Integration of GWAS and single-cell analysis for psoriasis identifies disease-associated cell subtypes and therapeutic targets

Abstract

Psoriasis is a heritable, common chronic autoimmune disorder characterized by cycles of remission and flare-ups. Here, we jointly analyze genomic and single-cell transcriptomic data to elucidate the genetic and molecular architecture of psoriasis. We perform a large-scale genome-wide meta-analysis of individuals of European ancestry (n = 1,131,685) and identify 125 independent susceptibility loci associated with psoriasis, including 17 previously unreported loci. Integrating these findings with single-cell transcriptome data identifies the predominant roles of myeloid and T cells in psoriasis and the enriched expression of disease-associated genes in keratinocytes and endothelial cell subsets. Finally, we prioritize 50 potential therapeutic target genes using multiple robust approaches and identify their cell subtype-specific activity patterns across non-lesional, lesional, and treatment conditions, thereby revealing layer-specific cross-cell-type interactions in psoriasis. These findings provide perspectives regarding the pathogenesis of psoriasis and highlight the role of immunometabolism in disease progression.

Integrating genetic data and skin single-cell data highlights psoriasis-associated cell types and cell type-specific functional mechanisms. These findings offer perspectives for identifying therapeutic targets for psoriasis at the cellular level.

Introduction

Psoriasis is a common chronic autoimmune disorder affecting approximately 2% of individuals of European descent. This condition is associated with various comorbidities, including psoriatic arthritis (PsA) in approximately 30% of patients and an increased risk of cardiovascular diseases such as atherosclerosis, hypertension, and myocardial infarction^1–4. Previous studies have shown that individuals with psoriasis have a higher prevalence of metabolic disorders, including obesity, diabetes, and hyperlipidemia than individuals without psoriasis^5–7. Notably, psoriasis is characterized by genetic factors. Previous twin studies indicated that 68% of the variation in psoriasis susceptibility can be attributed to genetic influences, underscoring the potential of genetic studies to elucidate the etiology of the disease and guide therapeutic strategies⁸.

Considering this growing interest, numerous genome-wide association studies (GWASs) have been conducted to identify psoriasis risk loci using increasing sample sizes, enabling the discovery of additional susceptibility variants^9–11. Notably, large-scale population-based datasets such as the UK Biobank (UKB), FinnGen, and 23andMe, Inc. have facilitated the exploration of genetic relationships across various phenotypes, including diseases and lifestyles^12–15. Meta-analyses integrating these extensive GWAS resources are crucial for increasing the statistical power to identify more disease-associated variants, thereby facilitating the interpretation of the genetic underpinnings of psoriasis through functional annotations.

Advances in single-cell sequencing technologies have enabled the generation of single-cell transcriptome data from the skin tissues of psoriasis patients^16,17. This approach allows the examination of individual gene expression profiles in various cell types including keratinocytes, fibroblasts, and diverse immune cells such as myeloid and lymphoid cells, facilitating the identification of rare disease-associated cell populations that were previously undetectable in bulk tissue analyses^18,19. Single-cell analyses have provided insights into the continuum of cell states within the skin, with trajectory inference methods serving as crucial tools for understanding transitions between states²⁰. The co-expression of gene networks within specific cell subtypes may play functionally significant roles by contributing to shared biological pathways. Therefore, the identification of cell type-specific gene network modules enhances functional interpretation. Therapeutic target selection benefits significantly from this approach because targeting cell type-specific genes minimizes off-target effects and improves both accuracy and safety compared to targeting genes expressed across multiple cell types. In summary, although GWAS facilitate the exploration of broad genetic susceptibility across large populations, integrating single-cell data enables a more in-depth examination of the causal genetic variant effects²¹. Recent studies have attempted to understand psoriasis mechanisms and to identify therapeutic target genes^11,22,23. However, most studies have analyzed single-cell gene expression data and genetic information separately and have not fully integrated multiple data modalities, including spatial context and cell-cell interactions, to understand the underlying biological mechanisms.

In this study, we utilized a joint analysis framework that integrated the large-scale GWAS and single-cell data. First, we conducted a GWAS meta-analysis of psoriasis in over one million individuals, representing the largest psoriasis GWAS to date. Next, we integrated these GWAS results with single-cell transcriptome data from skin samples from patients with psoriasis and healthy donors. This integration enabled us to examine the cell type-specific expression of psoriasis-associated genes and assess the relative polygenic disease enrichment in particular cell subpopulations. Finally, we prioritized potential drug target genes for psoriasis by leveraging insights from both GWAS and single-cell data. The aggregation of various methods to identify causal genes yields a robust list of candidate genes as therapeutic targets. Recently, immunotherapies employing biologics have demonstrated remarkable efficacy in the treatment of autoimmune diseases^24,25. We investigated how the functions of these prioritized genes changed during IL-23-targeted treatment in the skin microenvironment. This framework offers additional therapeutic possibilities for psoriasis at the cell type level and has potential applicability to other diseases.

Results

Genome-wide meta-analysis of psoriasis

A genome-wide meta-analysis of psoriasis in European populations was conducted by integrating four large-scale GWAS summary statistics, comprising data from over one million individuals: UKB (n = 408,619), FinnGen r10¹⁵ (n = 407,876), Stuart et al.¹⁰ (n = 44,161), and 23andMe (n = 271,029) (Fig. 1 and Supplementary Data 1). Following the approach outlined by Tsoi et al.⁹, Duffy’s method²⁶ was applied to correct for effect sizes and effective sample sizes in the UKB and 23andMe datasets, considering that these datasets included self-reported data (see Methods). Using the UKB samples of European ancestry, we compared polygenic risk scores (PRS) between ICD-diagnosed psoriasis cases, self-reported cases, and controls. Self-reported cases showed lower PRS values than ICD-diagnosed cases but remained enriched relative to controls, supporting the inclusion of self-reported cases with appropriate correction (Supplementary Fig. 1). Based on the risk allele frequencies in case samples and control samples, the total genetic liability across cases was consistent with estimated true-positive proportions of 73.85% for UKB cases and 37.5% for 23andMe cases, reflecting the weaker genetic signal of self-reported cases (Supplementary Data 2). After adjusting for the effect size, the effect size distributions of known psoriasis risk variants aligned closely with those reported in previous studies (Supplementary Fig. 2). To assess the impact of adjusting for self-reported phenotype definitions, we calculated PRSs in UKB samples using 23andMe psoriasis GWAS summary statistics before and after applying Duffy’s adjustment (Supplementary Fig. 3). The adjusted GWAS statistics resulted in a modest but consistent improvement in the PRS predictive performance. Adjustments for sample sizes yielded a meta-analysis comprising 41,983 cases and 1,089,702 controls, covering 10,339,620 variants with minor allele frequencies (MAFs) >1%. Based on the genome-wide significance threshold of p < 5 × 10⁻⁸, our meta-analysis identified a total of 198 independent SNPs from conditional and joint (COJO) association analysis across 125 distinct loci, including 17 previously unreported loci (Fig. 2 and Supplementary Fig. 4). Excluding the major histocompatibility complex (MHC) region, a total of 159 COJO SNPs were mapped to 122 loci (Supplementary Data 3). This GWAS included the largest sample size for psoriasis among the published studies, and a comparison of our GWAS with previous psoriasis GWASs based on the effective sample size (Supplementary Data 4) revealed an approximately linear increase in the number of significant distinct loci (Supplementary Fig. 5). Genes mapped to the identified psoriasis-associated loci were significantly enriched in functions related to the immune response (p = 4.34 × 10⁻¹⁵), mononuclear cell differentiation (p = 8.01 × 10⁻¹⁵), and cytokine-mediated signaling pathway (p = 1.81 × 10⁻¹³) in the MAGMA gene-set analysis (Supplementary Data 5). Previously unreported loci were mapped to genes implicated in immune-related functions (IRF2BP2, CXCR4, ULBP3, CRTAM, and SOCS3), mitochondrial functions (SDHD and BCO2), and cell growth and differentiation (CDC7 and PIK3R5). These loci reached genome-wide significance levels in the meta-analysis but not in any individual GWAS (Supplementary Fig. 6). For example, rs10191360 (chr2:136884679:C:T), mapped to CXCR4, exhibited consistent effect directions across all four GWAS datasets and achieved genome-wide significance (p = 1.77 × 10⁻⁸). This variant has been associated with multiple sclerosis in previous GWAS²⁷, suggesting its potential role in autoimmune disease pathogenesis.

Fig. 1. Overview of this study.

