Integrative Analysis of Epidermal Barrier Dysfunction Regulators in Psoriasis

Abstract

Introduction

Psoriasis is a chronic inflammatory skin disease characterized by keratinocyte hyperproliferation and compromised epidermal barrier function. Although tight junction disruption is a key pathological hallmark, the upstream transcriptional mechanisms driving the downregulation of core barrier proteins such as claudin-1 (CLDN1) remain insufficiently characterized.

Methods

We implemented an integrative multi-omics framework combining weighted gene co-expression network analysis (WGCNA), single-cell RNA sequencing (scRNA-seq), and machine-learning algorithms to prioritize core diagnostic features in psoriasis. Upstream transcription factors regulating CLDN1 were computationally predicted using the JASPAR, hTFtarget, and KnockTF databases. The regulatory effect of E74-like ETS transcription factor 3 (ELF3) on the CLDN1 promoter was evaluated using dual-luciferase reporter assays, and the inflammatory response of the axis was examined in TNF-α-stimulated HaCaT keratinocytes via quantitative real-time PCR.

Results

Machine-learning prioritization identified CLDN1 as a robust diagnostic feature that is significantly downregulated in psoriatic lesions. Multi-database integration converged on ELF3 as a candidate negative regulator of CLDN1. Dual-luciferase reporter assays demonstrated that ELF3 overexpression significantly suppressed wild-type CLDN1 promoter activity, whereas mutation of the predicted binding site abolished this repression. In TNF-α-stimulated HaCaT cells, ELF3 was markedly upregulated with a concurrent downregulation of CLDN1.

Discussion: 

These findings suggest that ELF3 functions as a candidate transcriptional repressor of CLDN1 promoter activity and provide preliminary in vitro and computational insights into the transcriptional alterations associated with epidermal barrier disruption in psoriasis, highlighting potential avenues for restoring structural barrier integrity.

Introduction

Psoriasis is a chronic immune-mediated inflammatory dermatosis characterized by keratinocyte hyperproliferation, disturbed differentiation, and impaired epidermal barrier function.¹ Although biologics targeting the IL-23/IL-17 axis provide substantial clinical efficacy, recurrent flares following treatment cessation underscore the contribution of intrinsic barrier defects alongside immune dysregulation.^2,3

The physical barrier relies heavily on tight junction (TJ) complexes within the stratum granulosum, where CLDN1 serves as an essential regulator of paracellular permeability.⁴ In psoriatic lesions, CLDN1 exhibits delayed assembly, aberrant localization, and reduced expression, which compromise paracellular sealing and facilitate antigen penetration.⁵ However, whether this reduction is merely a secondary consequence of inflammatory cytokines or actively driven by upstream transcriptional repression in keratinocytes remains unclear.

To investigate the regulatory networks associated with barrier disruption, we integrated bulk transcriptomics and single-cell RNA sequencing (scRNA-seq). Co-expression network analysis (WGCNA) identified a core disease-associated module whose activity was predominantly concentrated within the keratinocyte population. Using regularized machine-learning algorithms and SHAP feature attribution, CLDN1 was prioritized among candidates as a top-ranking diagnostic feature consistently downregulated in lesions.

Single-cell pseudotime trajectory inference further demonstrated a progressive decline in CLDN1 expression as keratinocytes transition toward the psoriatic state, illustrating a kinetic deterioration of cellular identity and barrier function. In addition, regional association mapping via LocusZoom illustrated overlapping signals between psoriasis GWAS loci and skin-specific eQTLs at the CLDN1 region.⁶ Subsequent single-cell virtual knockout simulations computationally modeled a network shift from structural maintenance toward pro-inflammatory cascades.

To identify upstream regulators of CLDN1, we integrated motif, ChIP-seq, and perturbation datasets (JASPAR, hTFtarget, and KnockTF), which converged on ELF3 as a candidate repressor. ELF3 was first characterized as a novel epithelium-specific member of the ETS transcription factor family,⁷ and its expression in the epidermis is restricted to the most differentiated suprabasal layers.⁸ In keratinocytes, ELF3 plays dual roles in modulating terminal differentiation⁹ and has been implicated in epithelial inflammatory responses.¹⁰ ELF3 can function as both a transcriptional activator and repressor depending on promoter context,^9,11 and its transcriptional activity is modulated by interactions with coactivators and corepressors¹² and NF-κB in human keratinocytes.¹³ A comprehensive review highlighted the multiple roles of ELF3 in epidermal development and disease.¹⁴ These observations position ELF3 as a potential mechanistic link between inflammatory stimuli and the transcriptional silencing of barrier genes.

We subsequently evaluated this regulation in vitro: dual-luciferase reporter assays demonstrated that ELF3 represses CLDN1 promoter activity in a sequence-dependent manner, and TNF-α stimulation in HaCaT keratinocytes markedly upregulated ELF3 while suppressing CLDN1 expression. These findings outline an inflammatory-responsive ELF3-CLDN1 regulatory axis connecting cytokine exposure to tight junction downregulation.

By combining multi-omics profiling with preliminary in vitro assays, this study prioritizes the ELF3-CLDN1 axis as a candidate transcriptional pathway underlying barrier-associated molecular deficits in psoriasis, providing insights for restoring epidermal integrity beyond immunosuppression.

