Multi-omics-based identification of atopic dermatitis target genes and their potential associations with metabolites and miRNAs

Abstract

Atopic dermatitis (AD), or atopic eczema, is one of the most common inflammatory skin diseases with up to 10% prevalence in adults, and approximately 15-20% in children in industrialized countries. As a result, there is an unmet need for faster, safer, and effective treatments for AD. AD pathogenesis represents a complex interplay between multiple factors, such as environmental factors or stimuli, genetic factors, immune dysfunctions. However, although multi-omics label studies have been very useful in understanding the pathophysiological mechanisms of AD and its clinical manifestations, there have been very few studies that integrate different labels of omics data. Here, we attempted to integrate gene expression and metabolomics datasets from multiple different publicly available AD cohort datasets and conduct an integrated systems-level AD analysis. We used four different GEO transcriptome data sets and, by applying an elastic net machine learning algorithm, identified robust hub genes that can be used as signatures, for example, H2AFX, MCM7, ESR1 and SF3A2. Moreover, we investigated potential associations of those genes by applying a pathway-based approach over metabolomics and miRNA datasets. Our results revealed potential novel associations between fatty acids and peroxisomal lipid metabolism pathways, as well as with several microRNAs.

Introduction

Atopic dermatitis (AD) is one of the most common inflammatory skin diseases with around 10% prevalence in adults, and approximately 20% in children [1]. AD is typically considered the result of a complex interplay between multiple factors, such as environmental factors or stimuli, genetic factors [2], immune dysfunctions, and host-microbiome interactions [3]. Omics analysis has been widely used in recent years to explore potential biomarkers and to gain a better understanding of the underlying pathobiology and pathophysiology of various clinical forms and different disease subtypes [3]. Numerous studies have been carried out in an effort to cater the molecular AD diagnosis, many of which entailing the analyses of high throughput omics experiments, namely transcriptomics [4-6], metabolomics [7,8], proteomics [9,10], etc. The primary objective of such studies is to identify markers that can be targeted for drug discovery or be used for the identification of biomarkers that cater disease stratification [11]. Although single-omics studies have been very useful for understanding the AD mechanisms and manifestation, multi-omics approaches offer the tantalising possibility of gaining an in-depth understanding of AD’s complex pathogenesis [3]. Currently, very few studies integrating different omics labels have been conducted [3]. Several studies examined the integration of lipidomics and transcriptomics [12], or transcriptomic and proteomic data [13], as well as of epigenetic and transcriptomic datasets [14] to explore the microenvironment and immune regulation in AD skin. Moreover, one recent study investigated the association of the infant gut microbiome with its metabolome in order to understand the impact of the microbiome on the risk of AD development [15]. While single omics approaches appear to be insufficient in identifying biomarkers for AD endotyping, novel multi-omics integration methods offer the potential of gaining a better understanding of the underlying AD molecular patterns, as well as of a potential interpretation of AD clinical variability [3].

In the present study, we attempted to integrate gene expression, miRNA and metabolomics datasets from multiple cohorts, and obtain an integrated systems-level overview of AD. Four different GEO transcriptome datasets were used and robust gene signatures determining core AD pathogenetic elements were identified. Furthermore, by adopting pathway-based approaches, it was demonstrated that metabolomics signatures were associated with particular gene signatures revealing novel genes, metabolites, and miRNAs associations.

Methods

Data collection

Table 1 lists the gene expression datasets obtained from the Gene Expression Omnibus (GEO) [16], as well as the metabolomics dataset retrieved from the MetabolomicsWorkbench [17]. All datasets were collected on 30th November 2020.

Table 1.

