Identification of human glucocorticoid response markers using integrated multi-omic analysis from a randomized crossover trial.

Background: Glucocorticoids are among the most commonly prescribed drugs, but there is no biomarker that can quantify their action. The aim of the study was to identify and validate circulating biomarkers of glucocorticoid action.
Methods: In a randomized, crossover, single-blind, discovery study, 10 subjects with primary adrenal insufficiency (and no other endocrinopathies) were admitted at the in-patient clinic and studied during physiological glucocorticoid exposure and withdrawal. A randomization plan before the first intervention was used. Besides mild physical and/or mental fatigue and salt craving, no serious adverse events were observed. The transcriptome in peripheral blood mononuclear cells and adipose tissue, plasma miRNAomic, and serum metabolomics were compared between the interventions using integrated multi-omic analysis.
Results: We identified a transcriptomic profile derived from two tissues and a multi-omic cluster, both predictive of glucocorticoid exposure. A microRNA (miR-122-5p) that was correlated with genes and metabolites regulated by glucocorticoid exposure was identified (p=0.009) and replicated in independent studies with varying glucocorticoid exposure (0.01 ≤ p≤0.05). Conclusions: We have generated results that construct the basis for successful discovery of biomarker(s) to measure effects of glucocorticoids, allowing strategies to individualize and optimize glucocorticoid therapy, and shedding light on disease etiology related to unphysiological glucocorticoid exposure, such as in cardiovascular disease and obesity. Funding: The Swedish Research Council (Grant 2015-02561 and 2019-01112); The Swedish federal government under the LUA/ALF agreement (Grant ALFGBG-719531); The Swedish Endocrinology Association; The Gothenburg Medical Society; Wellcome Trust; The Medical Research Council, UK; The Chief Scientist Office, UK; The Eva Madura's Foundation; The Research Foundation of Copenhagen University Hospital; and The Danish Rheumatism Association. Clinical trial number: NCT02152553.

Introduction

Glucocorticoids (GCs) have a key role in the metabolic, vascular, and immunological response to stress (Cain and Cidlowski, 2017; Oster et al., 2017). GC secretion from the adrenal gland is under tight dynamic control by the hypothalamic–pituitary–adrenal axis and is regulated in a classic circadian pattern (Cain and Cidlowski, 2017; Oster et al., 2017). Most actions of GCs are mediated by the ubiquitously expressed GC receptor (Cain and Cidlowski, 2017; Oster et al., 2017). The tissue-specific effects of GCs are regulated by many local factors, including pre-receptor metabolism of GCs and the interaction of the GC receptor with tissue-specific transcription factors, or through non-genomic mechanisms (Cain and Cidlowski, 2017; Oster et al., 2017). As a result of this complexity, circulating levels of cortisol relate poorly to tissue action of cortisol, and serum cortisol therefore has limited value as a biomarker for GC action (Karssen et al., 2001).

GCs are among the most commonly prescribed drugs, and GC treatment remains a cornerstone in the management of many rheumatic and inflammatory diseases despite the introduction of modern disease-modifying antirheumatic drugs and biological immunomodulatory treatment (Smolen et al., 2017). GC replacement is essential for survival in patients with various forms of adrenal insufficiency (Johannsson et al., 2015). However, metabolic and other side effects of GC treatment or replacement are common (Björnsdottir et al., 2011; Fardet et al., 2012), indicating that current methods to monitor their action and tailor their treatment are inadequate. Unphysiological GC exposure has been implicated in the etiology of several common diseases such as type 2 diabetes mellitus, hypertension, abdominal obesity, and cardiovascular disease (Ragnarsson et al., 2019).

Against this background, it is highly desirable to be able to measure and quantify GC action as this might be useful to refine current GC therapy. Biomarkers of GC action will also provide potential mechanistic understanding for the role of GCs in the etiology of many common diseases. Previous attempts to identify biomarkers using metabolomics have identified circulating metabolites associated with GC exposure (Alwashih et al., 2017a; Alwashih et al., 2017b). Integrated multi-omic analysis provides increased robustness over analysis of individual ‘omic data sets (Ideker et al., 2011). In particular, the identification of groups within one ‘omic ‘layer’ with shared co-regulation within another ‘omic layer implies a functional relationship that can be used both to assess the mechanistical relevance and to support the identification of biomarkers (Karczewski and Snyder, 2018; Misra et al., 2018).

The aim of this exploratory study was to define multi-omic patterns derived from independent tissues related to GC action and to use these patterns to search for clinically applicable circulating biomarkers of GC action. Subjects with primary adrenal insufficiency, Addison’s disease, lack GC production from the adrenal cortex and can therefore be considered a human GC ‘knock-down’ model (Figure 1A). An experimental study design including subjects with Addison’s disease, standardizing for diurnal variation and food intake, allowed a within-individual comparison between physiological GC exposure and GC withdrawal (Figure 1B). A multi-omic analysis strategy combining data from gene expression in circulation (peripheral blood mononuclear cells [PBMCs]) and an important metabolic tissue, adipose tissue, integrated with circulating microRNAs (miRNAs) and metabolites was used to identify putative biomarkers. The strongest putative biomarkers were then replicated in independent study groups with different GC exposure.

Figure 1.

Clinical and analytical part of the exploratory study and the replication step.

(A) Subjects with Addison’s disease (primary adrenal insufficiency, step 1) were studied in a random order during both physiological glucocorticoid (GC) exposure and GC withdrawal (step 2). Transcriptomics (whole-genome expression) in peripheral blood mononuclear cells (PBMCs) and adipose tissue (n = 28,869,869 genes), plasma miRNAomics (n = 252), and serum metabolomics in morning samples were analyzed (n = 164) (step 3). Integration of the multi-omic data derived a network including gene expression (derived from two independent tissues), microRNAs (miRNAs), and metabolites that were statistically differentiated between the two interventions (step 4). The miRNA findings, because of their centrality in the network, were replicated in subjects with different GC exposures (within the physiological range) from three independent studies (step 5). (B) Subjects with Addison’s disease (primary adrenal insufficiency) received in a random order intravenous (i.v.) hydrocortisone (HC) infusion mixed in 0.9% saline in a circadian pattern (physiological GC exposure) or the same volume of 0.9% saline alone (GC withdrawal) during 22 hr starting at 9 AM more than 2 weeks apart. During the GC exposure, HC(Solu-Cortef) was administered at a dose of 0.024 mg/kg/hr between 9 AM and 12 PM (first day), 0.012 mg/kg/hr between 12 PM and 8 PM (first day), 0.008 mg/kg/hr between 8 PM and 12 AM (first day), and 0.030 mg/kg/hr between 12 AM and 7 AM (second day). Samples for the ‘omics analyses were collected at 7 AM on day 2 of the intervention (morning samples). p.o.: oral.

Results

Clinical experimental study

Patient characteristics

Eleven subjects with well-defined Addison’s disease and no other endocrinopathies were recruited and included in the study between September 2013 and September 2015. One subject discontinued the study after randomization and before the first intervention because of persistent orthostatic hypotension. Ten subjects (four women with three of them post-menopausal) with a median age of 50 years (range, 25–57) and a median disease duration of 23.5 years (range, 1–33) completed all aspects of the study between May 2014 and October 2015. The median daily replacement dose of hydrocortisone (HC) prior to the study was 30 mg (range, 20–30), and 9 out of 10 subjects had treatment with fludrocortisone (mineralocorticoid) at a median daily dose of 0.1 mg (range, 0.1–0.2).

Clinical and biochemical outcomes

The main time points for sample collection in each intervention were at 9 AM on the first intervention day (‘before start’) and at 7 AM on the second intervention day (‘morning’) (Figure 1B). The subjects’ last ordinary oral HC dose was administered the day before admission to the study unit.

Infusion of HC mixed with isotonic saline (‘GC exposure’) had no effect on systolic and diastolic blood pressure, body weight, serum sodium and potassium, or plasma glucose concentrations compared to the same amount of isotonic saline infusion alone (‘GC withdrawal’) (Table 1). HC and saline infusion achieved the intended differences in GC exposure. Both median morning serum cortisol and cortisone during the HC infusion were within the physiological range (298 and 81.2 nmol/L, respectively) and markedly lower during the saline infusion (44.4 and 42 nmol/L, respectively, both p<0.001) (Figure 2). Serum cortisol and cortisone were detected in all subjects’ morning samples during the saline infusion, but both overnight (between 12 AM and 7 AM) urinary cortisol and cortisone excretion were below the limit of detection. Both HC and saline infusions were well-tolerated, and no serious adverse events were observed. Three subjects reported mild physical and/or mental fatigue, and one subject reported mild salt craving during the GC withdrawal period.

Table 1.

Clinical and biochemical tests assessed or collected immediately before each intervention and at 7 AM on the second intervention day during both interventions (n = 10).

		9 AM before start of intervention (first intervention day)		7 AM (second intervention day)
		Median (IQR)	p-Value	Median (IQR)	p-Value
S-cortisol (nmol/L)	HC	43.2 (38.0–55.1)	0.36	298 (228–359)	<0.001
	Saline	46.7 (43.1–61.7)		44.4 (36.8–52.5)
Overnight U-free cortisol (µg/7 hr)*,†	HC	–		678 (459–814)	<0.001
	Saline	–		<0.01‡
S-cortisone (nmol/L)	HC	40.7 (29.5–49.2)	0.76	81.2 (60.5–94.9)	<0.001
	Saline	42.1 (29.7–50.4)		42.0 (28.9–47.3)
Overnight U-free cortisone (µg/7 hr)*,†	HC	–		136 (106–151)	<0.001
	Saline	–		<0.01‡
SBP (mmHg)	HC	122 (111–134)	0.88	123 (107–139)	0.65
	Saline	127 (112–131)		124 (101–137)
DBP (mmHg)	HC	76 (63–83)	0.88	69 (59–76)	0.55
	Saline	72 (67–80)		65 (60–70)
S-sodium (mmol/L)	HC	141 (139–142)	0.85	140 (138–141)	0.97
	Saline	141 (138–142)		140 (138–141)
S-potassium (mmol/L)	HC	4.4 (4.3–4.7)	0.27	4.4 (4.3–4.8)	0.26
	Saline	4.2 (4.0–4.5)		4.4 (4.3–4.5)
P-glucose (mmol/L)	HC	5.0 (4.5–5.3)	0.79	5.6 (5.2–5.8)	0.08
	Saline	5.0 (4.6–5.1)		5.2 (5.2–5.4)
Body weight (kg)	HC	72.2 (69.6–77.3)	0.53	72.8 (69.3–77.7)	0.63
	Saline	73.5 (70.5–79.5)		74.0 (70.5–79.2)

*Overnight U-free cortisol and cortisone were collected from midnight to morning (12 AM to 7 AM) during physiological GC exposure (HC infusion) and GC withdrawal (saline infusion).

