information). We then used MetaboAnalyst (Xia & Wishart 2011) to analyse the metabolic pathways that were affected by fish kairomones. We put all putatively annotated KEGG compounds with VIP scores > 1 (based on the PLS-DA model including all three subpopulations) into MetaboAnalyst for metabolic pathway visualisation. Fisher’s exact tests were used for over-representation analysis (Toyotaet al. 2016) and out-degree centrality was used for pathway topology analysis (Xia & Wishart 2011). The FDR-corrected p values and impact values of all annotated pathways were plotted. Pathways were filtered based on the uncorrected p values (-log p > 0.5) and impact value (> 0.2) as those pathways were considered as potentially affected (Ratnasekhar et al. 2015).