GEO datasets employed in this study

GEO ID and reference	Data types	Platform used	Sample size	Tissue type
GSE36842 [20]	Microarray	GPL570, Affymetrix Human Genome U133 Plus 2.0 Array	Normal = 15, AD = 8, Total = 23	Skin Biopsies
GSE16161 [21]	Microarray	GPL570, Affymetrix Human Genome U133 Plus 2.0 Array	Normal = 9, AD = 9, Total = 18	Skin Biopsies
GSE6012 [22]	Microarray	GPL96, Affymetrix Human Genome U133A Array	Normal = 10, AD = 10, Total = 20	Skin Biopsies
GSE32924 [23]	Microarray	GPL570, Affymetrix Human Genome U133 Plus 2.0 Array	Normal = 10, AD = 10, Total = 20	Skin Biopsies
ST001431-Metabolomics workbench [66]	Metabolomics	LC-MS targeted	Normal = 24, AD = 25, Total = 49	Blood Serum

Differential expression analysis using GEO2R

We used the GEO2R [18], an online tool employing the limma package [19], to identify differential genes (DEGs) between normal vs. AD skin. For this analysis, two groups were created (normal vs. AD) for each individual dataset using the GEO ‘define groups’ feature. Genes without a corresponding gene symbol and genes with more than one probe set were removed. The statistical significance values were set as adjusted P-value ≤0.05. Three GEO series datasets were used, namely GSE36842 [20], GSE16161 [21], and GSE6012 [22], to identify differential expressed genes, and the GSE32924 dataset [23] was used for validation. An overview of the workflow (process) is presented in Figure 1.

Figure 1.

Overview of the study workflow and methods.

Differentially expressed genes networks

A network analysis of the selected differentially expressed genes (DEGs) was subsequently performed using the OmicsNet tool [24]. Using this tool, we mapped protein-protein interactions and transcription factor-gene interactions. Protein-protein interactions were mapped using a manually curated experimentally validated PPI database (IntAct). Moreover, the number of connections (node interactions) for each of the differentially expressed genes was calculated, and only genes with more than 100 interactions were considered for further analysis.

Gene marker selection using machine learning methods

We applied two feature selection methods, namely the Least Absolute Shrinkage and Selection Operator (LASSO) [25] and the Elastic Net (EN) [26] methods. Both EN and LASSO are able to automatically select the best features linked with an outcome variable (for example, normal vs. AD). Both EN and LASSO are considered penalty-based methods and hence provide a sparse solution. The penalty parameter, λ (range of λ: 0-1), was optimized using 10-fold cross-validation. For the case of LASSO, the penalty is applied to the sum of the absolute values of the regression coefficients (L1 norm). Elastic Net, on the other hand, employs a mixed version of both L1 and L2 penalty (Ridge penalty). We combined selected variables identified by both LASSO and EN and then applied a generalized linear model (GLM) to cater for the stability analysis of the selected features. The process was repeated 100 times and the features were ranked according to their respective selection frequency associated with each run. We then selected the first quartile of the combined LASSO and EN selected features over 100 runs. These selected features were then further modelled using logistic regression. Two area under the curve (AUC) distributions resulted from this analysis. The one related to random label sampling (i.e., the randomization of sample labels in each iteration, averaged over 100 iterations) was termed as ‘random AUC’. The second distribution was based on the bootstrapped samples.

Enrichment analysis and miRNA association analysis

The enrichment analysis of identified DEGs was then performed to assess the potential molecular pathways, gene functions or biological processes associations. Pathway analysis was subsequently performed over the differentially expressed genes using the OmicsNet tool [24] and the KEGG gene database. Pathways of interest were subsequently inspected to identify DEGs. The metabolomic enrichment analysis was performed using the IMPaLA (Integrated Molecular Pathway-Level Analysis) [27] web tool. We merged miRNA information with the genes selected from the previous step using the NetworkAnalyst v3.0 webtool [28].

Statistical analysis

R version 4.0.3 [29] was used for statistical analyses. For differential expressed genes between normal vs AD skin samples, Benjamini-Hochberg (BH) adjusted P-value ≤0.05 was considered. Metabolites were analysed using a Wilcoxon rank test and P-values of <0.05 were considered statistically significant.