†One of the 10 subjects was not included in the analysis because of a problem during sample collection.

‡Below the limit of detection.

DBP: diastolic blood pressure; GC: glucocorticoid; HC: hydrocortisone; IQR: interquartile range; P: plasma; S: serum; SBP: systolic blood pressure; U: urinary.

Figure 2.

Violin plots of serum and urinary glucocorticoids (GCs) during GC exposure and withdrawal.

Individual data and changes for morning (7 AM) serum cortisol and cortisone (A, B), and overnight (12 AM to 7 AM) urinary-free (U-free) cortisol and cortisone (C, D) from both interventions. Boxes represent interquartile range (IQR); whiskers mark spread of points within 1.5 times IQR; violins demonstrate distribution of results with the maximum width representing the highest density within each violin. Difference of median values between interventions is presented with p-values. HC: hydrocortisone.

Differentially regulated ‘omic elements associated with response to GCs

Similarity network fusion (SNF) was used to demonstrate overall similarity between subjects across and between ‘omic layers, prior to analysis (Appendix 1 and Appendix 1—figure 1). Differential gene expression was associated with GC response in both PBMC and adipose tissue (Appendix 1). Differential expression of metabolites and miRNA was identified in blood in relation to GC response (Appendix 1). Differentially expressed ‘omic elements (DEOEs) are presented in Table 2 and Supplementary file 1a–d. All DEOEs were used for integrated analysis, and false discovery rate (FDR)-corrected DEOEs were used for all other analyses (Table 2). DEOEs from the PBMC and adipose tissue transcriptomes were shown to have limited overlap in response to GC but were enriched for shared pathways, revealing an overlap that indicated shared mechanism in relation to GC exposure (Appendix 2 and Appendix 2—figure 3).

Appendix 1—figure 1.

Integration of unsupervised peripheral blood mononuclear cell (PBMC) transcriptomic, metabolomic, and miRNAomic data in subjects between glucocorticoid (GC) exposure and GC withdrawal at 7 AM using similarity network fusion.

Integration of all circulating ‘omic data (A) simultaneously and (B) stepwise. Each matrix represents a subject-by-subject comparison using similarity network fusion of change in signal of ‘omic data between GC exposure and GC withdrawal. The Euclidean distances between subjects based on their relative positions as determined by isomap dimensional scaling were used as a basis to determine similarity. Colors denote normalized continuous similarity score; red: dissimilar; yellow: similar. The complete data set for each specific ‘omic layer of data was used, transcriptome n = 53,617, GC-MS n = 82, LC-MS n = 82, miRNA n = 251. GC-MS, gas chromatography-mass spectrometry; LC-MS: liquid chromatography-mass spectrometry; miRNA, microRNA; S, serum.

Table 2.

Summary of differentially expressed ‘omic elements in association with response to glucocorticoids.

‘Omic data set	Total number of ‘omic elements	Number of significant elements (p<0.05)	Number of significant elements (FDR < 0.05)
PBMC transcriptome	28,869	4426	3997
Adipose tissue transcriptome	28,869	3520	3115
Plasma miRNAome	252	9	9
Serum metabolome	164	38	14

FDR: false discovery rate; miRNA: microRNA; PBMC: peripheral blood mononuclear cell.

Appendix 2—figure 3.

Gene Ontology related to glucocorticoid (GC) action in the peripheral blood mononuclear cell (PBMC) and adipose tissue transcriptome.

(A) Numeric summary of biological pathways (Gene Ontology) associated with differential gene expression between GC exposure and GC withdrawal in the PBMC and adipose tissue transcriptomes. (B) Biological pathways (Gene Ontology) associated with differential gene expression in both PBMCs and adipose tissue. Data demonstrate that while different genes may be involved in the transcriptomic response to GCs in the two independent tissues, there are common pathways through which GC action is manifested.

We assessed the impact of differential expression on the entire interactome to aid in the identification of similar GC-related function. Interactome network models were generated using differentially expressed genes (DEGs) from both the PBMC transcriptome and the adipose tissue transcriptome. These were shown to be consistent with one another (Appendix 2 and  Appendix 2—figures 1 and 2) despite the limited overlap of DEGs. GC-responsive genes were shown to have higher connectivity in the human interactome than expected by chance, demonstrated using 10,000 permutations of this network model (Appendix 2).

Appendix 2—figure 1.

Gene ontology and interactome network model of differential gene expression in peripheral blood mononuclear cells (PBMCs) between glucocorticoid (GC) exposure and GC withdrawal.

Differentially expressed transcriptomic data in PBMCs was used to determine (A) the canonical pathways involved and (B) infer an interactome network model using BioGRID database, consisting of 2467 nodes and 64 modules. (C) Modules of genes (octagons) within the interactome network model were determined using the Moduland algorithm and ranked by a hierarchy based on network centrality; central modules are functionally relevant as they can influence the activity of a wider proportion of the network. The modules are arranged (left to right) in columns of the top 10 genes in each ranked by network centrality (top to bottom). (D) Gene expression marked within the network model (red: up; green: down in hydrocortisone treatment). Causal elements (known literature associations implying regulatory function) are marked in blue.

Appendix 2—figure 2.

Correlated networks related to glucocorticoid (GC) action in the peripheral blood mononuclear cell (PBMC) and adipose tissue transcriptome.

The analyses were performed to identify similarities and differences in transcriptome (gene expression) associated to GC exposure in two different tissues. Network associations of GC response to genes with differential expression between GC exposure and GC withdrawal in (A) PBMC transcriptome and (B) adipose tissue transcriptome. Gray circles represent genes involved in response to GCs that are not differentially expressed. Sixteen genes up-regulated and four down-regulated in PBMCs. Thirteen genes up-regulated and eight down-regulated in adipose tissue.

Integration of PBMC and adipose tissue transcriptomes with plasma miRNAomic and serum metabolomic data

Hypernetworks are network structures where edges are not restricted to defining a relationship between two nodes but may be shared between many nodes. As such, these structures can be used to describe complex relationships that link multiple elements. Hypernetworks also allow for the same pair of nodes to be connected by multiple edges. This means that relationships between nodes can be ranked by the number of edges shared between them. Hypernetworks allow for the summary of correlation matrices, compressing the high-dimensional relationships between data points (transcripts/miRNA/metabolites) into a single metric of similarity. Hypernetworks facilitate integration of ‘omic data and can be used to define strongly associated elements. Elements with large numbers of shared edges are more similar and likely to be of functional relevance; clustering allows refinement of large ‘omic data sets to highly associated elements (Figure 3A, B). Hypernetworks are robust to random error and act to filter out false-positive correlations as these will not have a uniform pattern of correlation across all ‘omic elements.

Figure 3.

Hypernetwork analyses of integrated ‘omic data.

(A) Hypernetworks differ from traditional networks in that edges can connect more than two nodes. Nodes are represented by black circles, edges by colored lines and surfaces. This demonstration shows how one edge can connect (i) two nodes as a one-dimensional line, (ii) three nodes as a two-dimensional surface, and (iii) four nodes as a three-dimensional structure. Hypernetworks of ‘omic data can have edges shared between hundreds of nodes. (B) Hypernetwork diagram illustrating how a pair of nodes (a–h) can be connected by more than one edge. In this example, nodes e and d share two edges, as do b and d. (C) A hypernetwork plotted as a heat map can be used to investigate clustering of blood peripheral blood mononuclear cell (PBMC) transcripts, based on correlation to, for example, adipose tissue transcriptome. A central cluster, defined using hierarchical clustering, groups PBMC transcripts based on high numbers of shared edges (red square, n = 965). This approach was applied to define groups of PBMC transcripts with similar profiles when correlated against each other ‘omic layer. (D) Gene probe level overlaps between PBMC transcriptome clusters identified by hypernetwork shared with the other ‘omic data sets. PBMC transcriptomic changes are correlated with changes in miRNAome, adipose tissue transcriptome, and metabolome (gas chromatography-mass spectrometry and liquid chromatography-mass spectrometry overlaps combined); overlaps are common PBMC transcripts with correlation to the ‘omic data sets. Values in brackets represent the size of PBMC transcriptomic clusters drawn from all differentially expressed PBMC transcripts (n = 4426, p<0.05). Data demonstrates a fundamental relationship in glucocorticoid response between PBMC and adipose tissue (965 genes) and reinforces the presence of common pathways in these two independent tissues.

The analyses were performed to quantify groups of highly correlated ‘omic data sets. Hypernetwork heat maps show clusters of correlated peripheral blood mononuclear cell (PBMC) transcripts (total 4426, 3.9 × 10−8 < p<0.05) with differential expression (glucocorticoid [GC] exposure vs. GC withdrawal) shared with (A) adipose tissue transcriptome (red square, n = 965), (B) plasma miRNA (red square, n = 768), (C) serum metabolome measured gas chromatography-mass spectrometry (red square, n = 915), and (D) serum metabolome measured by liquid chromatography-mass spectrometry (red square, n = 312). Increasing numbers of shared correlations appear as a scale from blue to white as shown in the keys.