Genome-wide meta-analysis of four GWAS summary statistics for psoriasis was conducted. With single-cell data, modality-specific analyses and joint analysis such as genetic correlation with various diseases/traits, pseudotime trajectory, and scDRS were performed. Fifty genes were prioritized by applying additional gene prioritization approaches to genes expressed in disease-relevant cell types. The temporal modulation of the functional activity of these prioritized genes was examined using single-cell data from individuals during IL-23 targeted treatment. Created in BioRender. Won, H. (2026) https://BioRender.com/t23qghv.

Fig. 2. A karyotype plot for the psoriasis meta-GWAS result.

Significant loci identified from the joint and conditional analysis of the psoriasis meta-GWAS. Previously known loci and unreported loci are marked with yellow circle and cyan squares, respectively. Local Analysis of [Co]Variant Association (LAVA) analysis provides local heritability of predefined genetic regions. We used 2495 LD blocks derived from European 1000 Genomes data. The local heritability estimates are represented by the color scale of the karyotype regions. Local blocks with statistically significant local heritability (n = 132) are ranked by heritability, and psoriasis-associated significant loci within each block are highlighted in cyan (top panel). The red-outlined box represents the major histocompatibility complex (MHC) region which was excluded in the analysis.

Estimating genomic discovery potential in GWAS

Local heritability analysis was performed to assess the identification of unreported loci associated with psoriasis. Regions with high local heritability predominantly contained previously identified loci (Fig. 2). Notably, significant SNPs were also identified in the regions with low local heritability (Supplementary Data 6). We examined the relationship between the heritability of local LD blocks and the presence of significant SNPs within the blocks (Supplementary Fig. 7A). The local LD significance was associated with an increased likelihood of observing significant SNPs (chi-square test p = 2.42 × 10⁻⁸⁰). We first examined 42 regions in which LD local heritability was not significant but contained significant SNPs (referred to as LAVA-NS independent SNPs). When we compared these with 106 independent SNPs located in 72 LD blocks that showed statistically significant heritability (referred to as LAVA-S independent SNPs), we identified that LAVA-NS independent SNPs exhibited smaller effect sizes and lower statistical significance (Supplementary Fig. 7B, C). Next, 60 significant LAVA regions where no genome-wide significant SNPs were identified were categorized into four classes for each LD block to explore the potential for identifying additional loci (Supplementary Fig. 8A, see “Methods”). Except for nine MHC region-adjacent blocks, 13 of 51 (25%) blocks were previously reported as psoriasis-susceptibility loci in the GWAS Catalog across seven studies (Supplementary Fig. 8B, C and Supplementary Data 7). Notably, LD blocks classified as having high discovery potential showed a higher likelihood of harboring psoriasis-associated variants (5 of 15 (33%)). The LD block with the highest local heritability among regions not yet associated with psoriasis (chr10:118,416,050–119,528,730), which contains RPL21P16, PLPP4, and WDR11 genes, has previously been reported to be associated with metabolic traits such as HDL cholesterol, BMI, systolic blood pressure, prostate-specific antigen, and type 2 diabetes^28,29. These findings suggest that regions with high local heritability for psoriasis yet lacking significant loci may have the potential for additional discoveries.

Functional enrichment analysis and genetic correlation with diverse phenotypes

To explore the functional enrichment of psoriasis-associated loci, stratified LD score regression (LDSC) analysis was performed using tissue- and cell type-specific gene expression and chromatin modification annotations (see Methods). Psoriasis-associated loci exhibited significant enrichment in blood/immune-related tissues and cell types, with the highest enrichment observed in the synovial fluid (p = 9.17×10⁻⁹) (Fig. 3A and Supplementary Data 8). This result may be derived from the study cases, including individuals with PsA³⁰. Additionally, chromatin modification of T helper 17 (Th17) cell types showed the most significant enrichment (p = 1.15 × 10⁻¹²), followed by digestive system tissues such as the colon, esophagus, and sigmoid colon, suggesting shared mechanisms with inflammatory bowel disease (Supplementary Data 9).

Fig. 3. Genetic correlation analysis and functional enrichment of psoriasis.

A Tissue and cell type enrichment analysis using the stratified LD score regression (LDSC) analysis for psoriasis. Gene expression (top) and chromatin modification (bottom) of multiple tissues and cell types are used in this analysis. The dashed lines indicate Benjamini-Hochberg false discovery rate (FDR) thresholds (FDR < 0.05), and p-values were calculated based on one-sided tests. Points in dashed rectangles are the top-10 significant results and are represented in the right bar plots with details. B Genetic correlation of psoriasis with curated diseases/traits. p-values were calculated based on two-sided z-tests. The diseases/traits with Bonferroni-adjusted p-values < 0.05 are shown. Error bars represent the 95% confidence intervals calculated as r_g ± 1.96 × standard error. Details for used summary statistics are provided in Supplementary Data 10.

Genetic correlations between psoriasis and 113 traits were assessed using curated GWAS summary statistics, and genetic correlation analyses were conducted using the linkage disequilibrium score regression analysis (Supplementary Data 10). Among these, 41 diseases/traits exhibited significant genetic correlations with psoriasis after Bonferroni correction (p < 0.05/113; Fig. 3B). Notable positive correlations were observed with cardiometabolic and autoimmune diseases, including type 2 diabetes (r_g = 0.240; p = 7.97 × 10⁻¹²), ischemic heart disease (r_g = 0.207; p = 4.29 × 10⁻⁹), and inflammatory bowel disease (r_g = 0.284; p = 8.09 × 10⁻⁹), consistent with findings from previous observational studies³¹. In addition, psoriasis had a significant negative correlation with HDL cholesterol (r_g = −0.126; p = 7.8 × 10⁻⁶), whereas no associations were found with LDL cholesterol (r_g = 0.001; p = 0.981) and total cholesterol levels (r_g = −0.015; p = 0.608). Psychiatric conditions such as depressive disorder (r_g = 0.150; p = 8.03 × 10⁻¹³) and attention-deficit/hyperactivity disorder (r_g = 0.220; p = 4.27 × 10⁻⁹), as well as subjective health satisfaction (r_g = 0.244; p = 1.09 × 10⁻¹¹; coded higher values indicate greater dissatisfaction), also exhibited strong positive genetic correlations. Allergic traits, including allergies (r_g = 0.103; p = 0.034) and atopic dermatitis (r_g = 0.272; p = 0.0005), showed positive genetic correlations with nominal significance (p < 0.05).

Joint analysis of GWAS and single-cell data for psoriasis

Single-cell RNA sequencing (scRNA-seq) data from skin biopsies of patients with psoriasis and healthy controls were analyzed (see Methods). The dataset comprised samples from 14 patients and 8 healthy individuals, including 14 lesional (PP), 11 non-lesional (PN), and 8 normal skin (NS) samples. All samples were integrated into a single object with batch correction, and all cells were displayed in the shared Uniform Manifold Approximation and Projection (UMAP) space (Fig. 4A). Canonical markers were used to identify 10 major skin-resident cell types (Fig. 4B). Keratinocytes constituted the majority of cells in skin tissues, with fibroblasts being the second most abundant cell type in the dermal layer (Supplementary Data 11). Skin-resident immune cells, including myeloid cells, T cells, and mast cells, were present in relatively small proportions.

Fig. 4. Joint analysis of single-cell transcriptomics and GWAS data.

A UMAP plot of processed single-cell RNA-seq data including non-lesions from healthy controls (NS), non-lesions from psoriasis patients (PN), and psoriatic lesions (PP). B Dot plot for showing expression of representative cell markers. C Cell subtypes identified within each major cell type. D Lollipop plot displaying relative tissue enrichment. We compared the number of non-lesion cells (NS + PN) and lesion cells (PP) in each cell subtype and calculated the ratio of observed to expected (Ro/e). Values are shown as −1 × (non-lesion Ro/e) when non-lesion Ro/e > 1. Asterisks denote the false discovery rate (FDR) < 0.05 (two-sided permutation test, 10,000 permutations). E Cell subtype-level scDRS result. The outward bars represent FDR-adjusted one-sided p-values of the cell subtype-disease association. The inward bars represent FDR-adjusted within-cell subtype association heterogeneity. Bars with FDR < 0.10 were drawn with darker colors and thicker boundaries. F Two keratinocyte lineages (left), UMAP with RNA velocity (center), and FDR-adjusted p-values of cells based on one-sided Monte Carlo tests (right). UMAP and z-scaled scDRS scores derived from one-sided Monte Carlo tests for myeloid (G) and T cell (H) are shown.