Results

Hub genes identification

Publicly available GEO datasets were used to identify genes that are differentially expressed between AD vs. control samples. We mapped protein-protein interactions (PPI) using a validated PPI database (IntAct) [30] that allowed us to identify highly connected genes (called degree). Within the GSE36842 dataset, we identified the GABARAPL1 gene to have the highest degree with 519 connections. In total, 10 genes were selected with a degree of more than 100 from the GSE36842 dataset. For the GSE6012 dataset, we found 24 genes with a degree more than 100 that were considered hub genes. The MYC gene was found to have the highest degree with 769 connections. Similarly, for the GSE16161 dataset, GRB2 was identified to have the highest number of the degree, and in total nine genes were identified. No common genes shared across three datasets were found. One common gene, namely CALM3, was identified between the GSE36842 and GSE6012 datasets and one common gene, YWHAZ, was found between GSE6012 and GSE16161 datasets. The list of the genes and their corresponding degrees is presented in Supplementary Table 1.

Elastic net-based selection of predictive genes

We applied the EN method using the identified hub genes from each of the datasets separately to select the important discriminating genes between AD and control samples. From GSE36842, EN selected 10 genes out of 10 input hub genes. From GSE16161, EN selected seven genes out of nine input hub genes and finally from GSE6012, EN selected six genes out of 24 hub input genes (Table 2). To investigate the predictive performance of the identified genes, we used an external validation dataset, GSE3292. Nine genes were identified within the validation dataset and a logistic regression model was applied to obtain their individual as well as combined predictive power between control vs. AD samples. The individual AUCs for each of the genes with confidence interval (CI) were found: H2AFX (AUC: 1.000, CI: 1-1), MCM7 (1.000, CI: 1-1), ESR1 (0.971, CI: 0.88-1), SF3A2 (0.846, CI: 0.64-0.98), UBA1 (0.601, CI: 0.21-0.63), TP53 (0.577, CI: 0.37-0.81), EF1A1 (0.572, CI: 0.36-0.76), GIGYF2 (0.567, CI: 0.38-0.78), PPP2CA (0.510, CI: 0.23-0.75). Out of these genes, four, namely, H2AFX, MCM7, ESR1, SF3A2, were associated with regression models with AUC values higher than 0.80. Figure 2A depicts box plots of the predicted genes with associated P-values, while Figure 2B presents the AUC values of the combined regression model.

Table 2.

Hub genes and predictive genes

GEO-ID	Hub genes (>100 P-P connectivity)	Elastic Net predictive genes
GSE36842	10	10
GSE16161	9	7
GSE6012	24	6

Figure 2.

A. Box plots of the predictive genes across AD vs. control samples. Significant (P<0.05) gene up-regulation is represented in solid red for the healthy patients and gene down-regulation is highlighted in solid blue for the AD patients. B. An AUROC with 0.81 value and a CI (0.44-1) related to the performance of the model using 4 genes on the GSE32924 dataset.

Gene set enrichment analysis

A pathway and gene enrichment analysis was performed to identify potential pathway associations. The top five pathways identified were the Class I PI3K signaling events, the integrin family cell surface interactions, the Arf6 trafficking events, endothelins, the EGF receptor (ErbB1) signalling, as well as IL-3-, IL-5- IL-8- and IFN-gamma-related immune signalling pathways. Multiple genes were found to be associated with those pathways (Figure 3). A list of these genes and their associated pathways is provided in Supplementary Table 2. We further investigated genes involved in biological processes such as cell cycle regulation, apoptosis, signal transduction, DNA replication, immune response, etc (Figure 3).

Figure 3.

(A) Common pathways identified across multiple GEO data sets and (B) common biological processes identified across multiple GEO data sets.

miRNAs associations

NetworkAnalyst was then used to explore potential miRNA associations with the previously identified genes. The TP53 gene was identified as having the highest numbers of interactions (also called as degree) amongst them, interacting with 130 miRNAs. The degrees of the PPP2CA, ESR1, EEF1A1, and MCM7 genes were 68, 90, 61 and 90, respectively. Figure 4A provides a visualisation of the predictive genes and their associations with miRNAs. The miRNA-gene interacting pairs are reported in Supplementary Table 3.