Materials and Methods

Data Acquisition and Preprocessing

All computational and statistical analyses were performed within the R software environment (version 4.3.0).¹⁵ Transcriptomic profiles were obtained from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/).¹⁶ Specifically, bulk RNA-sequencing datasets GSE54456 (comprising 82 normal control samples and 92 psoriatic samples)¹⁷ and GSE30999 (comprising 85 normal control samples and 85 psoriatic samples)¹⁸ were downloaded and utilized as the training and independent validation cohorts, respectively. Raw expression matrices were log²-transformed and quantile-normalized. For single-cell resolution mapping, the scRNA-seq dataset GSE151177¹⁹ derived from human skin tissues was acquired. To evaluate genetic causality, psoriasis Genome-Wide Association Study (GWAS) summary statistics²⁰ and skin tissue-specific expression quantitative trait loci (eQTL) datasets²¹ were specifically curated based on Asian populations to avoid population stratification bias.

Weighted Gene Co-Expression Network Analysis (WGCNA)

The WGCNA R package (version 1.73)²² was employed to construct a co-expression network. The goodSamplesGenes function was first applied to assess data integrity. Sample clustering based on Euclidean distance was performed to exclude potential outliers, ensuring data homogeneity. A soft-thresholding power (β) was selected where the scale-free topology fit index (R²) reached a plateau (R² > 0.85). A topological overlap matrix (TOM) was generated from the adjacency matrix to minimize the effects of noise and spurious associations. Genes were hierarchically clustered based on the TOM dissimilarity (1 - TOM), and co-expression modules were identified using the Dynamic Tree Cut algorithm (minimum module size = 30, merge cut height = 0.25). The MEblue module was identified as the primary pathological driver (r = 0.94). Scatter plots were used to visualize the relationship between Module Membership (MM) and Gene Significance (GS). Functional enrichment of the core module genes was performed using the clusterProfiler R package.²³

Single-Cell Transcriptomic Analysis and Activity Mapping

Single-cell data processing was executed using the Seurat R package (version 5.4.0).²⁴ Stringent quality control (QC) was applied: cells with fewer than 200 or greater than 5000 detected genes, or with >10% mitochondrial transcripts, were filtered out. The data were normalized, and the top 2000 highly variable genes were identified. Following Principal Component Analysis (PCA), Uniform Manifold Approximation and Projection (UMAP) was applied using the top 20 principal components to visualize the cellular landscape. Cell clusters were annotated based on the top 5 differentially expressed signature genes identified via the Wilcoxon rank-sum test. To assess the module activity at the single-cell level, the AUCell R package (version 1.28.0)²⁵ was utilized to calculate the Area Under the Curve (AUC) scores for the WGCNA MEblue gene set to assess module activity. Based on the AUC density distribution, a strict threshold of AUC > 0.17 was applied to isolate the specific “Active” cell subpopulation (primarily keratinocytes).

Pseudotime Trajectory Inference

The developmental transition of keratinocytes was reconstructed using the monocle3 R package (version 1.4.26).²⁶ The Seurat object was converted into a cell_data_set (cds) object. Following dimensionality reduction and cell clustering within Monocle3, the learn_graph() function was used to fit a principal graph. The order_cells() function was applied, defining the “Normal” keratinocyte population as the root nodes. Kinetic expression profiles (smoothed normalized Z-scores) of the top 10 signature genes were plotted against disease pseudotime (0 to 20) to track continuous expression dynamics across disease progression stages.

Machine Learning and Diagnostic Signature Identification

An intersecting core candidate set of 314 genes was identified from WGCNA hub genes and scRNA-seq active markers. To prioritize robust diagnostic features from this high-dimensional gene pool, a comparative machine-learning workflow was implemented using the Mime R package (version 0.1.3).²⁷ Within this framework, various combinations of regularized dimensionality-reduction and ensemble classification algorithms were evaluated across the discovery cohort (GSE54456) using nested 10-fold cross-validation. Among the tested pipelines, the StepCox and CoxBoost combination exhibited optimal feature shrinkage efficiency and stable discrimination across validation iterations.

It should be clarified that while StepCox and CoxBoost were originally established for survival analysis with time-to-event outcomes, their core mathematical architecture operates via component-wise gradient boosting and penalized partial-likelihood shrinkage, effectively functioning as high-dimensional regularized regression models for collinear variable contraction. In our static case-control binary setting (psoriasis vs normal control), binary class labels were parameterized with uniform synthetic coordinate weights strictly to mobilize the penalized shrinkage operators for feature ranking, rather than to model time-dependent survival hazard. This heuristic implementation served solely as an exploratory feature-prioritization tool to isolate robust core genes across the 314 overlapping candidates. Model interpretability and the directional contribution of each prioritized feature were subsequently resolved using the fastshap R package (version 0.1.1).²⁸ SHAP summary beeswarm plots and global importance bar plots were constructed to quantify individual gene impact, in which CLDN1 emerged as a prominent feature negatively associated with the psoriatic state. Finally, the discriminative performance of the finalized 10-gene signature and individual genes was assessed using Receiver Operating Characteristic (ROC) curves via the pROC R package (version 1.19.0.1)²⁹ in both the discovery cohort (GSE54456) and the locked external validation cohort (GSE30999).

Genetic Colocalization Analysis

To evaluate the regional genomic architecture at the CLDN1 locus, LocusZoom was utilized to generate regional association plots integrating Asian psoriasis GWAS summary statistics and skin-specific eQTL signals. Asian psoriasis GWAS summary statistics (P < 5 × 10⁻⁸ for genome-wide significance) were obtained from the GWAS Catalog (https://www.ebi.ac.uk/gwas/), while Asian skin-specific eQTL data were sourced from relevant eQTL databases. By evaluating regional linkage disequilibrium (LD) patterns and the spatial proximity of lead single nucleotide polymorphisms (SNPs) at the CLDN1 locus, we assessed whether psoriasis susceptibility signals map to the same genomic region regulating CLDN1 expression.