Psoriatic keratinocytes exhibited significantly elevated expression of S100A7, S100A8, and S100A9 genes previously implicated in chronic inflammation and macrophage infiltration into the epidermal layer (Supplementary Data 12). A total of 55 cell subtypes were identified by independently analyzing the major cell types (Fig. 4C and Supplementary Fig. 9). Keratinocytes were categorized into five clusters based on their epidermal layer characteristics and enrichment in skin tissues (Supplementary Data 13). Notably, keratinocyte cluster 1 (KER_c1) prominently expressed KRT5 and KRT14, which are basal layer markers. Keratinocyte cluster 3 (KER_c3), cluster 2 (KER_c2) and cluster 4 (KER_c4) were characterized by expression of KRT1 and KRT10, which are markers of the spinous and granular layers (Supplementary Fig. 10). Specifically, the terminal subpopulations of the KER_c2 and KER_c4 clusters showed unique expression of the FLG gene, indicating that the terminal pseudotime points represent the outermost layer of the epidermis undergoing cornification. Additionally, KER_c2 was enriched with cells derived from lesion tissues, whereas KER_c4 was enriched with cells from normal tissues (Fig. 4D and Supplementary Data 14). These findings suggest that keratinocytes derived from psoriatic lesions show a higher proportion of differentiation toward a distinct lineage compared with normal-derived keratinocytes (Fig. 4F). The continuous trajectory observed in the UMAP plot, interpreted as pseudotime, reflects the spatial organization within the epidermal layers (Supplementary Fig. 11). Additionally, RNA velocity analysis identified the basal keratinocytes with high unspliced proportions, indicating an active transcription potential in less-differentiated states (Supplementary Fig. 12). Although the velocity arrows showed some noise due to sample heterogeneity, the overall directionality was consistent with the inferred pseudotime trajectories (Fig. 4F).

A single-cell disease relevance score (scDRS) analysis was then performed to integrate the processed scRNA-seq data with the GWAS summary statistics (Supplementary Data 14). Consistent with previous findings³², all T cell and myeloid subtypes exhibited strong disease enrichment based on the expression levels of GWAS-associated genes (Fig. 4E). Notably, a subcluster of endothelial cells (EC_c1), primarily expressing CDKN1A, and a subcluster of keratinocytes (KER_c1) showed modest significance (false discovery rate (FDR) < 0.10). Interestingly, keratinocytes undergoing cornification in the granular layer, enriched in normal tissues, showed higher disease relevance scores (Fig. 4F). Within-cell type heterogeneity refers to the presence of a consistent disease relevance score among cells within a cluster, without distinct differences between individual cells (Fig. 4E). While all myeloid and T cells exhibited significant associations, Langerhans cells (LCs) and cycling myeloid cells displayed higher disease relevance scores (Fig. 4G). Among the T cells, Tc17 cells (T_c3) and helper T cells (T_c0) had elevated scDRS scores. Notably, Tc17 cells exhibited low within-cell-type heterogeneity associations, indicating uniform characteristics across all cells within this subtype (Fig. 4E, H).

Potential implications of within-cell type heterogeneity

We further examined the within-cell-type heterogeneity of the scDRS scores across epidermal cell types. Cell types with significant scDRS heterogeneity showed larger differences in scDRS score distribution between normal and lesional skin (Supplementary Fig. 13A, B and Supplementary Data 15). Among the keratinocyte populations, the KER_c4 cluster exhibited higher scDRS scores in lesional skin than in normal skin, indicating increased disease relevance in the lesional state. We observed that cells with significant scDRS scores were preferentially localized in the outermost layers of the skin (Fig. 4F), consistent with the heterogeneity within the defined cell populations. Among the ECs, the EC_c1 cluster showed marginal scDRS significance at the cluster level. Within this cluster, a subset of cells displayed particularly high scDRS scores (Supplementary Fig. 13C, D). These high-scDRS cells showed increased expression of genes involved in the AGE-RAGE signaling pathway and negative regulation of TGF-β receptor signaling (Supplementary Fig. 13E, F), and the enrichment of these pathways increased progressively with higher scDRS scores (Supplementary Fig. 13G).

Robust gene prioritization for therapeutic targets based on multi-omic information

We identified candidate drug target genes for psoriasis by integrating GWAS summary statistics with single-cell transcriptomics data. Five criteria were applied: positional mapping (distance), gene-level significance (MAGMA), transcriptome-wide association study (TWAS), co-localization of GWAS signals with expression quantitative trait loci (eQTL), and single-cell expression (see “Methods”). As a result, 110 genes meeting at least four of these criteria were considered candidate drug targets and were further filtered using summary-based Mendelian randomization to assess the causal effects on psoriasis through gene expression in the skin or blood tissues (Supplementary Data 16). Finally, a total of 50 genes were prioritized as potential therapeutic targets for psoriasis (Fig. 5).

Fig. 5. Gene prioritization as psoriasis treatment targets.

The plot illustrates the gene prioritization process and the prioritized genes. Putative causal genes were filtered based on five criteria derived from GWAS summary statistics and single-cell expression. From left to right, (1) nearest genes of GWAS independent SNPs; (2) significant genes from MAGMA; (3) transcriptome-wide association study (TWAS) genes significant in at least one tissue of skin or whole blood; (4) Colocalization posterior probability (PP) > 0.8 in at least one tissue of skin or whole blood; and (5) genes expressed in at least 10% of cells in the interested cell types (Keratinocyte, Myeloid, T cell, and Endothelial cell). Genes meeting at least four criteria were retained. Among the 110 genes, 50 genes were further validated through summary-based Mendelian randomization (two-sided p < 0.05/110, Bonferroni correction). Druggability tiers from Open Targets Tractability were categorized into five tiers and displayed accordingly.

Notably, some prioritized genes exhibited tissue-specific regulatory patterns. For RGS14, NEK6, PRDX5, and FKBPL, the same alleles associated with increased psoriasis risk were linked to decreased expression levels in skin eQTL datasets, but increased expression in blood eQTL datasets. In contrast, for PRSS53, the risk allele was associated with increased expression in the skin and decreased expression in the blood. These results highlight the importance of tissue-specific gene regulation (Supplementary Data 17). Among these prioritized genes, four (IL23A, VKORC1, TYK2, and IL2RA) were associated with FDA-approved drugs, and approximately 50% of them are currently under investigation in the preclinical or clinical stages (Fig. 5 and Supplementary Data 18).

Cell type-specific psoriasis-associated gene co-expression networks

Prioritized genes identified as central to psoriasis pathogenesis are expected to exhibit specific functions at the individual cell-type level within the skin. To elucidate cell type-specific functions and interactions among gene sets in each cell type, we performed gene co-expression network analyses in keratinocytes, myeloid cells, and T cells. Focusing on co-expression modules containing at least two prioritized genes, we identified four modules (KER-M1 to KER-M4) in keratinocytes, six modules (MYE-M1 to MYE-M6) in myeloid cells, and two modules (T-M1 and T-M2) in T cells (Fig. 6A). In keratinocytes, the four modules were enriched in specific epidermal layers, respectively. For instance, module M1 exhibited the highest module eigengene score (module activity) in the basal layer cells, whereas module M3 showed elevated activity in the outermost cell subpopulation undergoing differentiation from the granular to the corneal layer (Fig. 6B–D). In myeloid cells, MYE-M2 and MYE-M5 were enriched in specific myeloid subtypes. Among myeloid cells, MYE-M2 showed specific enrichment in LCs (specificity = 39.6%), whereas MYE-M5 exhibited the highest specificity in type-2 conventional dendritic cells (cDC2A) (specificity = 58.2%; Supplementary Data 19). In summary, 32 prioritized genes were members of key modules associated with psoriasis (Fig. 6D). Importantly, within the MYE-M2 network, IL23A displayed notably high connectivity, as defined by strong Pearson correlations with other genes in the module, indicating its role as a hub gene. This finding suggests that IL-23–related pathways may play a pivotal role in disease pathogenesis through interactions between LCs and their adjacent cells (Fig. 6E).

Fig. 6. Cell type-specific gene co-expression modules associated with risankizumab treatment.

A Cell type-specific co-expression gene network identification using hdWGCNA, which identified 13, 49, 30 modules for keratinocyte, myeloid, and T cell, respectively. Each expressed gene was assigned to a single module. The bar plot represents the number of assigned genes that were defined as therapeutic targets. B Module eigengenes (MEs) of keratinocyte module 1 to module 4 (KER-M1 to KER-M4). C MEs of keratinocyte modules change over the pseudotime of two keratinocyte trajectories. Dots represent the ME score of single cells. Smoothed curves represent the locally estimated scatterplot smoothing (LOESS) algorithm-fitted mean values, and shaded bands represent the 95% confidence interval. scDRS scores of keratinocyte cells were plotted together (Dashed purple curve). D Cell type enrichment of 12 modules having two or more prioritized genes. The enrichments represent the proportion of average MEs for all cell subtypes in the major cell type. E Module memberships and connectivities of prioritized genes. Genes assigned to at least one of the 12 modules are shown. F Overview of procedure about risankizumab treatment data for psoriasis from Francis et al. G Temporal enrichment changes of MEs for non-lesion (NL) and lesion (L) skins over treatment. For each module, values were zero-centered on the Day 0 baseline NL condition and aggregated as pseudobulk expression per patient, using the five patients (n = 5) as replicates at each time point. Points and error bars represent the mean and 95% confidence intervals across pseudobulk replicates. H A correlation network for 12 modules from 20 samples of Francis et al. I Gene-set pathway analysis (one-sided hypergeometric test) for five modules of interest. Representative pathways with FDR-adjusted p-values < 0.10 were illustrated.