Figure 4.

A. Selected predictive genes with miRNA targets associations. B. Metabolite enrichment analysis using IMPaLA. The figure depicts the pathways found to be associated with AD (X-axis) and their associated-log10 (p) values (Y-axis).

Metabolite identification

A Wilcoxon rank test obtaining P-values (P<0.05) was performed over the serum metabolomics dataset (ID number ST001431). Cetylcarnitine, phosphatidylcholine diacyl, phosphatidylcholine diacyl, acylcarnitines, carnitine, acetylcarnitine, dicarboxylacylcarnitines were found to be significant. For these metabolites, a metabolic enrichment analysis was also conducted using IMPaLA. The top three metabolic pathways identified were the beta-oxidation of pristanoyl-CoA, peroxisomal lipid metabolism, and fatty acid metabolism pathways, while acetylcarnitine, phosphatidylcholine diacyl, carnitine and acylcarnitines were associated with those pathways significantly. A bar graph of the pathways identified and their associated (-log10) P-values are presented in Figure 4B.

Discussion

In this study, we aimed to identify candidate genes relevant to AD and to associate them with particular metabolites and miRNAs. We identified H2AFX, MCM7, ESR1, SF3A2, UBA1, TP53, EEF1A1, GIGYF2, PPP2CA as the hub and predictive genes. H2AFX is an H2A histone family member contributing to the chromatin-remodeling and the nucleosome-formation DNA repair preserving genetic integrity. H2AFX mutations have been shown to lead to breast cancer [31]. H2AFX has also been implicated in the hypoxia-driven neovascularization process [32]. The minichromosome maintenance complex (MCM) is known to be involved in pre-replication complex formation that helps in DNA denaturation and unwinds the replication fork [33]. The MCM complex is composed of six highly conserved MCM proteins [34,35]. Most of the MCM proteins are involved in carcinogenesis. Among them, MCM7 is considered to be an important unit of heteromeric MCM helicases which play a key role in controlling DNA replication and cell proliferation. A high level of MCM7 gene expression is also associated with tumour progression and metastasis and is considered an important cancer biomarker [36]. MCM7 has also been reported to be associated with hepatic carcinogenesis, esophageal squamous cell carcinoma, and prostate carcinoma [33,37,38]. Furthermore, MCM3 and MCM7 were demonstrated to be differentially expressed in primary T-cell cutaneous lymphoma skin samples [39]. MCM7 was also demonstrated to have higher levels of expression in skin proliferative disorders, such as keratoacanthoma, verruca vulgaris, and psoriasis vulgaris in comparison with normal skin [40]. ESR encodes estrogen receptor-alpha (ER-α), which is one of the potential targets in breast cancer treatment. It acts as a transcriptional regulator by interacting with estrogen and other proteins and has a role in breast cancer progression. ESR1 genetic variants are known contributors to breast carcinogenesis by inducing cell proliferation, genetic mutation, and necrosis [41]. The ESR-1 gene, encoding ESR-α, has been used as a target in therapy for breast cancer. Several miRNAs (miR-22, miR 301a-3p, miR206) were reported to suppress the expression of ER-α via direct binding with the 3’UTR (untranslated region) of ESR1 mRNA, which provided resistance to estrogen/ERα-targeted therapies [42,43]. Gao et al. have shown that expression of the ESR1 gene was regulated through a miRNA-mRNA regulatory pathway in ERα positive breast cancer [43]. They validated their results through a series of in silico studies and experimental analysis. Based on these findings, it was concluded that the miRNA targeted therapy emerged as a promising strategy for breast carcinoma. Moreover, the ERS1 gene has been shown to be differentially expressed in psoriasis, a chronic cutaneous inflammatory condition driven by IL-17 and IL-12/23 immune dysregulation [44]. As for SF3A2, a genome-wide screening in Drosophila and human cells showed the RNAi-mediated depletion of many different splicing factors (SF) to be associated with mitotic defects, ranging from abnormal chromosomal segregation to cytokinesis failure. SF are also involved in the mitotic division by disrupting the nuclear envelope. Pellacani et al. explored the role of SF3A2 in mitosis in Drosophila and humans reporting that the SF depletion influences the spindle formation and disrupts chromosomal segregation [45]. PPP2CA, protein phosphatase 2 catalytic subunit alpha gene, negatively controls cell growth and division. Also, in the study of Fang et al., it was proposed that PPP2CA played an important role in epidermis and hair follicles development in mice [46].