Single-Cell Virtual Knockout and Network Perturbation

The systemic regulatory impact of gene loss was simulated using the scTenifoldKnk R package (version 1.0.3).³⁰ Briefly, a single-cell Gene Regulatory Network (GRN) for the Wild-Type (WT) state was constructed using principal component regression (pcNet). A virtual knockout was then executed by setting the expression vector of CLDN1 to zero, followed by the generation of a KO-GRN. Non-linear manifold alignment was performed to compare the WT and KO networks. Downstream genes were ranked based on the magnitude of their regulatory structural shift, quantified as a Z-score. A threshold of Z ≥ 2 was applied to identify significantly perturbed effectors. clusterProfiler was subsequently used to perform GO enrichment analysis on these perturbed genes, contrasting the WT and KO functional landscapes.

Transcription Factor Prediction

Putative transcription factors regulating CLDN1 were identified by integrating three independent databases. JASPAR was used for motif-based prediction on the CLDN1 core promoter region (−2000 to +100 bp relative to the transcription start site).³¹ hTFtarget was queried for ChIP-seq physical binding evidence.³² KnockTF was utilized to retrieve transcription factors whose functional perturbation results in significant downstream CLDN1 expression changes (|log2FC| ≥ 2) across epithelial and relevant tissue models.³³ Transcription factors identified in at least two of the three databases were considered high-confidence candidates.

Dual-Luciferase Reporter Assay

A 2100-bp fragment of the human CLDN1 promoter (−2000 to +100) was amplified from genomic DNA and cloned into the pGL4.10 firefly luciferase reporter vector (Promega) to generate the wild-type construct. A mutant construct harboring targeted mutations within the predicted ELF3-responsive element was generated using a site-directed mutagenesis kit (Agilent) and verified by Sanger sequencing. The full-length human ELF3 coding sequence was synthesized and subcloned into the pcDNA3.1(+) expression vector (Invitrogen). All constructs were confirmed by DNA sequencing. HEK293T cells were cultured in DMEM supplemented with 10% fetal bovine serum. Cells were seeded into 24-well plates and co-transfected at 70–80% confluence with 400 ng of firefly reporter plasmid, 100 ng of effector plasmid (pcDNA3.1-ELF3 or empty vector control), and 10 ng of pRL-TK Renilla luciferase internal control plasmid using Lipofectamine 3000 (Invitrogen). Forty-eight hours after transfection, firefly and Renilla luciferase activities were measured using the Dual-Luciferase Reporter Assay System (Promega) on a GloMax luminometer. Relative luciferase activity was calculated as the ratio of firefly to Renilla luminescence and normalized to the vector control group. Data represent at least three independent experiments, each performed in triplicate, and are expressed as mean ± SD. Statistical comparisons were performed using a two-tailed unpaired Student’s *t*-test.

Cell Culture and TNF-α Stimulation

HaCaT human keratinocytes were obtained from the Cell Bank of the Chinese Academy of Sciences (Shanghai, China). HaCaT is a commercially available immortalized human keratinocyte cell line; therefore, no ethical approval or informed consent was required for its use in this study. Cells were cultured in DMEM supplemented with 10% fetal bovine serum and 1% penicillin/streptomycin at 37°C in a humidified atmosphere containing 5% CO₂. For TNF-α stimulation experiments, cells were seeded into 6-well plates at a density of 2×10⁵ cells per well and allowed to adhere overnight. Prior to stimulation, cells were serum-starved for 12 h in serum-free DMEM to minimize background signaling. Cells were then treated with 50 ng/mL recombinant human TNF-α (PeproTech) dissolved in phosphate-buffered saline for 24 h. An equal volume of PBS was added to control cells. After stimulation, cells were harvested for subsequent RNA or protein extraction. All experiments were performed in at least three independent biological replicates.

Quantitative Real-Time PCR

Total RNA was extracted from HaCaT cells using TRIzol reagent (Invitrogen) according to the manufacturer’s protocol. RNA concentration and purity were determined using a NanoDrop spectrophotometer. One microgram of total RNA was reverse-transcribed into cDNA using the PrimeScript RT reagent kit (Takara) in a 20 μL reaction volume. Quantitative real-time PCR was performed using SYBR Green Master Mix on an Applied Biosystems QuantStudio 5 Real-Time PCR System (Thermo Fisher Scientific, Waltham, MA, USA). Target gene accession numbers were retrieved from NCBI GenBank (NM_004433.5 for ELF3, NM_021101.5 for CLDN1, and NM_002046.7 for GAPDH). The primer sequences used were as follows: ELF3 forward, 5′-CAT GAC CTA CGA GAA GCT GAG C-3′ and reverse, 5′-GAC TCT GGA GAA CCT CTT CCT C-3′; CLDN1 forward, 5′-GTC TTT GAC TCC TTG CTG AAT CTG-3′ and reverse, 5′-CAC CTC ATC GTC TTC CAA GCA C-3′; GAPDH forward, 5′-GAA GGT GAA GGT CGG AGT C-3′ and reverse, 5′-GAA GAT GGT GAT GGG ATT TC-3′. The PCR cycling conditions were as follows: initial denaturation at 95°C for 3 min, followed by 40 cycles of 95°C for 15s and 60°C for 30s. Relative gene expression was calculated using the 2⁻ΔΔCt method with GAPDH as the internal reference gene. All reactions were performed in triplicate, and data represent at least three independent experiments. Results are expressed as mean ± SD.