We further analyzed the MYE-M2 module to characterize lesion-specific functional changes (Supplementary Fig. 14). Among the 895 MYE-M2 member genes, those co-expressed with IL23A (Topological overlap matrix pairwise similarity > 0.10) included established Langerhans cell markers (CD207, HLA-DQB2) as well as genes involved in lipid metabolism (ACOT7, LPAR3) (Supplementary Fig. 14A, B). Next, we calculated module connectivity separately in normal and lesional samples and compared the fold changes between the two conditions (Supplementary Fig. 14C). Pathway analysis of genes with decreased connectivity in lesional samples revealed the enrichment of the lipoxin (LX) biosynthesis pathway (adjusted p-value = 0.020) (Supplementary Data 20). This lower connectivity reflects a disruption of coordinated biological functions within the network. Given the established role of the lipoxin pathway in immunoregulation and wound healing³³, its disruption suggests that this anti-inflammatory mechanism may be impaired in psoriatic lesions. This dysregulation might contribute to the persistent inflammation and defective tissue repair characteristic of the disease.

Changes in module activity during IL-23 inhibitor treatment were investigated by analyzing temporal scRNA-seq data from five patients undergoing risankizumab treatment, as reported by Francis et al.¹⁷ (Fig. 6F and Supplementary Fig. 15). The dataset comprised non-lesional and lesional skin collected at baseline (day 0), as well as lesional skin collected at 3 and 14 days after treatment (days 3 and 14), with each patient contributing samples across all time points. To evaluate their therapeutic dynamics, we projected this independent treatment cohort onto pre-defined gene modules defined in the Ma et al. dataset (see Methods). Following the projection, module activity was compared between lesional and non-lesional samples collected at the initial visit (Fig. 6G and Supplementary Fig. 16). Notably, all keratinocyte modules, except KER-M3, exhibited hyperactivity in lesions at baseline, with further increases observed during treatment. Although KER-M3 initially displayed hypoactivity, its activity increased progressively following risankizumab administration. Correlation analyses of module activity revealed coordinated patterns of change across samples (Fig. 6H). For instance, KER-M1 enriched in basal layer keratinocytes clustered with MYE-M5 (cDC2A-specific) and T-M1, suggesting interactions between IL2RA-expressing T cells and cDC2A (Supplementary Data 21). These modules were characteristically enriched in processes such as extracellular matrix (ECM) remodeling and cell adhesion, indicating direct ECM-mediated interactions or crosstalk (Fig. 6I and Supplementary Data 22–26). Additionally, cornification and lipid metabolism processes specific to granular layer keratinocytes appeared positively regulated alongside LC-mediated T-cell activation during treatment.

Cross-cell-type communication analysis enables the identification of gene-level interactions between cell types. Using the MultiNicheNet framework, we performed differential analyses of ligand-receptor pairs to identify interactions that were differentially active between non-lesional and lesional skin, as well as between baseline lesional and treatment conditions (Supplementary Fig. 17A). The MYE-M5 module, which showed the strongest association with cDC2A, also exhibited a significantly enhanced interaction with the basal keratinocyte-associated module KER-M1. Notably, the DSC2-DSG3 interaction showed increased activity in lesional samples, suggesting potential cellular proximity or interaction between cDC2A and basal keratinocytes. Interactions between cDC2A and T cells were also upregulated in lesional skin, including CD80-CTLA4 interactions, reflecting a proposed cDC2A-mediated modulation of immune activity in the basal epidermal layer. Spatial transcriptomic validation using an independent psoriasis dataset provided additional context consistent with these predictions, showing close spatial proximity between basal keratinocytes and cDC2A specifically in lesional tissue (Supplementary Fig. 17B). cDC2A-associated spots were detected in most samples, and spatial co-localization between basal keratinocytes and cDC2A showed a significant positive association with Psoriasis Area and Severity Index (PASI) scores (Supplementary Fig. 17C). To assess the consistency across multiple datasets and the generalizability of the findings, we analyzed the Kim et al. dataset³⁴, which included samples of healthy controls and patients with psoriasis collected at week 28 after risankizumab treatment (Supplementary Fig. 18 and Supplementary Data 27). This dataset exhibited distinct cellular expression profiles compared to other single-cell datasets, including clearer separation of dendritic and other myeloid cells and the clustering of endothelial and fibroblast populations (Supplementary Fig. 18A–C). Using scDRS analysis, we revealed significant enrichment of genes associated with psoriasis in dendritic cell and T cell clusters, consistent with the patterns observed in the Ma et al. dataset (Supplementary Fig. 18D). Additionally, similar module-level cell-cell interaction patterns between lesional and non-lesional skin were observed in the long-term risankizumab treatment dataset (Supplementary Fig. 18E). Notably, unlike the relatively modest associations observed in short-term treatment samples, the long-term dataset revealed a significant shift of gene expression in post-treatment lesional samples toward control levels.

Discussion

A large-scale genome-wide association study encompassing over one million individuals of European ancestry identified reliable susceptibility risk variants for psoriasis. To our knowledge, this is the largest GWAS on psoriasis to date. We acknowledge the variation in psoriasis case definition among individual GWAS, which was confirmed by the difference in risk allele frequencies. By applying Duffy’s approach to the UKB and 23andMe data, as used in previous studies, the effect size was successfully adjusted, and the proportion of psoriasis cases was appropriately calibrated to approximately 2%. Based on two PRS analyses, we demonstrated the rationale for including self-reported cases and showed that these cases can be appropriately incorporated into GWAS analyses using Duffy’s adjustment method. A recent study revealed that higher genetic risk was linked to more severe psoriasis³⁵. This suggests potential heterogeneity in disease severity between self-reported and clinically diagnosed cases, implying that Duffy’s approach could be an approach to account for such heterogeneity arising from binary classification. Our analysis revealed that higher local genetic correlations within linkage disequilibrium (LD) blocks tended to correspond to a greater number of significant SNPs, suggesting that this scheme could be effectively applied in other studies. Furthermore, a subset of LD blocks showed high local heritability but lacked genome-wide significant signals. This observation suggests that further investigations are required to elucidate their biological relevance.

Utilizing psoriasis GWAS data, we employed recently developed post-GWAS techniques to examine functional aspects and genetic correlations with disease and trait categories. Our results underscore the pivotal role of T helper 17 (Th17) cells in psoriasis and highlight the significant heritability enrichment in the synovial fluid and immune-related cell types. Notably, in addition to associations with other autoimmune diseases, genetic correlations with cardiometabolic syndromes and psychiatric disorders were observed, which is consistent with previous research^2–4,36. Significant genetic overlap with diverse diseases suggests that a shared genetic architecture influences multiple conditions. Compared to previous GWAS-based studies, our analyses enabled a more precise and expanded assessment of the genetic relationships between psoriasis and other traits^11,22,23. With the increased statistical power of our GWAS, we detected statistically more robust genetic associations with conditions such as myocardial infarction, asthma, and anxiety disorders. Additionally, our genetic correlation analysis included a broader range of traits that have not been examined in earlier studies. We identified previously unreported genetic correlations with traits, such as attention deficit/hyperactivity disorder (ADHD), urate levels, health satisfaction, and time spent watching television. Time spent watching television is a form of leisure-time sedentary behavior that has been previously shown to be causally associated with cardiovascular disease³⁷. The observed positive genetic correlations between sedentary behavior, cardiovascular disease, and psoriasis may reflect shared genetic factors underlying these phenotypes.