The gene enrichment analysis showed that ESR1, TP53, and PPP2CA were linked with integrin family cell surface interaction, endothelins, and immune pathways of IL-3- IL-5-, and IL-8 mediated signaling. IL-3 is a multi-lineage colony-stimulating factor produced by lymphoid cells and eosinophils [47]. IL-5 is an eosinophil colony-stimulating factor activating eosinophils, while IL-8 is one of the proinflammatory chemokines [48]. Moreover, Ghosh et al. also proposed that integrin-linked kinase (ILK) signaling and endothelin signaling were involved in the pathogenesis of AD [3]. While ILK activation may be connected with S. aureus colonization of AD skin, endothelin-1 plays a role in mediating itch, one of the most prominent symptoms of AD [3].

Our gene enrichment analysis also depicted an association of MCM7 and PPP2CA with the TNF alpha/NF-κB signaling pathway. NF-κB is a crucial effector molecule of inflammation and was proven to be involved in AD immune pathologic signaling [49].

Genome-wide association studies (GWAS) have revealed the association of FLG and KIF3A with atopic dermatitis, resulting in a disrupted skin barrier and reduced clearance of allergens, respectively [3]. Moreover, the meta-analysis of AD-derived transcriptome datasets highlighted the importance of inflammatory markers (MMP12) as well as the genes involved in T-cell immunity activation (Th2/CCL18, Th1/IFN/CXCL10, Th17/PI3/elafin, Th17/Th22 S100A7/A8/A9), epidermal proliferation (KRT16, Mki67, CLDN8), markers of atherosclerosis signalling (IL-37, selectin E), and downregulation of epidermal lipids (FA2H, FAR2, ELOVL3) [50]. Ghosh et al. analysed five publicly available AD transcriptome datasets and used clustering analysis to identify the genes that were differentially expressed in AD skin [51]. They performed validation experiments over the findings and demonstrated that genes involved in the epidermal barrier function, inflammation-related genes, proteases and protease inhibitor genes, as well as genes of antimicrobial responses, were differentially expressed in AD [51]. Our predictive genes are different from the ones previously identified. This might be due to the different datasets and samples used, and potentially due to our study focus on the identification of the genes involved in multiple interactions with proteins and miRNAs.

Several studies have been conducted exploring the miRNAs involved in AD pathogenesis. MicroRNAs are non-coding RNA molecules involved in post-transcriptional changes, which can negatively modify gene expression by degrading mRNAs or lowering their quantity [52]. Dysregulation of miR-143, miR-146a, miR-151a, miR-155 and miR-223 has been reported in AD patients [53]. These miRNAs play a role in the regulation of proliferation and apoptosis of keratinocytes, immune signalling (e.g. cytokine and nuclear factor-κB-dependent inflammatory responses), as well as the activity of Th-17 and Treg cells [53]. Moreover, the analysis of child AD datasets revealed miR-451 to be a potential predictive biomarker for early diagnostics [54]. Nousbeck et al. reported that miRNA-451a targets the interleukin 6 receptor (IL6R) and the proteasome subunit beta type-8 (PSMB8), while IL6R and PSMB8 showed a negative correlation with miR-451a levels in peripheral blood mononuclear cells (PBMCs) [54]. Furthermore, miRNA-155 expression was found to correlate positively with the severity of AD, Th-17 cell percentage, IL-17 mRNA expression and its plasma concentration [55]. In our study, we identified mir-155-3p, mir-151a-5p and mir-223-3p to be associated with p53 gene expression; these miRNAs were reported to play a role in the regulation of immune pathways in AD [53].