Results

Identification of Psoriasis-Associated Co-Expression Modules

To identify critical gene networks driving psoriasis pathogenesis, WGCNA was performed on the transcriptomic data. Initial sample clustering based on Euclidean distance successfully identified and removed potential outliers, ensuring high data homogeneity for network construction (Figure 1A). A soft-thresholding power was determined at the point where the scale-free topology fit index (R²) reached a plateau above 0.85, satisfying the criteria for a scale-free network (Figure 1B). Hierarchical clustering based on topological overlap defined multiple distinct co-expression modules (Figure 1C). Evaluation of module-trait relationships revealed that the MEblue module possessed the most significant positive correlation with the psoriasis phenotype (Figure 1D). A robust correlation between MM and GS further indicated that genes centrally located within the blue module are highly relevant to the disease phenotype (Figure 1E). Functional enrichment analysis of this module highlighted key Gene Ontology (GO) and KEGG pathways, including “Nuclear division”, “Cell cycle”, and “Mitotic cell cycle phase transition”, which collectively reflect the hyper-proliferative state of keratinocytes in psoriatic lesions (Figure 1F). Taken together, these WGCNA findings reveal that the most prominent transcriptional alteration in active psoriatic lesions is concentrated within keratinocyte proliferation and differentiation gene networks, providing a defined modular framework for subsequent hub-gene screening.

Figure 1.

WGCNA identifying key gene modules associated with psoriasis. (A) Sample clustering to detect outliers. A hierarchical clustering dendrogram of samples was constructed based on Euclidean distance to identify and remove potential outliers, ensuring the homogeneity of the transcriptomic data for network construction. (B) Selection of the soft-thresholding power. The scale-free topology fit index (left, signed R²) and mean connectivity (right) were analyzed across various powers. A soft-thresholding power was selected where the R² reached a plateau (R²> 0.85), ensuring the construction of a scale-free network. (C) Gene dendrogram and module identification. Genes were clustered based on topological overlap, and co-expression modules were identified using the Dynamic Tree Cut method. Each color in the bottom row represents a distinct gene module. (D) Module-trait relationships heatmap. The correlation between module eigengenes and clinical traits (Control vs Psoriasis) was evaluated. The MEblue module shows the most significant positive correlation with the psoriasis phenotype, identifying it as the primary module of interest for further analysis. (E) Scatter plot of MM vs GS. A strong correlation (cor = 0.87) is observed between the membership in the blue module and the significance of genes for psoriasis. This confirms that genes highly connected within the blue module are also centrally involved in psoriasis pathogenesis. (F) Functional enrichment landscape. Gene Ontology (GO) and KEGG pathway enrichment analyses for the key module (blue module) genes. The top enriched terms, including “Nuclear division”, “Cell cycle”, and “Mitotic cell cycle phase transition”, reflect the hyper-proliferative state of keratinocytes in psoriatic lesions.

Single-Cell Mapping Localizes Hub Gene Activity to Keratinocytes

To dissect the cellular heterogeneity of the psoriatic microenvironment, single-cell transcriptomic profiling was visualized using UMAP. Seven major skin cell populations were annotated: Keratinocytes, T cells, Macrophages, Dendritic cells, Fibroblasts, Melanocytes, and Plasma cells (Figure 2A). Cell identity was validated by mapping the top 5 signature marker genes across each identified cluster (Figure 2B). Comparative analysis of cellular proportions demonstrated significant remodeling in diseased skin, characterized by a marked expansion of T cells, Macrophages, and Dendritic cells, alongside a relative decrease in the proportion of Keratinocytes (*P < 0.05, **P < 0.01) (Figure 2C and D). To contextualize the bulk RNA-seq findings, the activity of the WGCNA blue module hub genes was mapped across the UMAP space using AUCell scoring. The hub genes exhibited highly specific and concentrated activation exclusively within the Keratinocyte cluster (Figure 2E). An AUCell score threshold (> 0.17) based on density distribution was applied to isolate a highly active subpopulation of 3273 cells, confirming keratinocytes as the primary functional sites for the disease-associated module (Figures 2F, 3A and B).

Figure 2.

Single-cell transcriptomic landscape and cellular distribution of WGCNA hub genes in psoriasis. (A) UMAP visualization of the single-cell landscape. Dimension reduction analysis via UMAP identified seven major cell populations in skin tissues, including Keratinocytes, T cells, Macrophages, Dendritic cells, Fibroblasts, Melanocytes, and Plasma cells. (B) Marker gene expression across cell types. Dot plot illustrating the expression of the top 5 signature genes for each identified cell cluster. The dot size represents the percentage of cells expressing the gene, and the color intensity indicates the average expression level. (C) Cellular proportion shifts. Stacked bar plot comparing the relative abundance of cell types between the Control and Psoriasis groups, highlighting significant remodeling of the cellular composition in diseased skin. (D) Dynamics of global cell proportions. Box plots showing the statistical comparison of cell type frequencies. A significant decrease in the proportion of Keratinocytes and a marked expansion of T cells, Macrophages, and Dendritic cells are observed in psoriatic lesions (*P < 0.05, **P < 0.01, ns, not significant). (E) Activity mapping of WGCNA hub genes. Feature plot showing the AUCell scores of the WGCNA blue module hub genes across the UMAP space. The hub genes exhibit highly specific and concentrated activity within the Keratinocyte cluster. (F) Distribution of hub gene activity scores. Histogram and density plots of the AUC scores for the WGCNA hub gene set. The red vertical line indicates the threshold (AUC > 0.17), identifying a specific subpopulation of cells (3273 cells, primarily keratinocytes) with high expression of the psoriasis-related gene module.

Figure 3.