(A) Peripheral blood mononuclear cell (PBMC) transcriptome probe-level log-intensity distribution. (B) Adipose tissue transcriptome probe-level log-intensity distribution. Density distributions represent intensity per probe, normalized to proportion of overall measurements at each intensity. (C) PBMC transcriptome per-sample log-intensity distribution (D) Adipose tissue transcriptome per-sample log-intensity distribution. Boxplots represent distribution of intensity per sample; boxes represent interquartile range, with median values and whiskers represent the range. All figures were generated in R from measurements generated by the R package ‘oligo’.

Distribution before and following normalization of (A) gas chromatography-mass spectrometry (GC-MS) samples, (B) GC-MS metabolites, (C) liquid chromatography-mass spectrometry (LC-MS) samples, and (D) LC-MS metabolites. For all figures, distribution of concentration values is provided as boxplots and density distributions for individual samples and metabolites. Boxes represent interquartile range with median values, and whiskers represent the range; outliers are denoted as individual points. Density distributions represent overall distribution of concentrations combined across samples or metabolites. All figures were produced using R package ‘MetaboanalystR’.

Hypernetworks were generated, dichotomizing the data using a range of cut-off r-values, from 0.1 to 0.9 by steps of 0.05, in order to generate the hypernetwork incidence matrix . Overall chi-squared distance was calculated for each hypernetwork as the sum of chi-squared distance across the hypernetwork matrix . A general additive model was used to fit a curve through these data using ggplot2. This approach determines how edges are assigned between pairs of ‘omic elements and, therefore, the number of shared edges represented by the hypernetwork matrix . Chi-squared distance represents the signal window present in a hypernetwork by measuring the spread of values (numbers of shared edges). The greatest spread in values here was found when dichotomizing around an r-value of 0.5. Above and below this threshold, the range of values in the hypernetwork was reduced, limiting the ability to refine a set of interacting ‘omic elements supported by most edges.

To assess similarity, we defined the correlation coefficient between each differentially expressed ‘omic measurement and assessed as 'present' in the network model those correlations with an r-value standard deviations (sd). Edges were defined as PBMC transcripts with shared correlations, for example, two PBMC transcripts that are both correlated with the same three metabolites are connected by three edges. We summarized the shared correlations as a measure of similarity between each pair of GC-responsive PBMC transcripts, counting correlations across the other ‘omic data sets (Figure 3—figure supplement 1). The greatest number of correlations shared was between PBMC and adipose tissue transcriptome (525 genes, Figure 3C), reinforcing the observation that, while the gene-level overlap of differential expression was limited, common pathways are active in both tissues related to GC action, which involve similar networks of co-expressed genes. The rank order of the number of correlations shared with the GC-responsive PBMC transcriptome was adipose tissue transcriptome > plasma miRNAome > serum metabolome, and this was confirmed both by comparison of the heat maps (Figure 3—figure supplement 1) and by a Venn diagram (Figure 3D). The Venn diagram also reveals a strong correspondence between the serum metabolome and both PBMC and adipose tissue transcriptomes.

Figure 3—figure supplement 1.

Hypernetwork heat maps of integrated ‘omic data.

The analyses were performed to quantify groups of highly correlated ‘omic data sets. Hypernetwork heat maps show clusters of correlated peripheral blood mononuclear cell (PBMC) transcripts (total 4426, 3.9 × 10−8 < p<0.05) with differential expression (glucocorticoid [GC] exposure vs. GC withdrawal) shared with (A) adipose tissue transcriptome (red square, n = 965), (B) plasma miRNA (red square, n = 768), (C) serum metabolome measured gas chromatography-mass spectrometry (red square, n = 915), and (D) serum metabolome measured by liquid chromatography-mass spectrometry (red square, n = 312). Increasing numbers of shared correlations appear as a scale from blue to white as shown in the keys.

Identification and validation of a shared transcriptomic profile in both PBMCs and adipose tissue predicting GC response

Robustness testing was performed in which hypernetworks were generated to model dissimilarity based on the absence of correlations with PBMC transcripts. Any genes that were highlighted by these hypernetworks were removed from the downstream predictive analysis. Using this approach, we defined 271 of 965 PBMC transcripts with maximum predictive potential. This set of genes perfectly classified the HC- and saline-treated groups using partial least squares discriminant analysis (PLS-DA) (Figure 4A). We identified variables of importance using Random Forest and modeled the background experimental noise using permutation analysis (BORUTA) (Figure 4B). This identified a set of 59 genes as variables of importance with fold changes in the same directions in both transcriptomic data sets that perfectly classified HC from saline treatment (Supplementary file 1e). Nine of these genes were significantly differentially expressed in both PBMC and adipose tissue transcriptomes (Figure 4C), and, of these nine genes, six were associated with GC response via gene ontology (IL18RAP, JAK2, MTSS1, RIN2, KIF1B, and BCL9L) (Figure 4D). The gene set (n = 59) that we identified, which classified both PBMC and adipose tissue transcriptomes in relation to GC exposure, was validated (area under the curve [AUC] 0.70–0.96) by further testing in five other previous studies of GC action by other research groups in cellular models (Table 3). Further robustness of the random forest observations was provided by demonstrating that the minimal depth at which the variables of importance became active in prediction was small (Figure 4—figure supplement 1).

Figure 4.

An overlapping gene set in peripheral blood mononuclear cell (PBMC) and adipose tissue transcriptome can be used to classify glucocorticoid (GC) response.

These analyses were performed to depict the common predictive genes in PBMCs and adipose tissue. (A) Partial least squares discriminant analysis (PLSDA) showing complete separation of hydrocortisone (HC) infusion (GC exposure, blue points) from saline infusion (GC withdrawal, orange points) using 271 of 965 PBMC transcripts confirmed as robust in the hypernetwork by analysis of dissimilarity. X-variates 1 and 2: PLSDA components; expl. var: explained variance. (B) BORUTA feature selection identifies variables (genes) of importance in classification using a Random Forest approach to model experimental background noise (green: confirmed classification; yellow: tentative classification: red: rejected classification; blue: ‘shadow’ variable modeling experimental noise). Of 271 transcripts initially used, 59 were identified as important (confirmed [green] or tentative [yellow]) in separating GC exposure from GC withdrawal, as well as having the same direction fold change in both PBMC and adipose tissue transcriptomic data sets. (C) Predictive genes that are significantly differentially expressed between GC exposure and GC withdrawal in both PBMC and adipose tissue transcriptomes and display fold change in the same direction in both tissues (n = 9). (D) Association of predictive genes (six out of nine) with GC response through gene ontology. Data demonstrates the presence of a robust transcriptomic profile predicting GC response in two independent tissues.

Distribution of minimal depth among trees from random forest generated in Figure 4B. Minimal depth equals the depth of the node closest to the root, which is capable of splitting participants into groups. Minimal depth for variables in this tree is very low as all variables are significantly different between groups at an uncorrected p-value<0.05. Figure generated using R package ‘randomForestExplainer’.

Table 3.

Validation of the predictive genes from the current exploratory study against previous studies examining GC response in cellular systems.

Study title	GEO #	PMID	N	AUC (95% CI)	OOB* AUC†	OOB* error rate‡ (%)
Dexamethasone effect on epidermal keratinocytes in vitro	GSE26487	17095510 (Stojadinovic et al., 2007)	20	0.70 (0.51–0.89)	0.80	30
Dexamethasone effect on GC-resistant and -sensitive lymphoblastic leukemia cell lines	GSE22152	21092265 (Carlet et al., 2010)	24	0.71 (0.52–0.90)	0.78	29
In vivo GC effect on non-leukemic peripheral blood lymphocytes	GSE22779	21092265 (Carlet et al., 2010)	16	0.88 (0.63–1.0)	0.96	6
Osteosarcoma cell line response to activation of specific GC receptor alpha isoforms	GSE6711	17682054 (Lu et al., 2007) 22174376 (Jewell et al., 2012)	60	0.96 (0.89–1.0)	0.99	3
GC effect on lens epithelial cells	GSE3040	16319822 (Gupta et al., 2005)	12	0.83 (0.63–1.0)	0.72	17

*OOB data, the bootstrapping approach of Random Forest, ensures that every tree is built using ~63% of the available data, leaving ~ 37% that can be used for a validation test.

†AUC up to 0.96 demonstrates a high probability of correctly classifying a randomly selected sample from each study.

‡OOB error rate = prediction error using the OOB validation data.

The gene set that classified both PBMC and adipose tissue transcriptomes in relation to GC exposure with fold change in the same direction (see Figure 4B – 59 genes) was validated by further testing in five other publicly available studies of GC action in cellular systems.

AUC: area under the curve of the receiver operating characteristic; CI: confidence interval; GC: glucocorticoid; GEO: Gene Expression Omnibus; GEO #: study number deposited with GEO; N: study number size; OOB: out-of-bag; PBMC: peripheral blood mononuclear cell; PMID: PubMed ID number of the manuscript describing the data.

Figure 4—figure supplement 1.

Random forest distribution of minimal depth.

Distribution of minimal depth among trees from random forest generated in Figure 4B. Minimal depth equals the depth of the node closest to the root, which is capable of splitting participants into groups. Minimal depth for variables in this tree is very low as all variables are significantly different between groups at an uncorrected p-value<0.05. Figure generated using R package ‘randomForestExplainer’.

Integration of circulating ‘omic data sets leads to miRNA and metabolite markers of GC action

We further examined interactions between the circulating ‘omics data associated with GC exposure (Figure 3D). All of the circulating ‘omics data was combined to form a correlation matrix and hierarchical clustering used to identify ‘omic data points with similar correlation (Figure 5—figure supplement 1). Eleven clusters including transcriptomic, miRNAomic, and metabolomic data were identified, and these clusters were shown to have enrichment within the interactome network model (Supplementary file 1f and Appendix 2).

Figure 5—figure supplement 1.