Integration of GWAS and scRNA-seq data demonstrated that a joint analysis can yield valuable insights. Single-cell transcriptome data not only facilitate the discovery of previously unrecognized cell subtypes but also preserve the microenvironmental context within tissues. scRNA-seq data from both healthy and psoriatic skin samples, including lesional and non-lesional samples, were processed to determine the involvement of specific cell types in psoriasis. The scDRS algorithm enabled the identification of specific cell types potentially involved in the development of psoriasis by calculating a genetically driven disease relevance score for each cell type. Using the scDRS algorithm, strong associations between myeloid cells and T cells were consistently observed. We also identified subpopulations of keratinocytes and ECs that may be involved in psoriasis development. Notably, keratinocytes exhibited a continuous expression pattern from the basal to granular layers, reflecting the spatial organization of the epidermis. Specifically, cells in the outermost nucleated epidermal layer, which is characterized by pronounced cornification features, showed the strongest association with psoriasis. This implies that genetic susceptibility to psoriasis is fundamentally linked to the dysregulation of biological functions within the later-stage epidermal layer, reflecting its potential causal role in the pathogenesis of the disease. The EC_c1_CDKN1A cluster, which was enriched in normal skin tissue, showed marginal significance in psoriasis (FDR = 0.069). These cells were characterized by the specific expression of NFKBIA and SOCS3, which suppress cytokine responses and regulate inflammatory signaling via negative feedback mechanisms (Supplementary Data 13)³⁸. These findings suggest that dysfunction of this specific subset of endothelial cells may contribute to the pathogenesis of psoriasis. Within this cell cluster, cells with higher scDRS scores formed a distinct subcluster. This subpopulation showed higher expression of genes involved in the AGE-RAGE signaling pathway and negative regulation of TGF-β receptor signaling. Previous studies have reported that AGE-RAGE interactions are implicated in psoriasis and can suppress downstream TGF-β-mediated functions^39,40. Given the marginal statistical significance and increased analytical complexity, we excluded endothelial cells from downstream analyses. Further research will provide deeper insights into the role of endothelial cells in modulating immune cell function in psoriasis.

Using various gene prioritization methods that integrated GWAS and single-cell expression data, potential genetic targets for psoriasis treatment were identified. While traditional GWAS-based approaches have shown promise^41,42, filtering for genes with cell type-specific expression relevant to the disease may enhance screening precision and reduce unforeseen side effects⁴³. This selection was further refined by identifying the genes whose expression levels in the skin or whole blood had a causal relationship with the development of psoriasis. Although these strict approaches may reduce sensitivity in the screening of potential drug targets, they enhance specificity, allowing for a more efficient and precise prioritization of key targets.

Biological therapies targeting the IL-17/IL-23 axis have recently gained prominence in psoriasis treatment^44,45. Remarkably, risankizumab, an IL-23 inhibitor, has shown significant efficacy in clinical trials⁴⁶. In this study, we utilized additional longitudinal scRNA-seq data from patients undergoing risankizumab treatment. By projecting the data onto gene modules defined from untreated psoriatic skin, we captured dynamic changes in disease-relevant transcriptomic programs. We identified two key interaction clusters central to psoriasis pathogenesis. The first involved the KER-M1 module, enriched in basal layer keratinocytes, which interacted with cDC2A-specific MYE-M5 and T-M1 modules. Psoriasis is characterized by hyperproliferation of basal layer keratinocytes mediated through positive feedback interactions with inflammatory immune cells^47,48. cDC2A cells are a major source of IL-23, which activates IL-17-producing T helper subsets, thereby inducing the pathogenesis of psoriasis^49,50. KER-M1 is involved in ECM construction and cell adhesion; risk variants affecting the genes related to these functions may compromise the epidermal barrier and potentially facilitate immune cell infiltration⁵¹. The second interaction involved the KER-M3 module, enriched in the granular layer and the corneal layer keratinocytes interacting with LCs. IL23A, a gene critical to the IL-17/IL-23 axis, is predominantly expressed in LCs and cycling myeloid cells. This suggests that resident LCs may play a significant role in the pathogenesis of psoriasis by contributing to keratinocyte cornification. A recent study has reported that sustained increases in skin lipid content can impair LC autophagy functions, leading to excessive IL-23 expression⁵². This indicates that focusing on the regulation of immunometabolism, such as lipid metabolism, between keratinocytes and LCs may be a crucial strategy for psoriasis treatment and prevention. Furthermore, recent findings from a Mendelian randomization analysis that lipid-lowering drugs may causally reduce the risk of psoriasis reinforce the importance of lipid metabolism in disease management⁵³.

In summary, we propose two epidermal layer-specific pathogenic mechanisms in psoriasis based on integrative analyses of various IL-23 inhibition-based multimodal datasets (Supplementary Fig. 19). In the basal epidermal layer, enhanced physical interactions between keratinocytes, cDC2A, and T cells are associated with localized immune suppression, which may contribute to the excessive proliferation of basal keratinocytes. In the granular epidermal layer, disrupted interactions between keratinocytes and Langerhans cells are associated with the dysregulation of lipid metabolism in both cell types.

This study provides insights that may inform and serve as a foundation for further research. First, we employed three large-scale single-cell datasets specifically designed for psoriasis research. Increasing the scale of single-cell datasets may improve statistical power and enable a more comprehensive characterization of disease-associated cell types through integration with GWAS results, including those showing marginal significance in scDRS analysis, such as keratinocytes in the basal layer (KER_c1_KRT15, FDR = 0.065) and a subset of endothelial cells (EC_c1_CDKN1A, FDR = 0.069). Additionally, while previous studies have highlighted the role of fibroblasts in psoriasis progression¹⁶, this study did not identify any fibroblast subpopulations that exhibited significant disease enrichment from the GWAS signals. This trend was consistently confirmed by using an independent validation dataset (Supplementary Fig. 18D). Our study fundamentally aimed to investigate the genetic basis of psoriasis using a GWAS-based approach. One possible hypothesis is that the genetic risk for psoriasis is closely linked to the inherent functional programs of immune cells, keratinocytes, and endothelial cells. After disease onset, these genetically susceptible cells may secondarily (i.e., downstream) influence the gene expression profiles of fibroblasts within the skin tissue. In this context, the observed fibroblast-related changes may represent a consequence of psoriasis rather than a primary causal driver. Further experimental validation is required to confirm this hypothesis. For example, epidermal organoid models with genetic perturbation could be established to assess the effects of genetically susceptible cells on fibroblast transcriptional programs. Secondly, a systematic gene prioritization strategy was implemented to identify genes with putative causal roles in psoriasis. However, genes that did not meet prioritization criteria or were classified as least druggable may still have therapeutic potential. For example, IRF5 met all prioritization criteria but was classified as least druggable in current druggability databases. Nonetheless, IRF5 is a key regulator of autoimmune diseases, and many studies have actively explored IRF5 blockade as a therapeutic strategy^54,55. Similarly, Late Cornified Envelope (LCE) genes, which play a critical role in keratinocyte cornification within the outermost layer of psoriatic skin, did not meet our expression-based prioritization criteria⁵⁶. However, their potential as therapeutic targets should be further explored using complementary functional approaches.

This study has several fundamental limitations. First, the GWAS analysis was conducted exclusively on datasets of European ancestry, and additional investigations are required to evaluate the applicability of these findings to other populations. Second, a subset of psoriasis patients in our study likely has comorbid PsA. This may account for the observed functional enrichment within tissues such as synovial fluid, rather than reflecting a systemic or cutaneous pathogenesis. As previous studies explored the shared and distinct genetic architectures between cutaneous psoriasis and PsA, our results should be interpreted with caution, particularly in distinguishing cutaneous signals from PsA-driven effects^57,58. Third, many publicly available single-cell datasets did not provide comprehensive demographic information, including age and sex, which precluded us from performing a sex-stratified analysis. Furthermore, the datasets used in this study were heavily skewed toward male participants. Consequently, there are inherent limitations in the generalizability of our findings, as the analysis could not fully account for sex-specific effects. Previous research has indicated that male patients often exhibit greater disease severity compared to females⁵⁹. Moreover, studies have reported a significant link between sex hormones and immune regulation, noting that high estrogen levels correlate with clinical improvement, whereas low testosterone in males is associated with increased disease severity^60,61. Therefore, future investigations using large-scale and well-annotated single-cell cohorts are needed to elucidate the effects of demographic variables on psoriasis pathogenesis. Fourth, although our multi-omics framework identified key psoriasis-associated genes and putative cell-cell communication interactions, these findings remain computationally derived and lack direct mechanistic validation. Specifically, the functional relevance of the CD80-CTLA4 interaction and the basal keratinocyte-cDC2A-T cell trio in psoriasis pathogenesis requires further experimental confirmation. Despite our use of spatial transcriptomics to demonstrate the proximity of these cell types, this does not establish causality. As our findings were derived mainly from computational analyses, they should be interpreted with caution and not as definitive evidence of underlying mechanisms. For instance, future studies utilizing skin organoids or CRISPR-based functional screens are necessary to determine whether disrupting these specific interactions alters the inflammatory phenotype.

Taken together, our multi-omics framework provides a systematic approach to identifying cell type-specific molecular features and prioritizing biologically relevant pathways in psoriasis. This framework may help guide future experimental studies, facilitate the identification of potential therapeutic targets, and be extended to other complex diseases.