Identification of a robust diagnostic signature via integration of single-cell activation mapping and machine learning algorithms. (A) Module activation frequency across cell types. Quantitative analysis of the WGCNA blue module (psoriasis-related) activation frequency. Keratinocytes and Fibroblasts exhibit the highest proportions of “Active” status, confirming these cells as the primary functional sites of the identified hub genes. (B) UMAP projection of cellular activation status. Spatial distribution of cells with high WGCNA hub module activity. Red dots indicate “Active” cells, which are predominantly localized within the Keratinocyte cluster, demonstrating high cell-type specificity. (C) Venn diagram of candidate gene intersection. Intersection between WGCNA hub genes (derived from bulk-seq) and active markers (derived from scRNA-seq). 314 genes (10.5%) were identified as the overlapping core gene set, ensuring both population-level significance and cell-level specificity. (D) Performance evaluation of machine learning models. Heatmap illustrating the diagnostic accuracy (AUC) of various machine learning algorithm combinations (eg, StepCox, Lasso, CoxBoost, Ridge, RSF) across the training (GSE54456) and validation (GSE30999) datasets. The StepCox + CoxBoost model was selected for its superior and stable performance. (E) Gene importance assessment via SHAP. Top: SHAP beeswarm plot illustrating the impact of prioritized genes on disease prediction. Points are color-coded by relative feature value (purple for high expression, yellow for low expression). In this multivariate regularized model, high feature values of CLDN1 and UPP1 predominantly yield negative SHAP values pushing the classification toward the control state, whereas low expression levels contribute positively to the prediction of psoriasis, consistent with the observed downregulation of CLDN1 in psoriatic lesions. Bottom: Bar plot ranking the top 10 genes by global feature importance. CLDN1 and UPP1 rank among the top critical features in the diagnostic model, highlighting their pivotal diagnostic value in the molecular pathology of psoriasis.

Machine Learning Prioritizes CLDN1 as a Robust Diagnostic Feature

Intersecting the bulk-derived WGCNA hub genes with the scRNA-seq active markers yielded a highly specific core gene set comprising 314 candidates (10.5% overlap) (Figure 3C). To screen core candidate genes from the overlapping hub set, we evaluated several regularized feature-selection combinations across the discovery (GSE54456) and validation (GSE30999) cohorts (Figure 3D). Stepwise Cox regression coupled with gradient boosting (StepCox + CoxBoost) demonstrated stable feature contraction and ranking efficiency. SHAP attribution analysis was subsequently employed to quantify the marginal conditional contribution of prioritized genes within the multivariate model (Figure 3E). CLDN1 and UPP1 were identified as the top two contributors in the global importance hierarchy. Specifically, lower expression of CLDN1 contributed positively to the predicted probability of psoriasis, concordant with its role as a compromised barrier feature. For UPP1, while its raw univariate expression is markedly elevated in psoriatic lesions (as detailed below in Figure 4A), its conditional attribution in the multivariate model reflects adjusted marginal weights under high-dimensional collinear shrinkage. Together, these prioritized features demonstrate robust diagnostic and discriminatory power.

Figure 4.

External validation of the SHAP-selected top 10 diagnostic signature and its clinical performance. (A) Expression validation of the top 10 diagnostic signature. Box plots showing the relative expression levels (Z-score) of the 10 core genes identified by SHAP analysis across the training cohort (GSE54456) and the independent validation cohort (GSE30999). Most genes, including UPP1, and CCNE1, exhibit significant and consistent upregulation in psoriatic lesions compared to healthy controls, whereas CLDN1, PARD3 shows marked downregulation (****P < 0.0001). (B) ROC analysis in the training cohort (GSE54456). ROC curves evaluating the diagnostic efficacy of individual genes versus the integrated model. The combined Top-10 model achieves an ideal AUC of 1.000. Notably, CLDN1 and UPP1 individually show exceptional diagnostic performance with AUC values of 1.000 in this cohort. (C) ROC analysis in the validation cohort (GSE30999). Independent validation of the diagnostic model’s performance. The combined model maintains high diagnostic accuracy with an AUC of 0.991. Individually, CLDN1 demonstrates robust performance with an AUC of 0.940, confirming its high reliability as a molecular biomarker for psoriasis across different patient populations.

External validation confirmed the relative expression patterns (Z-score) of the top 10 core genes, showing significant and consistent upregulation of UPP1 and CCNE1, and marked downregulation of CLDN1 and PARD3 in psoriatic lesions (****P < 0.0001) (Figure 4A). ROC analysis demonstrated exceptional diagnostic efficacy. The combined 10-gene model achieved an AUC of 1.000 in the training cohort and 0.991 in the independent validation cohort. Individually, CLDN1 maintained robust performance with an AUC of 0.940 in the validation set, cementing its reliability as a molecular biomarker (Figure 4B and C). In summary, the machine-learning pipeline prioritized CLDN1 as an invariant, highly ranked diagnostic feature across independent patient cohorts, supporting the concept that tight junction downregulation represents an essential molecular hallmark of psoriasis and justifying its prioritization for downstream regulatory analysis.

Pseudotime Trajectory and Regional Association Mapping of CLDN1

The dynamic expression patterns of the signature genes during disease progression were modeled using Monocle3 pseudotime trajectory inference on the keratinocyte population. The continuous developmental path illustrated a clear transition from normal cells toward the psoriatic state (Figure 5A). Kinetic expression profiling along the continuous pseudotime path revealed distinct temporal dynamics among the top signature genes (Figure 5B and C). Structural and tight junction genes, including CLDN1, SMPD3, and PARD3, exhibited a steady and progressive decline, illustrating a gradual loss of terminal barrier differentiation as keratinocytes transition into the pathological state. In contrast, hyperproliferation and metabolic regulators such as UPP1 and CCNE1 showed marked and consistent induction, while CARHSP1 and PGAM1 displayed biphasic activation with peak expression during the intermediate transition phase (Figure 5C).