Heat map with clusters of circulating ‘omic data associated with glucocorticoid exposure identified using a correlation matrix.

Along the diagonal 11 clusters including peripheral blood mononuclear cell transcriptomic, miRNAomic and metabolomic data were identified (depicted in squares with their numbers). Color scale represents correlation coefficient. .

We then quantified the number of correlations between all the circulating ‘omic data associated with GC exposure (n = 336) using a hypernetwork. This approach was used to define a group of highly connected multi-omic elements with a relationship to GC exposure (Figure 5A).

Figure 5.

Integration of all circulating ‘omic data sets associated with glucocorticoid (GC) response.

These analyses were performed to lead to putative biomarkers of GC action. (A) Hypernetwork summary heat map of shared correlations between all circulating ‘omic elements (peripheral blood mononuclear cell transcriptome, plasma microRNA [miRNA], serum metabolome; n = 336) with differential expression between GC exposure and GC withdrawal. (B) Network representation of central cluster (red square in A [n = 139], of which 120 map to genes/miRNA/metabolites). Blue circles: genes with differential expression; red circle: differentially expressed metabolites; green: differentially expressed miRNA. Red box highlights DCK, one of the nine genes identified as a classifier of GC response (see Figure 4C). (C) Module decomposition of the hypernetwork. Genes modules (hexagons representing multiple highly connected genes) named by the most central gene in each module. miR-122-5p is present in the core of two modules (shown); color of modules represents centrality hierarchy: red: most central in the network; green: least central in the network.

Along the diagonal 11 clusters including peripheral blood mononuclear cell transcriptomic, miRNAomic and metabolomic data were identified (depicted in squares with their numbers). Color scale represents correlation coefficient. .

A hypernetwork model of the core group of 139 highly connected elements was generated (Figure 5B). DCK was the only gene shared with the GC-dependent adipose tissue transcriptome that also had predictive value (highlighted with a red square in Figure 5B). Deletion of the DCK gene region has been shown to be associated with increased sensitivity to GCs (Malani et al., 2017), an observation in alignment with the reduction in expression we found in both PBMC and adipose tissue transcriptomes in association with GC exposure (Figure 4C).

The hypernetwork model (Figure 5B) also highlighted a range of related miRNAs and metabolites. A hierarchical model of modules within the network was assessed using the measure of network centrality (Figure 5C). These modules revealed multi-omic relationships and demonstrated that miR-122-5p was the only miRNA present in higher order modules as measured by network centrality. miR-122-5p was correlated with cortisol exposure and the expression of FKBP5, a regulator of GC sensitivity (cluster 11 in Figure 5—figure supplement 1 and Supplementary file 1f).

Targeted replication of the plasma miR-122-5p fold change from the experimental study in subjects with Addison’s disease using an independent RNA separation procedure showed a marked down-regulation of miR-122-5p by increased GC exposure (p=0.009) (Figure 6). Two subjects did not show this miR-122-5p response, one man (disease duration 2 years, body mass index [BMI] 23.8 kg/m2; hydrocortisone 20 mg daily, fludrocortisone 0.1 mg daily) and one woman (disease duration 23 years; BMI 28.1 kg/m2; hydrocortisone 30 mg daily, fludrocortisone 0.2 mg daily) who both experienced mild mental fatigue during GC withdrawal.

Figure 6.

Replication of miR-122-5p as a putative biomarker of glucocorticoid (GC) action in the current biomarker discovery study.

Targeted replication of plasma miR-122-5p fold change from the current study population between subjects with Addison’s disease during GC exposure and GC withdrawal showed a significant down-regulation of miR-122-5p expression with increased GC exposure (p=0.009) conducted using an independent RNA separation procedure in the same samples.

Replication of miRNA findings in independent study groups

Based on (i) the functional association of a circulating miRNA with gene expression and metabolomics, and (ii) the correlation between the PBMC transcriptome and plasma miRNAome (Figure 3D), a targeted replication of the plasma miRNA findings was conducted using an independent RNA separation procedure. Twelve miRNAs were re-analyzed in the current study and in three other independent studies including subjects with different GC exposures: (i) in 60 subjects with rheumatoid arthritis with and without tertiary adrenal insufficiency after a short-term stop in their GC treatment (low vs. physiological GC exposure, respectively) (Borresen et al., 2017); (ii) in 20 subjects with Addison’s disease receiving HC replacement therapy and in 20 matched healthy control subjects (low vs. physiological GC exposure, respectively) (Bergthorsdottir et al., 2017); and (iii) acute low, medium, and excessive GC exposure in 20 healthy subjects (Stimson et al., 2017).

From this analysis, miR-122-5p was significantly associated with different GC exposure in all studies (Figure 7A– D). The expression of miR-122-5p was higher in subjects with rheumatoid arthritis and reduced GC exposure due to tertiary adrenal insufficiency (Figure 7A), and subjects with Addison’s disease had higher expression of miR-122-5p than healthy matched controls (Figure 7B). In the experimental study in healthy subjects, the expression of miR-122-5p was increased both after low and excessive high GC exposure compared to medium GC exposure at both high and low insulin levels (Figure 7C, D, respectively). The other 11 miRNAs (including miR-425-3p) did not show a relationship with GC exposure in the three replication studies.

Figure 7.

Replication of miR-122-5p as a putative biomarker of glucocorticoid (GC) action in independent patient groups with different GC exposure.

(A) The expression of miR-122-5p was higher in subjects with rheumatoid arthritis and reduced GC exposure due to tertiary adrenal insufficiency after a short-term stop of the GC treatment (AI) than in those without tertiary adrenal insufficiency (Normal). (B) Subjects with Addison’s disease (AD) had higher expression of miR-122-5p than healthy matched controls (Control). In an experimental study in healthy subjects, the expression of miR-122-5p was increased both after low and excessive high GC exposure (LowGC and ExcessGC, respectively) compared to medium GC exposure (MedGC) at both (C) high and (D) low serum insulin levels. Diamond: mean. Box = median ± interquartile range. Whiskers = upper and lower quartiles. miR-122-5p axis is presented as normalized expression.

Discussion

In a clinical experimental study designed to identify biomarkers of GC action, we succeeded in generating two profoundly different states of GC exposure within the physiological range in the same individual. The novelty of this study is the identification of pathways related to GC response and putative biomarkers of GC action in gene expression, metabolome, and miRNAs derived from integrated multi-omic analysis in two independent tissues. We identified a transcriptomic profile that was under similar GC regulation in both PBMC and adipose tissue transcriptomes, which was then validated by comparison to a range of previously published data by other research groups from cellular assays. We also identified a circulating miRNA, miR-122-5p, which was correlated with the circulating transcriptome and metabolome findings, suggesting for the first time a functional role in GC action. Moreover, the association between the expression of miR-122-5p and GC exposure was replicated in three independent study groups.

In order to identify putative biomarkers of GC action in humans, a clinical study was considered to be the most appropriate experimental setting. Addison’s disease or primary adrenal insufficiency is a rare disorder, but a unique clinical model for GC biomarker discovery due to absent or very low endogenous GC production (Gan et al., 2014; Sævik et al., 2020). Subjects with Addison’s disease were studied in a random order during physiological GC exposure and GC withdrawal. During GC exposure, infusion of HC delivered in isotonic saline via an infusion pump using a circadian pattern and saline alone (using the same volume and infusion pattern as during HC infusion) was administered during the GC withdrawal in order to prevent a state of sodium and fluid deficiency. This study design therefore allowed a within-individual comparison accounting for circadian rhythm and food intake. The marked difference in serum and urinary cortisol and cortisone, and the similar serum electrolytes, glucose, body weight, and blood pressure between the two interventions support the experimental success of the study design and strongly indicate that confounders related to metabolic changes or other secondary events related to the GC exposure or GC withdrawal were not influencing the output of the study. The measurable but very low concentrations of serum cortisol and cortisone throughout the GC withdrawal may be explained by a residual adrenal steroid secretion in some subjects (Gan et al., 2014; Sævik et al., 2020) and/or due to conversion of cortisone to cortisol in the liver and adipose tissue (Stimson et al., 2014).

Network models of ‘omic data can be used as a framework to assess the potential utility of biomarkers (Stevens et al., 2014). In this study, we have used a hypernetwork model of GC action based on differential gene expression in PBMCs as a basis to integrate adipose tissue transcriptome, plasma miRNA, and serum metabolomic data. Hypernetwork analysis leverages the power inherent in large data sets to assess interactions between ‘omic elements in a manner that is robust to false positives (Battiston et al., 2020). The associated interactome network derived from the PBMC transcriptome was shown to contain a number of genes with previously known GC-dependent binding of NR3C1 (the GC receptor) to regulatory elements, evidence that supports the specificity of the study design (Davis et al., 2018; Casper et al., 2018). Gene ontology analysis of the differential gene expression identified a range of pathways classically associated with GC action including GC-receptor signaling, immunoregulatory pathways such as those involving NF-κB, metabolic pathways, and cell cycle pathways. The plasma miRNA and serum metabolomic data was shown to map to the interactome network model of GC action, and this was taken as support for this data being putative circulating biomarkers functionally related to GC action.

Differential expression induced by GC treatment in both PBMCs and adipose tissue was indirectly associated with similar downstream elements by gene ontology analysis. These genes were not directly implicated with GC response, so, while the exact mechanisms may be different in each tissue, effects are coordinated through the same elements. Integration of the multi-omic data including both PBMC and adipose tissue transcriptomes was performed in order to increase the robustness of putative markers that could reflect action in other tissues such as adipose tissue, which is an important target organ for the metabolic actions of GCs. The 59 genes that behaved similarly in PBMC and adipose tissue were then validated in a range of studies examining GC response in different cellular systems. These included primary cell culture on keratinocytes (Stojadinovic et al., 2007) and lens epithelial cells (Gupta et al., 2005), along with PBMCs (Carlet et al., 2010) and cancer cells [both lymphoblastic leukemia (Carlet et al., 2010) and osteosarcoma (Lu et al., 2007; Jewell et al., 2012)]. The set of nine genes co-regulated in relation to GC exposure and GC withdrawal in both PBMC and adipose tissue transcriptomes can therefore be considered as putative markers of GC response. These could be used as a gene set to interrogate GC action in other experimental settings.