Methods

Ethics statement

The UK Biobank (UKB) received ethical approval from the National Research Ethics Committee (REC reference 11/NW/0382), and all research procedures were performed in accordance with the principles of the Declaration of Helsinki. For 23andMe data, GWAS data reported in a previous study⁹ was obtained. Participants provided informed consent under the 23andMe human subject protocol, which was reviewed and approved by Ethical & Independent Review Services.

Genome-wide association in UKB

We utilized genetic data from the UKB (version 3), which includes 487,409 individuals released in March 2018. Details of genotyping and imputation have been described in previous studies⁶². Genotyping was performed using Affymetrix UK BiLEVE Axiom or Affymetrix UK Biobank Axiom arrays (Santa Clara, CA, USA), covering more than 800,000 variants. Imputation was centrally conducted by the UKB using reference data from the 1,000 Genomes Project and the UK10K panel, with SHAPEIT3 employed for phasing and IMPUTE2 for imputation. Variant-level quality control (QC) filters were applied to the imputed data as follows: call rate <95%, Hardy–Weinberg equilibrium (HWE) p < 1 × 10⁻⁶, minor allele frequency (MAF) < 0.5%, or imputation quality score (INFO) < 0.4. In the present study, the analysis was restricted to individuals of European ancestry. Patients with psoriasis were defined as those with a self-reported diagnosis of L40 (Data field 131743) or an ICD-9/10 diagnosis code of 696.1/L40. The final dataset comprised 13,079 cases and 395,540 controls. GWAS was performed using a logistic mixed-model approach implemented in the SAIGE software version 1.1.3 (https://saigegit.github.io/SAIGE-doc)⁶³, with age, sex, genotyping array, and the first ten genetic principal components (PC1–PC10) included as covariates.

Meta-analysis of psoriasis GWAS

In addition to the UKB GWAS performed in this study, three additional GWAS datasets were collected: GWAS from Stuart et al.¹⁰, GWAS from FinnGen Release 10¹⁵, and GWAS from 23andMe⁹. The GWAS datasets from Stuart et al. and FinnGen were directly accessible. Access to the full, de-identified GWAS summary statistics from 23andMe was obtained through a data transfer agreement. Following the approach previously used in large-scale psoriasis GWAS⁹, we applied Duffy’s correction²⁶ to account for variations in case definition due to self-reports within the UKB and 23andMe datasets. This method accounts for discrepancies in risk allele frequencies (RAFs) between cases and controls by using well-defined psoriasis cohorts (Tsoi et al., 2012) as reference data⁶⁴. It was confirmed that RAFs in the controls were largely consistent across the UKB and 23andMe datasets. However, RAFs in these cases showed noticeable deviations, similar to those previously observed in the reference data⁶⁴ (Supplementary Fig. 1). Among the 36 genome-wide significant loci previously reported in the study by Tsoi et al., 17 loci in UKB and seven loci in 23andMe replicated at genome-wide significance. The proportion of estimated true-positive cases (q) was estimated in the UKB to be 0.7385, resulting in adjusted sample sizes of 9,659 cases and 398,960 controls. Similarly, for 23andMe, q was estimated to be 0.375, which resulted in an adjusted dataset of 6,045 cases and 264,984 controls. Estimated parameters for UKB and 23andMe are summarized in Supplementary Data 2.

A meta-analysis was performed using METAL released on 2011-03-25 (https://csg.sph.umich.edu/abecasis/Metal)⁶⁵ with an inverse-variance weighted approach. Prior to meta-analysis, an additional filter was applied to retain only common variants (MAF ≥ 1%) across the independent GWAS datasets. GWAS summary statistics that reported variants in hg38 genome coordinates were lifted over to hg19. Independent COJO SNPs were identified using genome-wide joint conditional analysis implemented in GCTA version 1.94.1 (https://github.com/jianyangqt/gcta)⁶⁶ with -cojo-slct option, a genome-wide significance threshold of p < 5 × 10⁻⁸, and a 10-Mb window, based on LD estimates from randomly selected 10,000 UKB individuals of European ancestry. Independent susceptibility loci were defined based on the ±1-Mb window size around each independent COJO SNP. Then, independent susceptibility loci were categorized based on their presence in prior GWAS studies or independent GWAS datasets. Loci that had already reached genome-wide significance within a 1-Mb region in previous studies (GWAS Catalog on October 2, 2025) or any of the four independent GWAS included in this study were classified as previously known loci. Conversely, loci that reached genome-wide significance only in the meta-analysis were designated as unreported loci. MAGMA⁶⁷ was used to conduct a gene-set enrichment analysis implemented in the FUMA web tool version 1.6.0 (https://fuma.ctglab.nl/)⁶⁸ with the default parameters.

Polygenic risk score analysis

To evaluate the impact of adjusting for psoriasis status and effect sizes in GWAS that included self-reported cases, we performed two complementary PRS analyses. First, we constructed GWAS summary statistics by performing a meta-analysis of the FinnGen r10 and Stuart et al. psoriasis GWAS using METAL, both of which did not include self-reported cases. The meta-GWAS summary statistics were applied to calculate PRSs in the UKB. Second, we assessed the predictive performance of the PRSs derived from the 23andMe psoriasis GWAS before and after applying Duffy’s adjustment method to the UKB samples. PRSs were calculated using the PRS-CS-auto algorithm with default parameters to infer the posterior SNP effect sizes. European ancestry samples from the 1000 Genomes Project were used as the LD reference panel.

Evaluation of potential for identifying additional loci in local LD blocks

We examined the association between local heritability significance and the presence of genome-wide significant SNPs across LD blocks using a contingency table (Supplementary Fig. 7A). Local heritability was calculated using the LAVA version 0.1.0 (https://github.com/josefin-werme/LAVA)⁶⁹ tool with default parameters across 2495 genomic loci. The statistical significance of each local LD block was defined as a Bonferroni-corrected p-value below 0.05 (0.05/2495). Each LD block was classified based on LAVA significance and the presence of genome-wide significant SNPs. The 60 LD blocks that were LAVA-significant but lacked genome-wide significant SNPs were further categorized according to their potential for additional discovery, considering the inflated heritability due to residual LD. First, LD blocks located within ±10 Mb of the MHC region were classified as the Near MHC region. Next, we classified each LD block based on its local heritability and the presence of nearby significant SNPs within a 5-Mb region: blocks for which another block within the 5-Mb region showed higher heritability were classified as Likely residual LD; blocks with previously identified significant SNPs in other blocks within the 5-Mb region were classified as Nearby GWS; otherwise, blocks that exhibited the highest heritability in the region and no nearby GWS had been reported were classified as having high potential for additional discovery.

LD score regression

A stratified LDSC analysis was performed to evaluate cell- and tissue-specific enrichment⁷⁰. This approach estimates the enrichment of psoriasis heritability in specific tissues and cell types using expression-derived function annotations. Precomputed annotation files from Finucane et al.⁷⁰ were used as reference datasets. Enrichment was tested using a one-sided test, and p-values were corrected using the Benjamini-Hochberg false discovery rate (FDR). Adjusted p-values < 0.05 were considered significant. To investigate potential genetic sharing between psoriasis and other diseases or lifestyle factors, we curated 113 GWAS summary statistics from diverse disease and trait categories and conducted an LDSC-based genetic correlation analysis. Each phenotype was classified into one of the following 10 categories: autoimmune, immune, cardiometabolic, neurological diseases, psychiatric disorders, mental health, cognitive function, laboratory and physical findings, leisure/social activity and lifestyles, and other physical health. Genetic correlation (r_g) was reported with 95% confidence intervals (CIs), and statistical significance was determined based on Bonferroni-adjusted p-value < 0.05/113. The list of GWAS used in this analysis is summarized in Supplementary Data 10.

Single-cell RNA-seq data analysis

To integrate psoriasis-associated gene expression across various cell types with the GWAS results, Single-cell RNA sequencing (scRNA-seq) data generated by Ma et al.¹⁶ was utilized. This dataset included normal skin samples (denoted as NS) from healthy individuals and non-lesional (denoted as PN) and lesional skin samples (denoted as PP) from patients with psoriasis.