Figure 5.

Single-cell trajectory analysis revealing the dynamic expression of top 10 signature genes during psoriasis progression. (A) Pseudotime trajectory of keratinocytes. Trajectory reconstruction was performed using Monocle3 to visualize the transition of keratinocytes. The UMAP plot is color-coded by pseudotime (0 to 20), illustrating a continuous developmental path. The black line represents the inferred trajectory, originating from normal cells and progressing toward the psoriatic state. (B) Heatmap of gene expression dynamics across disease stages. Expression patterns of the top 10 SHAP-selected signature genes are displayed along the disease progression continuum (Normal, Transition, Psoriasis). The heatmap highlights distinct clusters of genes that are progressively up- or down-regulated, reflecting the molecular remodeling of keratinocytes during the transition from healthy skin to psoriatic lesions. (C) Kinetic expression profiles of the top 10 signature genes. Smoothed normalized expression (Z-score) of individual genes plotted against disease pseudotime. Stages are categorized into Normal (blue), Transition (beige), and Psoriasis (pink). Downregulated barrier-associated trends: Tight junction and structural genes such as CLDN1, SMPD3, and PARD3 exhibit a progressive decline along the pseudotime continuum, reflecting gradual loss of barrier integrity and differentiation defects as keratinocytes progress toward the psoriatic state. Upregulated and dynamic trends: Pro-inflammatory, metabolic, and cell cycle-related genes, including UPP1, CARHSP1, EZH2, and CCNE1, show elevated or stage-specific kinetic profiles, reflecting metabolic reprogramming and hyperproliferative activation during disease evolution.

To assess the genomic context of CLDN1, regional association plots were constructed using Asian psoriasis GWAS summary statistics and skin-specific eQTL data (Figure 6). The regional association signals for psoriasis susceptibility mapped to the vicinity of the CLDN1 locus alongside local eQTL associations, illustrating overlapping genetic architecture at this chromosomal region.

Figure 6.

Regional association and genetic architecture at the CLDN1 locus. (A-E) Regional LocusZoom plots illustrating overlapping association signals between psoriasis GWAS loci and skin eQTLs at the CLDN1 region. Genomic positions are shown on the horizontal axis (Mb). Association statistical significance is plotted on the vertical axis as –log₁₀(P). The red diamond indicators highlight the lead (sentinel) SNPs at each locus, illustrating the spatial positioning of the most significant disease-susceptibility and expression-associated variants within the local linkage disequilibrium region.

Virtual Knockout Reveals Systemic Functional Remodeling

The systemic impact of CLDN1 loss was evaluated through a single-cell virtual knockout simulation using the scTenifoldKnk algorithm. By reconstructing a Gene Regulatory Network (GRN) and simulating CLDN1 deletion, genes were ranked by the magnitude of their regulatory shift (Z-score). CLDN1 exhibited the most substantial regulatory shift, driving massive downstream shifts in associated network genes such as FTH1, PLCG2, and small proline-rich proteins (SPRR2D, SPRR2A) above the Z ≥ 2 threshold (Figure 7A). Correlation analysis confirmed that genes with high initial absolute regulatory weight (connectivity) to CLDN1 in the Wild-Type (WT) network suffered the most substantial reorganization upon knockout (Figure 7B). Gene Ontology (GO) enrichment analysis revealed a stark functional transition: the WT network primarily governed barrier stability (“collagen-containing extracellular matrix”, “focal adhesion”), whereas the CLDN1-KO network shifted toward inflammatory remodeling, including “positive regulation of leukocyte activation” and “cytokine activity” (*P < 0.05, **P < 0.01, ***P < 0.001). Functional enrichment analysis of the perturbed genes indicated that in silico depletion of CLDN1 shifted the network from structural terms toward immune-related terms, such as “leukocyte activation” and “cytokine activity” (Figure 7C), providing a computational model of network-level barrier disturbance. Specifically, functional annotation of the differential networks (Figure 7C) highlighted a marked shift: while the wild-type network was predominantly enriched in structural stability and cell-matrix integrity terms such as “focal adhesion” and “collagen-containing extracellular matrix”, the simulated CLDN1-knockout network exhibited significant enrichment in immune-stimulatory and inflammatory cascades, including “positive regulation of leukocyte activation” and “cytokine activity” (*P < 0.05, **P < 0.01, ***P < 0.001). This computational transition illustrates how CLDN1 functional deficiency at the network level is predicted to trigger a shift from structural barrier maintenance toward a pro-inflammatory microenvironment.

Figure 7.

