Acute interaction between hydrocortisone and insulin alters the plasma metabolome in humans

With the aim of identifying biomarkers of glucocorticoid action and their relationship with biomarkers of insulin action, metabolomic profiling was carried out in plasma samples from twenty healthy men who were administered either a low or medium dose insulin infusion (n = 10 each group). In addition, all subjects were given metyrapone (to inhibit adrenal cortisol secretion) + /− hydrocortisone (HC) in a randomised crossover design to produce low, medium and high glucocorticoid levels. The clearest effects of insulin were to reduce plasma levels of the branched chain amino acids (BCAs) leucine/isoleucine and their deaminated metabolites, and lowered free fatty acids and acylcarnitines. The highest dose of hydrocortisone increased plasma BCAs in both insulin groups but increased free fatty acids only in the high insulin group, however hydrocortisone did not affect the levels of acyl carnitines in either group. The clearest interaction between HC and insulin was that hydrocortisone produced an elevation in levels of BCAs and their metabolites which were lowered by insulin. The direct modulation of BCAs by glucocorticoids and insulin may provide the basis for improved in vivo monitoring of glucocorticoid and insulin action.


Pre-treatment
Prior to multivariate analysis, data were transformed using (log2) and Pareto variance scaled (Par) where the responses for each variable are centred by subtracting its mean value and divided by square root of its standard deviation.

Data visualisation and biomarkers identification
Prior to modelling the data, Hotelling's T 2 and DModX limits were employed to detect outliers of samples which could possibly affect the whole model. Samples were removed when above the 99% red line (action limit) of Hotelling's T 2 or exceeded the 95% orange line (warning limit) of Hotelling's T 2 plus Dcrit (critical limit) of DModX 2 . Principal Component Analysis (PCA) is an unsupervised model employed to explore how variables clustered regardless of Y class 3 .
Orthogonal projections to latent structures (OPLS) is a supervised model that predicts Y from X and can separate variation in X that correlates to Y (predictive) and variation in X that is uncorrelated to Y (orthogonal/systemic). OPLS-DA is a discriminant analysis of OPLS and employed to examine the difference between groups while neglecting the systemic variation. P values of biomarkers were corrected using a false discovery rate (FDR) 4 using Metaboanalyst 3.0 software (http://www.metaboanalyst.ca/). Variable importance in the projection (VIP) was employed, which served to indicate the contribution of each variable in the metabolomic change to a given model compared to the rest of variables 5 ; the average VIP is equal to 1.
Based on that value, a variable larger than 1 was deemed to have more contribution in explaining y 6 . A 95% confidence interval was calculated for each metabolite based on jack-knifes of uncertainty, which estimated the prediction error rate based on the cross validation rule used 6 . The biomarker selection workflow was as follows: 1. After analysis using SIMCA-P, the metabolites were filtered based on their p-values and 95% CI of mean difference; if a metabolite had a p-value > 0.05 and/or its 95% CI crossed 0, then it was filtered out.
2. All the significant metabolites were processed by using Metaboanalyst in order to get FDR corrected p-values and area under the ROC curves (AUC); if the metabolite had an FDR > 0.05 and/or AUC < 0.7, it was filtered out.
3. The remaining significant metabolites were tested again using split-plot analysis of variance (ANOVA) in order to examine the possibility of significant interaction between both interventions.

Diagnostics and validation
R 2 and Q 2 are diagnostic tools used in both supervised and unsupervised models; R 2 represents the percentage of variation explained by the model, Q 2 indicates the percentage of variation in response to cross validation 7 which meant that it was capable of predicting with a much greater level of accuracy than chance. Cross validation by SIMCA-P -by default -leaves 1/7 th of the data out at each iteration, hence, the appropriateness of the cross-validation was examined by a plot of Y observed vs Y predicted; the R 2 value was used to optimise the number of latent variables (orthogonal axis) 8 . A permutations test was applied to supervised models to evaluate whether the specific grouping of the observations in the two designed classes was significantly better than any other random grouping in two arbitrary classes 9 .
Model validity was assessed using cross validated ANOVA (CV-ANOVA), corresponding to H0 hypothesis of equal cross validated predictive residual of the supervised model in comparison with variation around the mean 10 . The area under curve (AUC) of a receiver operating characteristic (ROC) was used to assess the predictability of the classifier with a rough guide as follows: 0.9-1.0 = excellent, 0.8-0.9 = good, 0.7-0.8 = fair, 0.6-0.7 = poor, 0.5-0.6 = fail (36). Fisher's prob. 9.6e-009    observations. The model consists of one predictive x-score component; component t [1] and one orthogonal x-score components to [1]. t [1] explains 38.5% of the predictive variation in x, to [1] explains 28.3% of the orthogonal variation in x, R 2 X (cum) = 0.66, R 2 Y (cum) = 1, R 2 (cum) = 0.76.

Figure S6. Heat map shows metabolites that significantly affected by both interventions insulin and HC.
The plot shows heat map of the putative biomarkers ( Table 7)