Tonacci et al. reviewed the role of different microRNAs in AD and ASD (Autism Spectrum Disorder) patients [56]. They reported that the main dysregulated miRNAs in AD are miR-186, miR-361-3p, miR-605, mir-150, mir-455-3p, and mir-302c-3p. Out of these miRNAs, our analysis identified miR-361-3p and mir-455-3p to be interacting with FFF1A1, mir-605 to be interacting with TP53, and mir-302c-3p to be interacting with MCM7.

From the metabolomic analysis conducted, we identified phosphatidylcholine diacyl, carnitine, acetylcarnitine and acylcarnitines to have either higher or lower abundance in AD patients. Those metabolites are associated with beta-oxidation of pristanoyl-CoA, peroxisomal lipid metabolism, and fatty acid metabolism pathways, which are essential for energy metabolism [57,58]. It is known that epidermis exhibits extensive levels of lipid metabolism to maintain the barrier function [59]. Furthermore, peroxisomes play a role in the proliferation, differentiation of cells as well as in the modulation of inflammatory responses [60]. Few metabolomic studies have been conducted in the context of AD. In the study by Huang et al. [61], which was focused on the metabolome of children with AD, it was reported that tryptophan and indolelactic acid (tryptophan metabolism pathway members), taurochenodeoxycholate/taurochenodeoxycholic acid, taurocholate/taurocholic acid (bile acid metabolism pathway members), free fatty acids and carnitines were increased in the AD cohort with higher reported IgE levels, revealing potential energy metabolism disruptions [61], which is in line with our findings. Furthermore, hydroxyl octadecadienoic acids, most of the hydroxy eicosatetraenoic acids (belonging to polyunsaturated fatty acids pathway), as well as glycocholate/glycocholic acid (belonging to bile acid metabolism pathways), were shown to be increased in AD versus healthy controls, independently of patient serum IgE levels (e.g. in cohorts of AD patients with both high or normal levels of serum IgE) [61]. Other studies proposed that the suppression of fatty acids beta-oxidation occurred in mice with AD [62], and mitochondrial dysfunction tended to decrease fatty acid oxidation and to be associated with higher serum levels of IgE [63].

The complex AD etiopathogenesis is characterized by specific microbiome patterns. The decreased diversity of microbial species with enhanced colonization of staphylococcal species, Staphylococcus aureus (S. aureus) and Staphylococcus epidermidis (S. epidermidis) in particular, is present in AD skin due to inflammation, disrupted skin barrier and reduced levels of antimicrobial peptides [64]. Recent findings demonstrated that overabundant S. epidermidis colonisation, which was previously believed to be beneficial, may have a damaging effect on the skin barrier [65]. Few studies also proposed that gut microbiome disturbances could be related to altered host immune function in AD [3].

Some of the limitations of this study include the lack of experimental validation (for example, quantitative polymerase chain reaction (qPCR) experiments for the identified hub genes) of our findings and the inclusion of a relatively low quantity of AD patient samples. Furthermore, the GEO datasets analysed were classified as AD skin versus healthy skin, with no differentiation between lesional AD and non-lesional AD skin samples, as well as high and normal IgE patient serum levels.

Conclusions

Using the multi-omics approach over AD patient samples datasets, we have identified several novel associations between genes regulating the cell cycle and cell proliferation. Further qRT-PCR validation studies are required to gain additional insights into the role of the identified hub genes in AD pathogenesis.

Disclosure of conflict of interest

None.

Abbreviations

AD

Atopic Dermatitis

ASD

Autism Spectrum Disorder

BH

Benjamini-Hochberg

DEGs

Differential expression genes

EN

Elastic Net

GEO

Gene expression data

LASSO

Least absolute shrinkage and selection operator

IMPaLA

Integrated Molecular Pathway-Level Analysis

PPI

protein-protein interactions

miRNA

MicroRNA

qPCR

Quantitative polymerase chain reaction