All the miRNA findings in this study are novel. While emerging experimental evidence indicates impact on regulation of GC action at several points by miRNAs (Clayton et al., 2018), this is the first time that miRNAs are shown to be globally correlated to GC action in humans. Both the hypernetwork analysis and the interactome network model implied the functional significance of some miRNAs, particularly miR-122-5p. In our hypernetwork model, the expression of miR-122-5p was correlated with clusters of genes that were centrally coordinated by expression of both RNF157 and TBXAS1, the former suggested to be a key regulator of both PI3K and MAPK signaling pathways, commonly perturbated in cancer and metabolic disorders (Dogan et al., 2017). Expression of TBXAS1 is pharmacogenomic linked to inhaled GC exposure in asthma (Dahlin et al., 2020). miR-122 is precursor transcript of mature miRNAs, including miR-122-5p (Carthew and Sontheimer, 2009; Bartel, 2004). miR-122 is expressed in the liver in humans (Tsai et al., 2009; GTEx Consortium, 2015; GTEx Consortium, 2013) and mice (Tsai et al., 2009). Hepatocyte nuclear factor HNF4A (Li et al., 2011; Xu et al., 2010), along with HNF3A (FOXOA1), HNF3B (FOXOA2), and HNF1A (Xu et al., 2010; Coulouarn et al., 2009), has been shown to be a key regulator of miR-122 expression in human cells. Down-regulation of miR-122 in murine models has been associated with non-alcoholic fatty liver disease (Alisi et al., 2011) and diabetes mellitus (Guay et al., 2011), and in humans, miR-122-5p has also been associated with fatty liver disease (Raitoharju et al., 2016).

miR-122-5p may be a functional link between unphysiological GC exposure and metabolic and cardiovascular disease. Increased exposure to GCs impairs glucose tolerance and may induce type 2 diabetes (Hackett et al., 2014). Indeed, reduced miR-122-5p expression has been seen in animal models of diabetes, and the reduction of this miRNA in response to increased GC exposure may suggest that miR-122-5p is a functional link between GC action and metabolism. In support of these findings are observations showing that miR‐122-5p regulate insulin sensitivity in murine hepatic cells by targeting the insulin‐like growth factor (IGF) 1 receptor (Dong et al., 2019). Recent human studies have also suggested that miR-122-5p is an indicator of the metabolic syndrome, with reduced expression in response to weight loss in overweight/obese subjects (Hess et al., 2020). miR-122-5p has also been suggested as a biomarker of coronary artery stenosis and plaque instability (Wang et al., 2019; Singh et al., 2020; Ling et al., 2020). As unphysiological GC exposure has been associated with obesity, diabetes, and cardiovascular disease (Walker, 2007), it is possible that miR-122-5p is reflecting different GC exposure in these disorders. The subjects with Addison’s disease in our clinical experimental study had no other comorbidities previously known to be associated with miR-122-5p expression, and therefore the presence of such confounders in our miR-122-5p finding seems to be unlike.

Specific miRNAs circulating in a stable, cell-free form in plasma or serum may serve as biomarkers in some diseases (Kroh et al., 2010), and, in our integrated analysis, they seem to be a realistic and clinically useful marker of GC action. We therefore focused on the replication of the miRNA findings from the discovery study. For this purpose, we performed a targeted analysis of 12 putative miRNAs and analyzed them in 120 subjects from independent study groups with different GC exposure in terms of dose, duration of exposure, and route of administration. The rationale for selecting these groups was that their GC exposure mostly remained within the normal physiological range. Despite the experimental differences between these studies, and the fact that these studies were not designed to study miRNA biomarkers of GC action, miR-122-5p was down-regulated by increased GC exposure in all of them. One exception was when short-term excessively high GC exposure was studied in afternoon samples in 20 subjects. There is no clear explanation for this, except the possibility that high non-physiological GC exposure has other secondary effects that may affect the levels of miR-122-5p.

The network analysis also identified putative metabolomic markers of GC action. GCs have a key role in metabolic regulation of stress by mobilizing energy through glucose, protein, and lipid metabolism. Previous studies have found an association between different GC doses and levels of branched-chain amino acids, fatty acids, some acyl carnitines, and tryptophan and its metabolites (Alwashih et al., 2017a; Sorgdrager et al., 2018). In our study, the amino acid tyrosine and the pyrimidine base uracil had a central position in the hypernetwork, which defined a group of highly connected multi-omic relationships within physiological GC exposure. Some of the other metabolomic data from our study was also in line with previous metabolomic studies in patients with adrenal insufficiency (Alwashih et al., 2017b; Sorgdrager et al., 2018). Excessive exposure to GCs in healthy subjects has, on the other hand, shown a strong, immediately and long-lasting impact on numerous biological pathways in the metabolome that may be either direct or indirect through the metabolic and cardiovascular action of pharmacological doses of GCs (Bordag et al., 2015).

There are some study limitations that need to be acknowledged. The low number of subjects included in the clinical experimental study could have reduced the power to detect a putative marker in individual ‘omic data sets, but this limitation was compensated for by the crossover study design and the integration of multi-omic layers. Another limitation is that we have only studied markers collected in the morning during physiologically peak cortisol exposure. However, the strengths of our study are the experimental study design, consideration of diurnal variation in GC action and impact of food intake, and the within-individual comparison, which minimizes confounders, as well as the fact that the putative markers that we have replicated are associated with known GC-responsive genes in two different tissues, suggesting their functional importance in GC action. Moreover, the integration of multi-omic layers allows for the reduction of background noise (Huang et al., 2017) and forms the basis for a detailed model of GC action. Hypernetwork summaries of correlation networks are recognized as providing signatures of mechanism (Pearcy et al., 2016; Johnson, 2011; Butte et al., 2000; Oldham et al., 2006) and, as such, are useful to assess both function and define markers of direct action.

In this clinical biomarker discovery study, we identified genes, miRNA, and metabolites that are differently expressed during GC exposure and GC withdrawal in subjects with Addison’s disease. The multi-omic data showed a high degree of coherence, and network analysis identified transcriptomics and metabolites that were closely correlated. The final outcome of the study is identification of a miRNA that is regulated by GC exposure and correlated with genes and metabolites that are also regulated by GCs in this study, indicating its functional relevance. The replication of this miRNA in three independent study groups increases the likelihood that the discovered miRNA, miR-122-5p, could become a biomarker of GC action to be used in clinical settings.

Materials and methods

Experimental study design

Study design

The study was a prospective, single-center, single-blind, randomized, two-period/crossover clinical trial.

Study subjects

Men and women with Addison’s disease for >12 months on stable cortisol replacement (with HC 15–30 mg/day) for ≥3 months followed at the Center for Adrenal diseases in the Out-patient Clinic at the Department of Endocrinology-Diabetes-Metabolism, Sahlgrenska University Hospital (tertiary referral hospital), Gothenburg, Sweden, were eligible for inclusion. Other inclusion criteria were age 20–60 years, body mass index 20–30 kg/m2, and ability to comply with the protocol procedures. Exclusion criteria were GC replacement therapy for indication other than Addison’s disease, any treatment with sex hormones including contraceptive drugs, treatment with levothyroxine, renal or hepatic failure, significant and symptomatic cardiovascular disease, diabetes mellitus, current infectious disease with fever, and pregnancy or breastfeeding. Recruitment was stopped when all eligible subjects had been asked to participate.

Power calculation was not performed because of the exploratory nature of the study. Power calculations were also difficult in the context of ‘omic analysis as there may be variable effect sizes over different ‘omic elements.

The study was approved by the Ethics Review Board of the University of Gothenburg, Sweden (permit no. 374-13, 8 August 2013) and conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all subjects before participation. The study was registered at ClinicalTrials.gov with identifier NCT02152553.

Study treatment

HC infusion was prepared by adding 0.4 mL of Solu-Cortef 50 mg/mL to 999.6 mL 0.9% saline, which resulted in 1 mg HC per 50 mL intravenous infusion. HC infusion was adjusted in accordance with previous observations in healthy males (Kerrigan et al., 1993) and interventions in both sexes (Løvås and Husebye, 2007; Figure 1B). The aim was to achieve a near-physiological circadian cortisol curve with early morning rise in serum cortisol that would peak at 7 AM and trough concentrations at midnight. In the GC-withdrawal intervention, 0.9% saline infusion alone was administered using the same volume as during the HC infusion. Thus, a person weighing 75 kg received 2 L of intravenous infusion over 22 hr during each intervention.

Interventions

All subjects were admitted after an overnight fast to the in-patient Endocrinology Department at the Sahlgrenska University Hospital at 8 AM (first intervention day) and were discharged at 12 PM the following day (second day). Subjects were randomized using a free randomization plan (generated at http://www.randomization.com/ on 27 April 2014) before the first intervention to receive either HC infusion or only saline infusion in a single-blind, crossover manner at least 2 weeks apart (Figure 1B). The researcher responsible for the clinical study generated the randomization plan, enrolled the study subjects, and assigned participants to interventions. Female subjects (when fertile) were studied during the early follicular phase (days 5–10) of their regular cycle under both interventions. Subjects were told not to take their ordinary mineralocorticoid dose on the day before each intervention but to take their ordinary HC dose. Subjects received standard meals at fixed times during both interventions. Their consumption of coffee or tea was recorded during the first intervention in order to consume the same amount and at the same time points during the second intervention.

