Statistics
Five replicates were done for all conditions for ionomics, isotopes,
metabolomics and four replicates were done for proteomics. Supervised
multivariate analysis of omics data was carried out by orthogonal
projection on latent structure (OPLS) (Bylesjö et al. , 2007) with
Simca (Umetrics), using K, Ca and putrescine as predicted Y variables
and metabolites (or proteins) as predicting X variables. The absence of
statistical outliers was first checked using a principal component
analysis (PCA) to verify that no data point was outside the 99%
confidence Hotelling region. The goodness of the OPLS model was
appreciated using the determination coefficient R² and the predictive
power was quantified by the cross-validated determination coefficient,
Q². The significance of the statistical OPLS model was tested using a χ²
comparison with a random model (average ± random error), and the
associated P -value (P CV-ANOVA) is reported
(Eriksson et al. , 2008). Best discriminating features were
identified using volcano plots whereby the logarithm of theP -value obtained in univariate analysis (two-way ANOVA; factors
used: K x Ca or K x putrescine) was plotted against the
rescaled loading (pcorr) obtained in the OPLS. In such a
representation, best biomarkers have both maximal –log(P ) and
pcorr values. Univariate analysis of statistical classes
in bar plots was performed using a two-way ANOVA (Fisher statistics),
with a threshold of P as indicated in figure legends. When all
conditions were compared at once, a one-way ANOVA was conducted.