Systemic Regulatory Network Perturbation and Functional Remodeling Following Virtual Knockout of CLDN1. (A) Global regulatory perturbation landscape upon CLDN1 knockout. The scTenifoldKnk R package was utilized to perform a virtual knockout of CLDN1 by reconstructing a single-cell Gene Regulatory Network (GRN) and simulating its deletion. Genes are ranked along the x-axis based on the magnitude of their regulatory shift (Z-score). CLDN1 exhibits the highest perturbation, followed by top-ranked downstream effectors such as FTH1, PLCG2, and members of the small proline-rich protein family (SPRR2D, SPRR2A). A threshold of Z ≥2was applied to identify significantly perturbed genes. (B) Correlation between original connectivity and perturbation impact. Left: The bar plot displays the absolute regulatory weight (connectivity) of the top perturbed genes with CLDN1 in the Wild-Type (WT) network. Right: The corresponding Z-scores represent the magnitude of the regulatory shift in the Knockout (KO) state. This comparison reveals that genes with high initial connectivity to CLDN1 (eg, FTH1, PLCG2, MTRNR2L1) undergo the most substantial regulatory reorganization upon the loss of CLDN1 function. (C) Differential functional landscapes between Wild-Type and CLDN1-KO states. Gene Ontology (GO) enrichment analysis highlights the top 5 terms with significant differences across Biological Process (BP), Cellular Component (CC), and Molecular Function (MF). Wild-Type network (blue bars): Functional activity is primarily associated with barrier stability and innate defense, including “regulation of innate immune response”, “focal adhesion”, and “collagen-containing extracellular matrix”. CLDN1-KO network (red bars): The virtual knockout shifts the functional focus toward inflammatory processes and adhesion remodeling, such as “positive regulation of leukocyte activation”, “positive regulation of cell-cell adhesion”, and “cytokine activity” (*P < 0.05, **P < 0.01, ***P < 0.001).

Integrative Prediction and in vitro Evaluation of ELF3 as a Transcriptional Repressor of CLDN1

To elucidate the upstream regulatory mechanisms driving CLDN1 downregulation in psoriasis, we performed a comprehensive transcription factor (TF) prediction by integrating data from three independent databases: JASPAR (motif sequence), hTFtarget (ChIP-seq physical binding), and KnockTF (functional perturbation). This cross-validation yielded key regulatory candidates, including ELF3, SOX4, and STAT3 (Figure 8A). We next evaluated their expression profiles in bulk RNA-seq cohorts. Across both the GSE54456 and GSE30999 datasets, ELF3 and STAT3 were significantly upregulated in psoriatic lesions compared to healthy controls, whereas SOX4 was downregulated (Figure 8B). To determine the functional relationship between these TFs and CLDN1, we performed Pearson correlation analysis. ELF3 exhibited the most robust and consistent negative correlation with CLDN1 expression across both independent cohorts, strongly indicating its role as a transcriptional repressor, whereas STAT3 and SOX4 showed weaker or inconsistent correlations (Figure 8C). Single-cell RNA sequencing (scRNA-seq) mapping further confirmed the distinct spatial expression landscape of these TFs, localizing ELF3 within relevant cellular populations (Figure 8D).

Figure 8.

Computational prediction and in vitro evaluation of ELF3-mediated CLDN1 promoter repression. (A) Venn diagram illustrating the integration of candidate transcription factors targeting CLDN1 across three independent databases: JASPAR (Sequence motif), hTFtarget (ChIP-seq evidence), and KnockTF (Perturbation data). (B) Expression levels of candidate TFs (ELF3, SOX4, STAT3) in bulk RNA-seq cohorts GSE54456 and GSE30999. Boxplots demonstrate that ELF3 is consistently and significantly upregulated in psoriatic tissues compared to healthy controls (****P < 0.0001). (C) Scatter plots showing the correlation between candidate TF expression and CLDN1 expression. ELF3 exhibits a strong and consistent negative correlation with CLDN1 across both GSE54456 and GSE30999 datasets, supporting its role as a repressor. (D) UMAP visualizations displaying the single-cell expression landscape of ELF3, SOX4, and STAT3. (E) Dual-luciferase reporter assay evaluating sequence-dependent transcriptional repression. Overexpression of ELF3 (ELF3 OE) significantly reduces CLDN1 promoter activity compared to the empty vector. Mutating the predicted ELF3 binding site (Mut + ELF3 OE) rescues luciferase activity, confirming sequence-dependent promoter repression (****P < 0.0001, ns, not significant). (F) In vitro validation of the inflammatory regulatory axis. qPCR analysis reveals that TNF-α stimulation significantly upregulates ELF3 mRNA expression while concurrently downregulating CLDN1 expression, mimicking the psoriatic pathological state (****P < 0.0001).

To experimentally validate this regulatory axis, we conducted a dual-luciferase reporter assay. Overexpression of ELF3 (ELF3 OE) significantly suppressed the transcriptional activity of the wild-type CLDN1 promoter. Crucially, mutation of the predicted ELF3 binding motif abrogated this inhibitory effect, indicating sequence-dependent transcriptional repression of the CLDN1 promoter (Figure 8E). Finally, In HaCaT keratinocytes, TNF-α treatment significantly increased ELF3 transcript levels while reducing CLDN1 expression (Figure 8F). Together, these observations suggest that ELF3 can act as a candidate transcriptional repressor targeting CLDN1 promoter activity under inflammatory conditions.

Discussion

While current biologic therapies targeting the IL-23/IL-17 axis have significantly improved the clinical management of psoriasis, disease recurrence upon treatment cessation remains a persistent challenge. This highlights the necessity of addressing underlying epidermal structural defects alongside immune dysregulation. In this study, we integrated multi-omic analyses and computational modeling to explore the regulatory role of the skin barrier in psoriasis. Our findings indicate that CLDN1, a critical component of tight junctions, serves as a key susceptibility gene whose progressive downregulation during keratinocyte differentiation may exacerbate local inflammation.