During each intervention, the subjects’ blood pressure, body temperature, and weight were monitored. Because of the study design and the variations in circadian rhythm, blood sampling was collected at exactly the same time before the start of intervention, at midnight (12 AM), and in the morning of the second intervention day (7 AM). Urine was collected between midnight and morning (overnight), and abdominal subcutaneous fat was collected in the morning of the second intervention day immediately after blood and urine sampling. Adipose tissue was collected after local injection with lidocaine under the umbilicus on the right side of the abdomen during saline infusion and on the left side during HC infusion. The study was unblinded for each study subject after the completion of all aspects of the study (the second intervention).

Replication studies

Baseline samples in subjects treated with prednisolone for rheumatoid arthritis

This was a cross-sectional clinical study of prednisolone-induced adrenal insufficiency undertaken at the Department of Medical Endocrinology and Metabolism, at University Hospital, Rigshospitalet, Copenhagen, Denmark, between 2012 and 2018 (Borresen et al., 2017). In the current replication analysis, 60 subjects were included. All subjects had rheumatoid arthritis, received long-term prednisolone treatment (minimum 6 months), and treated with a current prednisolone dose of 5 mg/day. Of the 60 subjects, 23 had an insufficient response to the Synacthen test (GC-induced adrenal insufficiency, AI group) and 37 had a normal response (normal group). The samples included in the replication analysis were collected in the morning after an approximately 48 hr pause of prednisolone dosing (before the Synacthen test) and after overnight fasting. Plasma miRNA analysis of frozen samples was performed at Exiqon Services, Denmark.

Case–control study in subjects with or without Addison’s disease

This was an observational, cross-sectional, single-center, case–control study undertaken in our unit in Gothenburg, Sweden, between 2005 and 2009 (Bergthorsdottir et al., 2017). In the current replication analysis, the subgroup of 20 subjects with Addison’s disease under daily replacement therapy with oral HC ≥ 30 mg (AD group) and their 20 healthy control subjects with no GC therapy matched for age and gender (control group) were included. The samples included in the replication analysis were collected in the morning between 8 AM and 10 AM after an overnight fast, and for the cases after morning administration of their oral HC, which means a very low cortisol exposure during the night before sample collection. Plasma miRNA analysis of frozen samples was performed at Exiqon Services, Denmark.

Randomized, crossover study in healthy subjects

This was a randomized, double-blind study in 20 lean healthy male volunteers undertaken at the Edinburgh Clinical Research Facility between July 2010 and April 2012. The full protocol has been published previously (Stimson et al., 2017). Volunteers were randomized to receive either a low- or medium-dose insulin infusion (10 subjects in each group) and attended on three occasions after overnight fasting. Subjects received metyrapone (to inhibit adrenal cortisol secretion) with and without HC infusion (over 6.5 hr) in order to produce low, medium, or excessive GC levels (Low/Med/ExcessGC during high insulin and low insulin cohorts, respectively). The samples included in the replication analysis were collected in the afternoon at the end of each intervention (approximately 6.5 hr after start) on three occasions (low, moderate, or excessive high GC levels). Plasma miRNA analysis of frozen samples was performed at Exiqon Services, Denmark.

Generation and preparation of ‘omic data

Plasma cortisol and cortisone were analyzed using liquid chromatography-mass spectrometry (LC-MS), and urinary-free cortisol and cortisone were analyzed using gas chromatography-mass spectrometry (GC-MS) at the Mass Spectrometry Core Laboratory, Centre for Cardiovascular Science, University of Edinburgh, Edinburgh, UK. PBMCs were isolated on-site from whole blood using a gradient-based separation procedure and Ficoll-Paque PREMIUM (GE Healthcare).

A microarray gene expression analysis using Affymetrix Human Gene 2.0 ST arrays in both PBMC and adipose tissue was performed at the Array and Analysis Facility, Science for Life Laboratory at Uppsala Biomedical Center (BMC), Sweden.

The untargeted miRNA analysis in plasma was performed at Exiqon Services, Denmark. The targeted miRNA analyses in plasma (including the replication samples) were performed at Exiqon Services, Denmark, at a later date than the untargeted analysis. The 14 miRNAs included in the analysis based on the findings from the untargeted analysis were miR-425-3p, miR-186-5p, miR-15b-5p, miR-95-3p, miR-16-1-3p, miR-576-5p, miR-122-5p, miR-200a-3p, miR-193b-3p, miR-424-5p, miR-574-3p, miR-148a-3p, miR-18a-5p, and let-7g-5p.

Metabolic profiling of serum by GC-MS and LC-MS was performed at the Swedish Metabolomics Center in Umeå, Sweden.

Preprocessing of ‘omics data sets was carried out in the following ways. PBMC and adipose tissue transcriptomes were normalized using robust multichip average (RMA) via the R package oligo (Carvalho and Irizarry, 2010), which corrects for background variation, quantile normalizes, and summarizes features to gene-probe set level (Figure 3—figure supplement 2). GC-MS and LC-MS metabolomic data sets were analyzed using the R package MetaboanalystR (Chong and Xia, 2018), which filters variables based on ranked interquartile range, normalizes metabolites to sample median, and log transforms the resultant intensities (Figure 3—figure supplement 3). Qlucore Omics Explorer (version 3.3, Lund, Sweden) was used to scale and mean center miRome data. How all these analyses were performed is described in detail in Appendix 3.

Figure 3—figure supplement 2.

Transcriptome quality control measures.

(A) Peripheral blood mononuclear cell (PBMC) transcriptome probe-level log-intensity distribution. (B) Adipose tissue transcriptome probe-level log-intensity distribution. Density distributions represent intensity per probe, normalized to proportion of overall measurements at each intensity. (C) PBMC transcriptome per-sample log-intensity distribution (D) Adipose tissue transcriptome per-sample log-intensity distribution. Boxplots represent distribution of intensity per sample; boxes represent interquartile range, with median values and whiskers represent the range. All figures were generated in R from measurements generated by the R package ‘oligo’.

Figure 3—figure supplement 3.

Metabolome quality control measures.

Distribution before and following normalization of (A) gas chromatography-mass spectrometry (GC-MS) samples, (B) GC-MS metabolites, (C) liquid chromatography-mass spectrometry (LC-MS) samples, and (D) LC-MS metabolites. For all figures, distribution of concentration values is provided as boxplots and density distributions for individual samples and metabolites. Boxes represent interquartile range with median values, and whiskers represent the range; outliers are denoted as individual points. Density distributions represent overall distribution of concentrations combined across samples or metabolites. All figures were produced using R package ‘MetaboanalystR’.

Data analysis of differential gene expression

Principal component analysis (PCA) was performed to provide further quality control and define the relationship of variance between samples, allowing structure within the data set to be defined (Qlucore Omics Explorer 3.3). Quality control of transcriptomic data was performed using PCA with cross-validation and data consistency was confirmed. No outliers were identified. Differential gene expression was determined by a paired t-test comparing the two interventions. Network analysis of DEGs was performed using Advaita Bio’s iPathwayGuide (https://www.advaitabio.com/ipathwayguide); gene ontology performed using this software analysis tool implements the ‘Impact Analysis’ approach that takes into consideration the direction and type of all signals on a pathway, and the position, role, and type of every gene (Ahsan and Drăghici, 2017).

Gene ontology, gene expression regulated by miRNA, and causal network analysis

Gene ontologies were associated with differentially regulated gene lists (Ingenuity Pathway Analysis [IPA], Qiagen, Redwood City, CA). miRNAs were paired with genes that were theoretically regulated by specific miRNAs using IPA. The databases used for this mapping were TarBase (Vlachos et al., 2015), miRecords (Xiao et al., 2009), and peer-reviewed biomedical literature, as well as predicted miR–mRNA interactions from TargetScan (Agarwal et al., 2015).

The Encyclopedia of DNA Elements (ENCODE) data (Rosenbloom et al., 2013) was used to map genes in the interactome network model of GC action that had been previously shown to have dexamethasone dose-dependent DNA binding of NR3C1, the GC receptor gene.

Causal network analysis (CNA) allows the identification and prioritization of regulatory system elements within transcriptomic models. CNA was performed within IPA (Krämer et al., 2014). CNA identifies upstream molecules, up to three steps distant, that potentially control the expression of the genes in the data set (Krämer et al., 2014). A prediction of the activation state for each regulatory factor (master regulator), based on the direction of change, was calculated (Z-score) using the gene expression patterns of the transcription factor and its downstream genes. An absolute Z-score of ≥|1.4| and a corrected p-value<0.05 (Fisher’s exact test) were used to compare the regulators identified.

Network model construction and comparison

Lists of DEGs were used to generate network models of protein interactions in Cytoscape 2.8.3 (Smoot et al., 2011) by inference using the BioGRID (3.4.137) database (Chatr-Aryamontri et al., 2015).

The Cytoscape plug-in Moduland (Kovács et al., 2010; Szalay-Beko et al., 2012) was applied to identify overlapping modules, an approach that models complex modular architecture within the human interactome (Chang et al., 2013) by accounting for the non-discrete nature of network modules (Kovács et al., 2010). Modular hierarchy was determined using a centrality score and further assessed using hierarchical network layouts (summarizing the underlying network topology). The central module cores (metanode of the 10 most central elements) was determined and used as a basis to integrate the miRNA and metabolomic data. Transcriptomic and metabolomic data were combined to form a single network model using the Metscape (Karnovsky et al., 2012) plug-in for Cytoscape. Differential ‘omic data was compared and clustered in a correlation matrix using the corrplot plug-in (Murdoch and Chow, 1996) for R (R Development Core Team, 2020).

Similarity network fusion

Subject-level similarity network fusion (SNF) (Wang et al., 2014) was performed on ‘omic data as a test for similarity. To perform SNF, the SNFTool R-package was used (Wang et al., 2014). First, Euclidean distances were calculated between gene probe sets, and these were then combined using a nonlinear nearest neighbor method over 20 iterations. The fused data was subjected to spectral clustering and presented as a heat map.

Hypernetworks

We modeled the dynamics of potentially relevant PBMC and adipose tissue transcripts, miRNAs, and metabolites by assessing their activity as measured by the number of shared correlations against the background of all ‘omic elements called present after data processing.

