Transcriptomic Subtyping of Atopic Dermatitis Reveals Distinct Drug Response Signatures

Transcriptomic Subtyping of Atopic Dermatitis Reveals Distinct Drug Response Signatures | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Transcriptomic Subtyping of Atopic Dermatitis Reveals Distinct Drug Response Signatures Sang Hyun Moh, Jiyeon Kim, Jisoo Han, Sun Hee Hwang, Bumho Yoo, and 6 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8477069/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Atopic dermatitis (AD) is a chronic inflammatory skin disease characterized by marked clinical heterogeneity and variable responses to systemic therapies. Although transcriptomic studies have revealed substantial molecular diversity within AD lesions, how this heterogeneity relates to therapeutic responsiveness remains incompletely understood. Here, we performed integrative transcriptomic analyses of lesional and non-lesional skin from patients with AD to define molecular subtypes and examine their biological and therapeutic relevance. Unsupervised clustering of lesional skin transcriptomes identified two distinct lesional subtypes with divergent molecular features. These subtypes differed markedly in immune activation, cell-cycle–associated programs, and immune cell composition. We next assessed subtype-specific drug responsiveness by comparing lesional gene expression signatures with transcriptomic changes induced by dupilumab and cyclosporine treatment. The major lesional subtype exhibited strong inverse transcriptional concordance with drug-induced expression changes, whereas the minor lesional subtype showed attenuated responses, particularly to cyclosporine. Analysis of an independent cyclosporine-treated cohort further demonstrated that clinical non-responders displayed transcriptomic features resembling the minor lesional subtype, including enrichment of cell-cycle–associated programs. Together, these findings demonstrate that AD lesions comprise biologically distinct transcriptomic subtypes with differential immune composition and systemic drug responsiveness, providing a molecular framework for understanding heterogeneity in AD and supporting transcriptome-based stratification for precision treatment strategies. Health sciences/Biomarkers Biological sciences/Computational biology and bioinformatics Health sciences/Diseases Biological sciences/Immunology Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Atopic dermatitis (AD) is a common chronic inflammatory skin disease affecting up to 10–20% of children and 2–10% of adults worldwide [1]. The disease is characterized by recurrent eczematous lesions, intense pruritus, and a relapsing clinical course, leading to substantial impairment in quality of life and increased psychosocial burden [2]. In moderate-to-severe cases, AD often requires long-term systemic therapy, underscoring the clinical need for effective and durable treatment strategies. From a pathogenic perspective, AD has long been regarded as an immune-mediated disorder, with type 2–driven inflammation playing a central role in disease development and progression [3–5]. However, recent transcriptomic studies have revealed considerable inter-patient variability in the degree and composition of immune involvement [6,7]. Single-cell transcriptomic analyses of human AD skin have demonstrated heterogeneous enrichment of immune cell subsets together with distinct stromal and fibroblast populations, highlighting marked cellular diversity within lesional tissue [6]. Consistent with these observations, bulk transcriptomic profiling across patient cohorts has shown that AD lesions can exhibit variable combinations of Th2- and Th17-associated immune programs rather than a uniform type 2–dominant signature [7]. Moreover, longitudinal transcriptomic analyses suggest that AD lesions comprise both stable, patient-specific molecular signatures and dynamic inflammatory programs that fluctuate with disease activity, underscoring intrinsic molecular heterogeneity beyond transient inflammation [8]. This intrinsic molecular heterogeneity has important clinical implications, particularly with respect to therapeutic response. Despite the availability of targeted and systemic treatments, clinical outcomes in AD remain highly variable, with substantial differences in treatment efficacy observed across patients receiving the same therapy [9,10]. This variability in treatment response has been well documented in both controlled clinical trials and real-world settings. In the pivotal phase 3 LIBERTY AD CHRONOS trial, dupilumab treatment resulted in marked overall clinical improvement at the population level; however, only a subset of patients achieved stringent response criteria. At week 16, Investigator’s Global Assessment (IGA) 0/1 with a ≥2-point improvement was observed in approximately 39% of dupilumab-treated patients, while EASI-75 responses were achieved in 64–69%, indicating that a considerable proportion of patients exhibited partial or suboptimal responses despite effective type 2 pathway blockade. [9]. Similar patterns persisted at week 52, underscoring sustained inter-individual variability in treatment outcomes. A large systematic review and meta-analysis encompassing over 3,300 AD patients across 22 observational studies reported pooled response rates of 85.1% for EASI-50, 59.8% for EASI-75, and only 26.8% for EASI-90 after 16 weeks of dupilumab therapy, indicating that a substantial fraction of patients experience partial or incomplete clinical improvement despite treatment [10]. These findings suggest that, beyond overall efficacy, inter-individual differences in underlying molecular and immunological programs likely contribute to differential therapeutic responsiveness in AD. In addition to biologic therapies, similar heterogeneity in treatment response has been reported for conventional systemic immunosuppressants such as cyclosporine. Although cyclosporine is widely used for moderate-to-severe AD and can induce rapid clinical improvement, both clinical trials and real-world studies have demonstrated substantial inter-individual variability in treatment efficacy and durability. Substantial inter-patient heterogeneity in real-world effectiveness can be quantified by discontinuation patterns (“drug survival”) and the proportion of patients stopping therapy due to lack of benefit. In a large daily-practice cohort of adults with severe AD (n=356), the