We performed preprocessing following a standard single-cell pipeline using raw count matrices. Low-quality cells were filtered out based on the following criteria: (1) cells with <500 unique molecular identifiers (UMIs). (2) cells with <100 genes detected. (3) cells with >10% mitochondrial gene expression. Additionally, cells displaying outlier characteristics (i.e., UMI counts exceeding three median absolute deviations (MADs) from the global median) were excluded. To further remove doublet cells, DoubletFinder (version 2.0.3)⁷¹ was applied to each sample, and the identified doublets were removed. All the preprocessing and cell type annotation steps were conducted using the Seurat package version 5.1.0 (https://satijalab.org/seurat/articles/install_v5)⁷² in R. After merging the samples, gene expression values were normalized using the SCTransform (SCT) algorithm⁷³, which reduces technical noise from sequencing artifacts. Mitochondrial percentage was included as a covariate using the vars.to.regress parameter. Principal component analysis (PCA) was performed on the SCT-normalized expression values using the RunPCA function, and batch effects were corrected using RunHarmony for the top 30 principal components (PCs). To identify major cell types, we performed Uniform Manifold Approximation and Projection (UMAP) and graph-based Louvain clustering using the top 20 Harmony components. The cell types were annotated based on the specific expression of representative marker genes (Supplementary Data 11). For DEG analysis, the Wilcoxon rank-sum test implemented in the FindAllMarkers function with min.pct=0.1 was used. To further characterize the subpopulations within each major cell type, we re-applied SCTransform normalization, Harmony batch correction, Louvain clustering, and UMAP. Each cell subtype was labeled as Celltype_ClusterID_CellSubtype (if expressing canonical cell subtype markers)_MarkerGene. The annotation of the cell subtypes expressing canonical cell subtype markers was based on the results of a previous study that generated and published a dataset¹⁶. To calculate the gene-set-based score for each cell, we used the AddModuleScore function with parameter nbins=12 and other default parameters.

For each cell subtype, we estimated the expected cell counts by constructing a contingency table between the tissue group (lesion vs. non-lesion) and cell membership in the subtype of interest. The expected cells were calculated using the following formula, employed in the chi-square test:

$$Expectedcells=\frac{\left(Totalcellsofthesubtype\right)\times \left(Totalcellsinthetissuegroup\right)}{Totalnumbercells}$$ 1

Finally, the ratio of observed to expected (Ro/e) values was calculated for lesion and non-lesion tissues as follows:

$$Ro/e=\frac{Observedcells}{Expectedcells}$$ 2

If the Ro/e value for the non-lesion was greater than 1, the cell subtype was defined as normal-enriched. Otherwise, the subtype was defined as lesion-enriched if the Ro/e value for lesional tissue was greater than 1. Statistical significance of Ro/e values was assessed by 10,000 random permutations of tissue labels with two-sided empirical p-values. After multiple testing correction, FDR < 0.05 was considered statistically significant.

To quantify the relative disease enrichment of psoriasis risk genes at the single-cell level, scDRS analysis version 1.0.2 (https://github.com/martinjzhang/scDRS)⁷⁴ was performed. The scDRS algorithm translates the GWAS-based genetic susceptibility to psoriasis into cell-level disease relevance scores by projecting GWAS signals onto single-cell expression profiles. This allowed us to quantify the cells and cell types that showed stronger genetic associations with psoriasis. This method uses MAGMA z-scores from psoriasis GWAS with preprocessed scRNA-seq data to calculate the aggregated disease scores. The scDRS score was weighted using the MAGMA z-scores and inversely weighted using technical noise in single-cell gene expression. The mitochondrial UMI percentages and tissue types (NS, PN, and PP) were used as covariates. To assess statistical significance, 10,000 matched control sets of cell-specific control genes were generated using one-sided Monte Carlo permutation tests. The observed scores of the GWAS-identified genes were then compared with control distributions. Cell types were considered disease-relevant if the Benjamini-Hochberg FDR-adjusted p-value < 0.1 across all cell types. Differences in scDRS scores between cells from normal and lesional tissues were assessed for each cell type using a permutation test with 10,000 permutations, and p-values were adjusted using the Benjamini-Hochberg FDR.

Within the keratinocyte population, the cells exhibited a continuous spatial organization along the epidermal layer in the UMAP space. Keratinocyte differentiation from the basal layers toward the outer epidermis is well established⁷⁵. This pattern suggests a gradual differentiation process, and we applied trajectory inference using a Slingshot library⁷⁶ with parameter start.clus = KER_c1_KRT15. This analysis allowed us to model pseudotime trajectories, with basal layer keratinocytes serving as the starting point, progressing toward the spinous and granular layers. To further validate keratinocyte differentiation, RNA velocity analysis was performed using scVelo⁷⁷ and veloVI⁷⁸ for functional validation. RNA velocity estimates future transcriptional states of cells by modeling the relative abundance of unspliced and spliced transcripts in the single-cell expression profiles. Higher proportions of unspliced reads indicated active transcription and were associated with less differentiated cellular states. For veloVI analysis, the model was implemented using default hyperparameters as specified in the documentation, with the training process conducted for a maximum of 500 epochs.

Prioritizing therapeutic target genes

Genes with potential causal effects on psoriasis were prioritized by integrating GWAS and single-cell data. For the initial prioritization, five independent criteria were applied. Genes that met at least four of these criteria were considered putative susceptibility genes for psoriasis. First, we identified genes located within 10 kb of the COJO-identified SNPs using the FUMA web tool⁶⁸. Second, we assessed gene-level statistical significance using MAGMA⁶⁷, which calculates significance based on the distribution of p-values from GWAS-significant SNPs within coding regions, as implemented in FUMA. Third, we performed a transcriptome-wide association study (TWAS) using S-Predixcan (https://github.com/hakyimlab/MetaXcan)⁷⁹ to infer transcriptome-wide significant genes from eQTL data in whole blood and two skin tissue types (sun-exposed and not sun-exposed) in GTEx. Pretrained models from Multivariate Adaptive Shrinkage Models (MASHR-M) were downloaded and applied to our GWAS summary statistics. Fourth, we conducted a Bayesian colocalization analysis using coloc version 5.2.3 (https://github.com/chr1swallace/coloc)⁸⁰ to evaluate whether the independent COJO SNPs for psoriasis colocalized with eQTL signals based on marginal effects from GWAS summary statistics. This colocalization framework assumes a single causal variant underlying the tested association signals at each locus. Genes with a posterior probability (PP.H4) > 0.8 were considered colocalized, based on eQTL data from GTEx (two skin tissue types and whole blood) and eQTLGen (whole blood). For gene prioritization, genes in the MHC region were excluded because of the highly complex LD structure. Fifth, we selected genes expressed in at least 10% of cells from any of the four cell types: keratinocytes, myeloid cells, T cells, and ECs, which showed significant scDRS results (FDR < 0.1).

Summary-based Mendelian randomization

To assess the causal effects of the prioritized genes on psoriasis, a summary-based Mendelian randomization (SMR) version 1.3.1 (https://yanglab.westlake.edu.cn/software/smr) analysis⁸¹ was conducted for 110 putative susceptibility genes. We utilized expression quantitative trait loci (eQTL) data from two skin tissues and whole blood from the Genotype-Tissue Expression (GTEx) project⁸², as well as whole blood eQTL data from eQTLGen. Genes were considered to have a causal relationship with psoriasis if they met the Bonferroni-adjusted p-value threshold (0.05/110) and passed the HEIDI test (p-value < 0.01), which evaluates the likelihood of pleiotropic effects due to linkage disequilibrium. Results with HEIDI p-values between 0.01 and 0.05 require careful interpretation due to potential heterogeneity, with further details provided in Supplementary Data 17.

Druggable genes

For the druggability assessment, we leveraged the most recent tractability information from Open Targets (released in June 2024). This resource evaluates the potential suitability of targets for drug discovery based on the characteristics of known compounds. Following the classification proposed by Rasooly et al.⁸³, drug targets were categorized into five groups: approved drugs (bucket 1 for all modalities); drugs in clinical development (buckets 2 and 3 for all modalities); drugs in the preclinical phase (buckets 4 and 5 for small molecules); predicted druggable targets (buckets 6 to 8 for small molecules and buckets 4 and 5 for antibodies); and otherwise least druggable targets.

Co-expression gene network analysis

We performed a weighted gene co-expression network analysis (WGCNA) on each cell type using the hdWGCNA package version 0.4.1 (https://github.com/smorabit/hdWGCNA)⁸⁴ with Ma et al.’s single-cell data¹⁶. Prior to constructing the co-expression networks, similar cells were aggregated into meta-cells using a k-nearest neighbors algorithm (k = 25) applied to harmony embeddings. To define connections between genes, the optimal soft-thresholding power was determined by fitting a scale-free topology model. The TestSoftPowers function in the hdWGCNA automatically estimated this value by identifying the lowest soft power threshold that achieved a scale-free topology model fit exceeding 0.8. Subsequent steps, including network construction, computation of module eigengenes, and assessment of eigengene-based connectivity (kME), were conducted using a standard hdWGCNA pipeline. This analysis was independently applied to keratinocyte, myeloid, and T-cell populations, which were previously identified as significant in the scDRS analysis. All other parameters were set to defaults following the recommendations of the tool developers, including the calculation of kME using the Pearson correlation.