A matrix (m rows and n columns) was generated of correlation distances (r-values) between the significantly differentially expressed multi-omic data (forming m rows) and all ‘omic data called present (forming n columns). The r-values were normally distributed.

A similarity matrix was defined by dichotomizing the correlation distance based on an r-value threshold of sd (if sd of , then value = 1; if sd of , then value = 0); the new matrix was termed M and represents the incidence matrix of the hypernetwork. An element of M, , where and are elements of and respectively, is defined as follows:

To generate the hypernetwork, we multiplied  by the transpose of , (Johnson, 2011; Ha et al., 2020), the elements of the resulting square matrix (, an matrix) are the number of correlations shared by each pair of interacting ‘omic elements; this is also the number of edges connecting each pair of nodes. was clustered using hierarchical clustering to identify the group of highly connected ‘omic elements.

The dichotomization parameters were shown to correspond to maximum signal window in the data using chi-squared distance metric (Figure 3—figure supplement 4). The chi-squared distance () was defined aswhere is the order of the matrix , is the element, , of , and is the expected value of an element of . The expected value of an element of  was calculated at any chosen dichotomization threshold by dividing the total number of correlations by the order of the matrix.

Figure 3—figure supplement 4.

Chi-squared distance distribution from a range of dichotomization thresholds.

Hypernetworks were generated, dichotomizing the data using a range of cut-off r-values, from 0.1 to 0.9 by steps of 0.05, in order to generate the hypernetwork incidence matrix . Overall chi-squared distance was calculated for each hypernetwork as the sum of chi-squared distance across the hypernetwork matrix . A general additive model was used to fit a curve through these data using ggplot2. This approach determines how edges are assigned between pairs of ‘omic elements and, therefore, the number of shared edges represented by the hypernetwork matrix . Chi-squared distance represents the signal window present in a hypernetwork by measuring the spread of values (numbers of shared edges). The greatest spread in values here was found when dichotomizing around an r-value of 0.5. Above and below this threshold, the range of values in the hypernetwork was reduced, limiting the ability to refine a set of interacting ‘omic elements supported by most edges.

Differential expression analysis was performed to refine genes for hypernetwork analysis. This approach serves to identify potentially relevant ‘omic elements. FDR-corrected p-values for all elements selected for hypernetwork integration are presented in Supplementary file 1g. We identified 4426 DEGs in PBMCs, 3520 adipose tissue DEGs, 38 metabolites (17 LC-MS, 21 GC-MS), and 12 miRNAs below an uncorrected p-value of 0.05. Data was analyzed across nine matching samples (normalized log2 score was inverted between GC exposure and GC withdrawal, i.e., +1 and –1, respectively).

A hypernetwork is inherently robust as individual correlations are not considered significant; rather hypernetworks model higher order interactions between nodes (‘omic elements) based on large numbers of shared edges (correlations). This approach only highlights ‘omic elements that are supported by the majority of the data and, as such, is robust to a wide range of r-value thresholds as well as small sample sizes.

Further, robustness of the hypernetwork observations was determined using a dissimilarity matrix derived from the original similarity matrix (i.e., the complement of the similarity matrix). The elements assessed as dissimilar were subtracted from those defined as similar. Elements within the output of the dissimilarity analysis that were also similar were eliminated from further predictive analysis.

The BORUTA R package (Kursa, 2014; Kursa and Rudnicki, 2010) was used for feature selection of transcriptomic data with predictive value. Random Forest (Breiman, 2001) was implemented in R using 5000 trees to determine the predictive value expressed as the area under the curve of the receiver operating characteristic.

Statistical analyses

Unsupervised analysis of metabolomic and transcriptomic data to assess how GC exposure grouped the study subjects was performed using Orthogonal Projections to Latent Structures Discriminant Analysis in SIMCA 13.0 (Sartorius) or PLS-DA MixOmics plug-in (Rohart et al., 2017) for R.

For quantitative variables with normal distribution, we performed independent samples t-test. Mann–Whitney U-test was performed for non-normally distributed variables. Chi-squared test or Fisher’s exact test, as appropriate, was used for categorical variables. Wilcoxon rank test was used for detecting differences between the two interventions in quantitative non-normally distributed variables. All statistical tests were two-sided, and p<0.05 was considered to be statistically significant. Further robustness for ‘omic data analysis was provided by considering the findings as clusters of co-expressed findings (Cleary et al., 2017). Statistical analyses were performed using SPSS (Statistical Package for Social Science) program, version 24 software for Mac.

Unless otherwise stated, all other statistical analyses were performed in R version 4.0.2 for Windows. Figures were plotted using ggplot2 (Wickham, 2016), gplots (Warnes et al., 2020), ggpubr (Kassambara, 2020), and reshape2 (Wickham, 2007).

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Acceptance summary:

The authors have identified candidate markers of glucocorticoid action that can now be investigated in larger cohorts. This work has generated a vast amount of data, including transcriptomics data from both adipose tissue and peripheral blood mononuclear cells, plasma micro RNA data and serum metabolomics data.

Decision letter after peer review:

Thank you for submitting your article "Identification of human glucocorticoid response markers using integrated multi-omic analysis" for consideration by eLife. Your article has been reviewed by three peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Clifford Rosen as the Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Natalie Bordag (Reviewer #2).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

We would like to draw your attention to changes in our revision policy that we have made in response to COVID-19 (https://elifesciences.org/articles/57162). Specifically, when editors judge that a submitted work as a whole belongs in eLife but that some conclusions require a modest amount of additional new data, as they do with your paper, we are asking that the manuscript be revised to either limit claims to those supported by data in hand, or to explicitly state that the relevant conclusions require additional supporting data.

Our expectation is that the authors will eventually carry out the additional experiments and report on how they affect the relevant conclusions either in a preprint on bioRxiv or medRxiv, or if appropriate, as a Research Advance in eLife, either of which would be linked to the original paper.

Summary:

The research focuses on the identification of markers of glucocorticoid action both at the transcriptomics level and the metabolite level. The availability of robust markers would be valuable in allowing the optimisation of glucocorticoid dosage. Ten subjects with primary adrenal insufficiency were used to enable biomarker discovery following treatment with hydrocortisone. While the cohort was small many of the markers observed had a high level of significance making this a useful preliminary study. A micro RNA (MiR-122-5p) was identified as being strongly correlated with glucocorticoid exposure and this finding was replicated in independent studies.

Essential revisions:

– Molecular data are often very noisy and require extensive cleaning. Which pre-processing approaches were applied? PCA is normally used to explore the sample structure and identify possible abnormalities while normalisation is needed to rescale the expression values to reduce noise. More explanations are needed on how PCA was used and whether any array normalisation was conducted.

– Certain results are not well explained. For example, those reported in the subsection “PBMC and adipose tissue transcriptomes have limited overlap in response to GC but are enriched for shared pathways” for PBMCs (the same applies to adipose tissue), how many differentially expressed genes (DEGs) did you find? This number is important to know before using those genes in pathway analysis. Did you use FDR or other forms of multiple testing correction to determine significance? It is unclear if you then used those DEGs to test pathway enrichment and you found a connectivity of n=4426 which is 3.7-fold stronger than using random genes. I am not sure what this means, or maybe 4426 is the list of DEGs? If so what is 3.7 connectivity? Please explain. Again, for pathways or GO terms enrichment, FDR should be used to correct for multiple testing. Overall, this analysis (Figure 2) does not appear to add much to the manuscript. Without multiple testing correction, relevant analyses seem to be those based on further gene filtering (i.e. retaining only genes confirmed across tissues and platforms such as those in Figures 4-8).

– Another example of low clarity is Figure 3. What is shared correlation? Heatmaps have no contrast and are uninformative. The dichotomisation by 0.5 is inappropriate, a higher cut-off (e.g. 0.8) might increase the contrast. Maybe the Venn diagram alone (Figure 3E) would be enough to convey your message if better explained?

– Discussion. miR-122-5p is not a novel circulating miRNA, it has been widely monitored as a diagnostic marker. Rephrase.

– There is a very extensive literature on miR-122-5p as diagnostic marker which I don't think has been thoroughly covered in the Discussion. For elevated miR-122-5p has been proposed as predictor on myocardial infarction which in turn has been associated with depressed hydrocortisone levels. Some additional discussion in this direction might be useful.

– Given that many conditions seem to affect miR-122-5p it would be important differentiate the effect of GC exposure from other conditions that affect its levels.

– In the metabolomics data the fold changes are quite small to confident on their robustness in such a small data set. Were the values of the metabolomic markers normally distributed?

Essential revisions:

– Molecular data are often very noisy and require extensive cleaning.

We recognize the point made by the editors/reviewers and have now made clear the extensive work we performed to establish the limits of noise and signal in our multi-omic data sets.

Note that we chose to use multiple approaches to ‘omic data-processing where possible. These approaches are now clearly outlined in the text both in the Materials and methods section and Appendix 1.

Which pre-processing approaches were applied?

This was performed as follows:

a) Transcriptome: RMA (performed using R package oligo) was used to background correct and quantile normalise, prior to summarising to gene-probe set level (subsection “Generation and preparation of ‘omic data” and Appendix 1).

b) Metabolome: Metaboanalyst (via R package MetaboanalystR) was used to filter variables based on ranked interquartile range, normalise metabolites to sample median and log transform the resultant intensities (subsection “Generation and preparation of ‘omic data” and Appendix 1).

c) miRNAome: Qlucore Omics explorer was used to scale and mean centre miRNAome data (subsection “Generation and preparation of ‘omic data”).

PCA is normally used to explore the sample structure and identify possible abnormalities while normalisation is needed to rescale the expression values to reduce noise. More explanations are needed on how PCA was used and whether any array normalisation was conducted.

PCA was performed on all ‘omic data sets to assess the presence of outliers using cross validation. This is now clarified in the Materials and methods section and Appendix 1). No outliers were identified and no batch correction was required as the experimental work was carried out in a single batch.