Our initial WGCNA and scRNA-seq spatial mapping localized the primary psoriasis-associated gene module (MEblue) predominantly within the keratinocyte population (Figures 1 and 2). This observation supports the perspective that keratinocytes actively participate in the local inflammatory milieu.³⁴ Crucially, our multi-layered analysis revealed a consistent expression deficit of CLDN1 throughout disease progression. Bulk tissue analysis and machine learning models (StepCox + CoxBoost) identified CLDN1 as a significantly downregulated core diagnostic feature in psoriatic lesions (Figures 3 and 4). Concurrently, Monocle3 pseudotime trajectory inference further confirmed that as individual keratinocytes transition into the pathological state, CLDN1 expression exhibits a dynamic and progressive decline (Figure 5). This highly consistent downregulation across both macroscopic tissue levels and microscopic single-cell trajectories strongly indicates that individual cells fail to properly mature and express functional barrier proteins during the pathological process.³ This cellular-level deficit of physical barrier components underscores the functional impairment of the tight junction seal in diseased skin.

To explore whether this consistent downregulation of CLDN1 is a primary genetic risk factor or merely a secondary response to inflammation, we performed genetic colocalization. The alignment of GWAS susceptibility loci with skin-specific eQTL signals at the CLDN1 locus (Figure 6) suggests a genetic basis for its abnormally low expression in psoriasis.⁶ Building on this, our virtual knockout simulation using scTenifoldKnk (Figure 7) provided computational insights into the systemic consequences of CLDN1 functional loss. The simulation predicted a notable regulatory shift from structural maintenance (eg, focal adhesion) to immune-related processes (eg, leukocyte activation and cytokine activity). While these findings are derived in silico, they align closely with in vivo observations where CLDN1-deficient models exhibit severe barrier dysfunction,⁴ computationally supporting the hypothesis that compromised physical barriers can act as potent amplifiers for inflammatory cascades.

To elucidate the upstream mechanisms driving this genetically and pathologically linked downregulation, we shifted our focus to transcriptional regulatory networks. By integrating sequence motif predictions (JASPAR), physical binding evidence (hTFtarget), and perturbation databases (KnockTF), we identified ELF3 as a high-confidence candidate. ELF3 is an epithelium-specific transcription factor highly restricted to differentiating suprabasal keratinocytes, known to play dual roles in regulating terminal differentiation and mediating inflammatory responses. Our bulk cohort analysis and single-cell mapping confirmed that ELF3 is significantly upregulated in psoriatic lesions and is strongly negatively correlated with CLDN1 expression. Importantly, our dual-luciferase reporter assays provided in vitro evidence that ELF3 overexpression significantly suppresses wild-type CLDN1 promoter activity in a sequence-dependent manner (Figure 8). Furthermore, TNF-α stimulation in keratinocytes induced ELF3 while concomitantly reducing CLDN1 transcript levels (Figure 8). These observations suggest a candidate regulatory connection between inflammatory signaling and the transcriptional downregulation of tight junction genes in psoriasis.

Several limitations of this study must be acknowledged. First, the feature prioritization was conducted using retrospective GEO cohorts representing established chronic plaques; real-world clinical utility across early or atypical lesions remains to be determined. Second, while the dual-luciferase reporter assay supported sequence-dependent promoter repression by ELF3, this experiment was carried out in HEK293T cells; formal confirmation of physical binding in keratinocytes via chromatin immunoprecipitation (ChIP-qPCR) or electrophoretic mobility shift assays (EMSA) is still required. Third, our study lacks functional rescue experiments—such as ELF3 knockdown or overexpression in cytokine-exposed keratinocytes—to definitively prove whether inflammatory CLDN1 downregulation is mediated exclusively by ELF3. Another limitation is that while we documented significant transcriptional suppression of CLDN1, this study did not include functional biophysical assays (such as TEER measurements, paracellular tracer flux, or confocal immunofluorescence tracking CLDN1 membrane localization). Therefore, our conclusions are based on molecular and transcriptomic alterations rather than direct physiological proof of permeability barrier disruption. These mechanistic and functional questions will be the subject of future experimental investigations.

Conclusion

In summary, our integrated computational and experimental analysis highlights the key regulatory importance of CLDN1 deficiency in the pathogenesis of psoriasis. By combining insights gained from single-cell trajectory analysis with predictions of transcriptional activity based on multiple databases and in vitro validation, we determined that the ELF3-CLDN1 axis represents a candidate mechanism linking inflammatory stimuli to the breakdown of the physical barrier. This study provides a comprehensive, data-driven framework that enhances our understanding of barrier dysfunction in psoriasis and suggests that targeting upstream transcriptional repressors of structural proteins could represent a promising complementary approach to current immunosuppressive therapies.

Funding Statement

No funding was received for this study.

Abbreviations

WGCNA, Weighted Gene Co-expression Network Analysis; scRNA-seq, single-cell RNA sequencing; SHAP, Shapley Additive Explanation; CLDN1, Claudin-1; TJ, tight junction; GO, Gene Ontology; GRN, Gene Regulatory Network; WT, Wild-Type; BP, Biological Process; CC, Cellular Component; MF, Molecular Function; SNPs, single nucleotide polymorphisms; GWAS, Genome-Wide Association Study; eQTL, expression quantitative trait loci.

Data Sharing Statement

Data requests should be directed to Dr. Wenxin Wang.

Ethics Declarations

This study analyzed publicly available, de-identified human transcriptomic and genomic datasets from the Gene Expression Omnibus (GEO) repository (accession numbers: GSE54456, GSE30999, and GSE151177) and commercially available immortalized cell lines (HaCaT and HEK293T). According to Items 1 and 2 of Article 32 of the Measures for Ethical Review of Life Science and Medical Research Involving Human Subjects promulgated by the National Health Commission of China (February 18, 2023), research utilizing legally and publicly available human data or anonymous data that does not involve identifiable personal information or biological risk is exempt from local Institutional Review Board (IRB) review.

Author Contributions

Yingdi Liu and Yinglin Wang contributed equally to this work. All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors declare no competing interests.