We further quantified the enrichment of cell subtypes within modules containing at least two of the 50 prioritized genes. For a given module M the raw average module eigengene (ME) of cells in cell subtype i was computed as:

$$S_i=\frac{{\mathrm{\Sigma }}_{c\in C_i}M{E}_{\mathrm{c}}}{∣C_i∣}$$ 3

where $ME$ denotes the module eigengene, S_i represents the average MEs of cell subtype i, and |C_i | represents the number of the cells of cell subtype i. Because $S_i$ values can be negative, min-max scaling was performed to normalize the module scores $S_i$ within the [0, 1] range across all cell subtypes.

$$S_i^{\prime }=\frac{S_i-\min \left(S\right)}{\max \left(S\right)-\min \left(S\right)}$$ 4

Using these normalized module scores across cell subtypes, the proportion of normalized module scores was defined as the module M enrichment of a given cell subtype i, calculated as:

$$\text{Specificity}{\text{Score}}_i=\frac{S_i^{\prime }}{{\sum }_{j=1}^n{S}_j^{\prime }}$$ 5

where n denotes all cell subtypes belonging to the major cell type analyzed.

IL-23-targeted treatment single-cell data analysis

We analyzed single-cell gene expression changes over time during IL-23-targeted therapy using data from Francis et al.¹⁷, which involved skin tissue sampling from five patients diagnosed with severe psoriasis across three visits. We utilized scRNA-seq data from non-lesional and lesional skin collected at baseline (day 0) and lesional skin at 3 and 14 days after treatment (day 3 and day 14). To integrate this dataset with the already processed scRNA-seq data, we performed the same preprocessing steps. Major cell types were annotated using representative markers, focusing on keratinocyte, myeloid, and T cell subsets. We employed the MapQuery function in Seurat to perform reference mapping using SCT-normalized expression and Harmony embeddings. Module mapping for independent cell types was conducted using the ProjectModules function in the hdWGCNA library.

Additionally, an independent scRNA-seq dataset from Kim et al.³⁴ was included as the replication dataset. This dataset included skin scRNA-seq data from healthy controls and patients with psoriasis, as well as samples collected at a long-term post-treatment time point (week 28 after risankizumab treatment). All the preprocessing steps were performed using the same pipeline as described above. Cell type annotation for this dataset was conducted independently using canonical marker genes reported in the original study rather than reference mapping because substantial heterogeneity in data characteristics driven by distinct expression profiles resulted in reference-based mapping being disproportionately assigned to a single cluster. We performed scDRS analysis on the Kim et al. dataset for external validation. To ensure consistency with our primary analysis of the Ma et al. data, we included pre-treatment (baseline) psoriasis samples and healthy control samples. The analysis utilized the same MAGMA-derived gene weights and parameters that we applied to the Ma et al. data. Mitochondrial percentages and sample types (lesional or healthy) were used as covariates.

To assess the temporal enrichment changes in module eigengenes (MEs), we offset the average MEs of non-lesional cells to zero for each module. Subsequently, the values were scaled based on the absolute maximum values of the average MEs from the lesional samples across the three visits. To visualize modules with similar patterns across samples, we computed the pairwise Pearson correlation of average MEs of 20 samples across all modules and adjusted the computed p-values using Benjamini-Hochberg FDR. A graph was constructed using the igraph R package with the Fruchterman-Reingold layout algorithm. Gene-set pathway enrichment analysis was conducted using the enrichr R package⁸⁵ with default parameters, including pathways such as GO:BP, GO:CC, GO:MF, KEGG, and REACTOME.

Cell–cell interaction analysis

Cross-cell type communication analysis was performed using MultiNicheNet R package version 2.1.0 (https://github.com/saeyslab/multinichenetr)⁸⁶, a framework designed to identify ligand-receptor pairs that are differentially active between conditions of interest. In this study, we estimated condition-specific cell-cell communications by performing pairwise contrasts using IL-23 inhibition treatment datasets, focusing on keratinocytes, myeloid cells, and T cells. Specifically, two contrasts were tested: lesion versus non-lesion to identify disease-associated interactions, and treatment versus lesion to identify treatment-induced changes. Ligand-receptor analyses were restricted to interactions between the predefined module pairs highlighted in Fig. 6H, including KER-M3 and MYE-M2, as well as KER-M1, MYE-M5, and T-M1. Data preprocessing and differential tests were conducted using the default parameters, following the recommendations of the tool developers. For each module pair, ligand-receptor interactions were evaluated by matching the ligand genes expressed in the sender cell modules with the receptor genes expressed in the receiver cell modules. Differential tests were conducted using the default MultiNicheNet workflow, which employs a generalized linear model based on sample-level pseudobulk analysis to determine log2 fold changes and p-values. Owing to the sparsity of single-cell data, results from differential tests were filtered using a nominal P  <  0.05 threshold, while adjusted P values were also provided in the results.

Spatial transcriptome data

Spatial transcriptomic data obtained from the study by Castillo et al.⁸⁷ was used to validate the results of cell-cell communication analysis. We used preprocessed data provided in the original publication. Nine non-lesional samples from psoriasis patients were excluded from this analysis, resulting in seven normal skin samples and 14 lesional skin samples included in the analysis. Spatial spots corresponding to basal keratinocytes and cDC2A were defined based on the co-expression of two marker genes for each cell type (Basal keratinocyte: DSG3 and KRT5; cDC2A: DSC2 and LAMP3). For each spot, a composite expression score was calculated using the product of the log-normalized expression values of the corresponding marker genes as a measure of cell-type-specific signal enrichment. To assess the spatial proximity between the two cell types, we calculated the bivariate spatial autocorrelation using Lee’s L statistic⁸⁸. Lee’s L quantifies the degree of spatial co-localization between two spatially distributed variables across neighboring spatial units. The statistic ranges from −1 to 1, where values close to 0 indicate spatial randomness, positive values indicate spatial co-localization, and negative values indicate spatial segregation. Lee’s L statistic was calculated for each sample to evaluate spatial associations between defined cell-type-specific spots.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

H.J. and H.-H.W. designed the study, coordinated the GWAS, processed the single-cell data, and conducted the analyses. W.-Y.P. and H.-H.W. provided the UKB dataset. H.J., B.K., and I.S. processed the UKB genotype data. H.J., Y.A., and S.K. curated the GWAS summary statistics for genetic correlation analysis. H.J., Y.A., and H.-H.W. coordinated figures and tables. M.S., H.K., and H.-H.W. provided advice on the study design and statistical analysis. S.H. managed and resolved issues with the server and software used in this study. All authors contributed to the interpretation of the results and agreed to the final version of the manuscript.

Peer review

Peer review information

Nature Communications thanks Fan Zhang, who co-reviewed with Lauren Vanderlinden, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding statement

This study was supported by grants from the National Research Foundation of Korea (NRF), funded by the Korea Government (Ministry of Science and ICT; RS-2022-NR070328 and RS-2023-00223277).

Data availability

The UKB GWAS summary statistics and meta-analysis summary statistics generated in this study (excluding the 23andMe dataset) have been deposited in the GWAS Catalog (https://www.ebi.ac.uk/gwas/downloads) under accession numbers (GCST90984667, GCST90984668, GCST90984669, GCST90984670). The 23andMe GWAS summary statistics are available under restricted access to protect participant privacy. Access can be obtained by qualified researchers by applying directly to 23andMe (https://research.23andme.com/collaborate/), in accordance with 23andMe’s data sharing policy.

The publication sources and download URLs for the GWAS summary statistics used in the genetic correlation analyses are listed in Supplementary Data 10.

The following GWAS summary statistics are available at the corresponding URLs: FinnGen r10¹⁵ (https://www.finngen.fi/en/access_results), Stuart et al.¹⁰ (https://www.ebi.ac.uk/gwas/studies/GCST90019016).

Single-cell transcriptomics data are available at the Gene Expression Omnibus: Ma et al.¹⁶ (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE173706), Francis et al.¹⁷ (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE228421), Kim et al.³⁴ (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE278330).

Spatial transcriptome data from Castillo et al.⁸⁷ are available at the Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE202011). All other data supporting the findings of this study are available within the article and Supplementary Information files.

Code availability

Free software and frameworks are used for all analyses.

Custom scripts for this study are available on GitHub (https://github.com/HyeonbinJoHCho/Psoriasis_Jo2026) at Zenodo (https://zenodo.org/records/21256344)⁸⁹.

Competing interests

W.-Y.P. is an employee of GENINUS, which is unrelated to this work. The remaining authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77309-2.