– Certain results are not well explained.

We have clarified all experimental results, provided extra detail and added extra explanation in the revised manuscript. The implementation of hypernetwork analysis to integrate and define relevant findings represents a novel aspect of this work and we have taken the opportunity to develop reader’s understanding of how we have built this analysis pipeline (Results and Materials and methods).

For example, those reported in the subsection “PBMC and adipose tissue transcriptomes have limited overlap in response to GC but are enriched for shared pathways” for PBMCs (the same applies to adipose tissue), how many differentially expressed genes (DEGs) did you find? This number is important to know before using those genes in pathway analysis. Did you use FDR or other forms of multiple testing correction to determine significance?

Candidate transcripts, miRNAs and metabolites used in hypernetwork analysis were identified initially using a supervised method without FDR correction. We subsequently reanalysed these candidates and determined that [3997/4426] PBMC transcripts, [3115/3520] adipose tissue transcripts and [17/38] metabolites were significantly differentially expressed below an FDR corrected p value of 0.05. FDR adjusted differential expression is not required for hypernetwork analysis as this approach is used as a method of refining a set of potentially relevant ‘omic elements in which to test network dynamics in an unsupervised manner with an assessment of robustness added downstream (see later description of hypernetwork analysis). This has been clarified in the text (subsection “Hypernetworks”), along with the additional explanation in the Results section and the addition of Supplementary file 1G, defining FDR corrected p-values for transcripts and metabolites used in hypernetworks.

Appendix 1 interrogates the similarities between PBMC and adipose tissue transcriptomes, using a false discovery rate modified p-value cut-off of <0.05. These lists are different to those used in hypernetwork analysis and p-value distributions that can be found in Supplementary file 1A.

It is unclear if you then used those DEGs to test pathway enrichment and you found a connectivity of n=4426 which is 3.7-fold stronger than using random genes. I am not sure what this means, or maybe 4426 is the list of DEGs? If so what is 3.7 connectivity? Please explain.

This point has been explained further. The analysis provides data showing an enrichment in connectivity of the differentially expressed genes in the interactome model and supports functional relevance. We have combined these data with the section from the start of the description of the hypernetwork data. These are minor pieces of circumstantial evidence and have now been added to Appendix 2.

Again, for pathways or GO terms enrichment, FDR should be used to correct for multiple testing. Overall, this analysis (Figure 2) does not appear to add much to the manuscript. Without multiple testing correction, relevant analyses seem to be those based on further gene filtering (i.e. retaining only genes confirmed across tissues and platforms such as those in Figures 4-8).

We agree that Figure 2 in the previous version of the manuscript does not add much to the manuscript and have added these results to Appendix 2 (Appendix 2—figure 2 and 3). As we were comparing the same pathways in both PBMC and adipose tissue we have reported p-value rather than FDR to facilitate identification of possible overlap. The data presented in the old Figure 2 highlight the hypothesis that there is a shared relationship in both PBMCs and adipose tissue with GC regulation and as such is minor point. We agree with the reviewer that the filtered analysis presented later is more important for interpretation of the main message of the manuscript.

– Another example of low clarity is Figure 3. What is shared correlation? Heatmaps have no contrast and are uninformative. The dichotomisation by 0.5 is inappropriate, a higher cut-off (e.g. 0.8) might increase the contrast. Maybe the Venn diagram alone (Figure 3E) would be enough to convey your message if better explained?

We have changed all heatmaps to a blue/yellow contrasting colour scheme in alignment with best practise.

We recognise that the hypernetwork concept is a novel concept and have taken the opportunity to spell out the interpretation in the Results section that describes the integrated ‘omic findings.

We now provide a figure to highlight the structure of a hypernetwork and describe more precisely how the analysis works (Figure 3A and B). We specifically explain shared correlation and its relationship to the underlying similarity analysis, although we leave the more precise mathematical description of the matrix structures for explanation in the Materials and methods section. We also provide further clarity in the Materials and methods section on how this method uses the correlation distance to simply define similarity based on the assessment of values of the standard deviation of the distribution of all r-values. The correlation distance is calculated in a similar manner to a Pearson correlation coefficient but the interpretation is as a distance metric. Note that we could “improve” the r-values by using Spearman’s correlation coefficient but doing this removes sensitivity to variation in the data and, importantly, loses a direct relationship to the Euclidean distances between the expression values of different genes. We chose this threshold based on the standard deviation equates to a Pearson correlation coefficient of 0.5 as this value maximised the signal window in the hypernetwork analysis ensuring specific analysis downstream. The hypernetwork signal window was assessed experimentally and these data are now presented (Figure 3—figure supplement 4). Using these approaches, we reflect the small size of the study in the absolute magnitude of the r-value in an unbiased manner but leverage the depth of the multi-omic data sets to increase confidence in the findings. With this experimental pathway the analysis is resistant to low r-value thresholds and robustness testing via comparison of the analysis with the complement of the underlying similarity matrix (i.e. the dissimilarity matrix) adds a final level of confidence in the analysis. The use of the dissimilarity matrix of the data allows a subtraction of those interactions defined as dissimilar from those that are similar and results in a highly robust non-statistical assessment of “presence” in the analysis (subsection “Hypernetworks”). We propose that this approach is more robust than permutation analysis that is also a possible method to determine robustness.

I addition, we have now focussed the description of the results on the Venn diagram which has been simplified for clarity as requested.

– Discussion. miR-122-5p is not a novel circulating miRNA, it has been widely monitored as a diagnostic marker. Rephrase.

We agree with the reviewer and the text in the Discussion section is rephrased accordingly.

– There is a very extensive literature on miR-122-5p as diagnostic marker which I don't think has been thoroughly covered in the Discussion. For elevated miR-122-5p has been proposed as predictor on myocardial infarction which in turn has been associated with depressed hydrocortisone levels. Some additional discussion in this direction might be useful.

We thank the reviewer for bringing this to our attention. New findings concerning the role of miR-122-5p in different disorders are added in the Discussion section. Moreover, a discussion that our miRNA finding (miR-122-5p) may be the functional link between unphysiological glucocorticoid exposure and disorders such as type 2 diabetes mellitus, obesity and cardiovascular disease is added in the Discussion section.

– Given that many conditions seem to affect miR-122-5p it would be important differentiate the effect of GC exposure from other conditions that affect its levels.

In conjunction with the response to the previous comment, we have added in the Discussion section a statement about the potential influence of the known conditions affecting miR-122-5p into our interventions.

– In the metabolomics data the fold changes are quite small to confident on their robustness in such a small data set. Were the values of the metabolomic markers normally distributed?

Quality control was done for both transcriptome and metabolome and normality was demonstrated. We now provide these data in the Materials and methods section and Figure 3—figure supplements 2 and 3.

Key resources table

Reagent type (species) or resource	Designation	Source or reference	Identifiers	Additional information
Recombinant DNA reagent	GeneChip WT PLUS Reagent Kit	Affymetrix Inc	Other	P/N 703174 Rev. 1
Commercial assay or kit	Human Gene 1.0 ST array	Affymetrix Inc	–	–
Commercial assay or kit	Human Gene 1.1 ST array	Affymetrix Inc	–	–
Commercial assay or kit	Human Gene 2.0 ST array	Affymetrix Inc	–	–
Commercial assay or kit	miRCURY LNA Universal RT microRNA PCR, Polyadenylation, and cDNA Synthesis Kit	Exiqon	–	–
Commercial assay or kit	miRCURY RNA Isolation Kit-Biofluids	Exiqon	–	–
Chemical compound, drug	Solu-Cortef	Pfizer Inc	–	–
Software, algorithm	Agilent Masshunter Profinder	Agilent Technologies, Inc	Other	Version B.08.00
Software, algorithm	SPSS	SPSS	RRID:SCR_002865	–
Software, algorithm	R	R Project for Statistical Computing	RRID:SCR_001905	–
Software, algorithm	Rstudio	Rstudio	RRID:SCR_000432	–
Software, algorithm	MetaboAnalystR	–	RRID:SCR_016723	–
Software, algorithm	Moduland algorithm	–	https://www.linkgroup.hu/docs/ModuLand-ESM1-v5.pdf	–
Software, algorithm	Cytoscape	Cytoscape	RRID:SCR_003032	–
Software, algorithm	Qlucore	Qlucore Omics Explorer	https://www.qlucore.com/bioinformatics	–
Software, algorithm	Robust Multi-Array Average algorithm	–	http://www.molmine.com	–
Software, algorithm	ChromaTOF	LECO	https://www.leco.com	–
Software, algorithm	MATLAB R2016a	Mathworks	https://www.mathworks.com	–
Software, algorithm	Roche LC software	Roche Molecular Systems, Inc	–	–
Software, algorithm	NormFinder	Aarhus University Hospital, Denmark	RRID:SCR_003387	–
Software, algorithm	NIST MS 2.0 software	NIST	https://chemdata.nist.gov	–
Other	LightCycler 480 Real-Time PCR System	Roche Molecular Systems, Inc	RRID:SCR_020502	–
Other	Agilent 1290 Infinity UHPLC-system	Agilent Technologies, Inc	https://www.agilent.com	–
Other	Agilent 2100 Bioanalyzer system	Agilent Technologies, Inc	RRID:SCR_018043	–
Other	Agilent 6550 iFunnel Q-TOF LC/MS	Agilent Technologies, Inc	RRID:SCR_019433	–
Other	ENCODE	Stanford University	RRID:SCR_015482	–
Other	UCSC Genome Browser	UCSC	RRID:SCR_005780	–
Other	TarBase	DIANA Tools	RRID:SCR_010841	–
Other	miRecords	Biolead.org	RRID:SCR_013021	–
Other	TargetScan	Whitehead Institute for Biomedical Research	RRID:SCR_010845	–
Other	BioGRID	TyersLab.com	RRID:SCR_007393	–
Other	Ingenuity Pathway Analysis	Qiagen	RRID:SCR_008653	–