overall cyclosporine drug survival declined steeply over time (34%, 18%, 12%, and 4% remaining on cyclosporine at 1, 2, 3, and 6 years, respectively; median overall drug survival 256 days), and discontinuations reflected multiple clinical trajectories—including disease control (26.4%), treatment-limiting side effects (22.2%), and ineffectiveness (16.3%), with an additional 6.2% stopping due to combined side effects plus ineffectiveness—indicating that a meaningful subset of patients either cannot tolerate cyclosporine or do not achieve adequate clinical response despite treatment [11]. Consistent estimates were reported in a 10-year experience from two academic centers: among cyclosporine courses that were discontinued (234/267), approximately 15% were stopped as “not effective,” while 24% were discontinued due to adverse events and 29% due to controlled disease (median treatment duration 258 days), reinforcing that cyclosporine response is heterogeneous—ranging from clear clinical control to early termination for either insufficient efficacy or toxicity [12]. The pronounced inter-individual variability in clinical responses to both dupilumab and cyclosporine observed across previous studies suggests that patients diagnosed with AD may harbor distinct underlying molecular states despite shared clinical features. Such heterogeneity in treatment outcomes is unlikely to be fully explained by differences in drug exposure or adherence alone, and instead points to intrinsic molecular differences that may shape therapeutic responsiveness. These observations highlight the need to move beyond a uniform view of AD as a single inflammatory entity and to systematically characterize patient-level molecular heterogeneity. In this context, defining transcriptional subtypes of lesional skin and linking them to differential drug responses represents a critical step toward improving treatment stratification and advancing precision medicine approaches in AD. In several complex and heterogeneous diseases, systematic molecular stratification has provided a powerful framework for capturing biological diversity and informing therapeutic decision-making beyond conventional clinical classification. In breast cancer, for example, transcriptome-based subtyping schemes such as the PAM50 classifier have established biologically distinct molecular subtypes with reproducible differences in prognosis and treatment response, and are now routinely integrated into clinical practice [13,14]. By contrast, although atopic dermatitis exhibits marked molecular heterogeneity and variable responses to systemic therapies, a comparably standardized and clinically actionable transcriptomic subtyping framework has not yet been established. This gap highlights the need for molecularly informed approaches to stratify AD patients and to better understand heterogeneity in therapeutic response. On the basis of these considerations, we sought to systematically delineate molecular subtypes of AD using lesional skin transcriptomic profiles and to investigate whether such subtypes are associated with distinct biological characteristics and therapeutic responses. Using an integrative transcriptomic approach, we identified two lesional subgroups with divergent molecular features. These lesional subtypes differed in the extent of immune activation, metabolic and cell-cycle–associated programs, immune cell composition, and responsiveness to systemic therapies. Together, these findings support the existence of biologically distinct AD subtypes at the transcriptomic level and provide a molecular framework for understanding heterogeneity in disease mechanisms and treatment response. Methods Data Sources and Preprocessing Transcriptomic profiling was performed using the RNA-seq dataset GSE157194, which contains paired lesional (AL) and non-lesional (AN) skin biopsies collected prior to any systemic treatment. After quality filtering and metadata verification, 57 AL samples and 54 AN samples at baseline were included for subtype discovery and differential expression analysis. The dataset also includes 55 follow-up lesional samples collected three months after treatment with either dupilumab or cyclosporine, which were used to assess drug-induced transcriptional changes. Raw count matrices were processed using DESeq2 (v1.46.0), and both variance-stabilizing transformation (VST) and counts per million (CPM) values were generated to obtain stabilized and comparable expression measures for downstream analyses. The microarray cohort GSE58558, which profiles skin biopsies from atopic dermatitis patients treated with cyclosporine, was downloaded using the GEOquery package. The dataset provides log₂-normalized probe-level intensities together with metadata specifying lesional versus non-lesional sampling and treatment response categories. In total, the cohort includes 50 responder and 6 non-responder lesional samples, as well as 48 responder and 5 non-responder non-lesional samples, spanning multiple treatment time points (day 1, week 2, week 12). Probe identifiers were mapped to HGNC gene symbols using the hgu133plus2.db annotation database. When multiple probes mapped to the same gene symbol, the first matching probe listed in the annotation database was retained. Differential Expression Analysis and Clustering Differentially expressed genes (DEGs) were identified using DESeq2 applied to raw count matrices for GSE157194. Differential expression was assessed by contrasting non-lesional (AN) and lesional (AL) samples, and genes with |log₂FC| ≥ 1 and an adjusted p-value < 0.05 were considered significant. The same criteria were applied to define subtype-specific DEGs for the major and minor AL subtypes by performing non-lesional versus lesional comparisons within each subtype. Hierarchical clustering was performed using seaborn.clustermap applied to CPM-normalized DEG expression matrices. Before clustering, gene-wise values were standardized (standard_scale=0), and robust scaling was applied (robust=True) to reduce the influence of outliers. Clustering was computed using the default Euclidean distance metric and average linkage method. Principal component analysis (PCA) was applied to the CPM-normalized DEG expression matrix to assess global transcriptional structure. Dimensionality reduction was performed using the standard PCA implementation in sklearn (n_components=2), and the resulting PC1–PC2 coordinates were used to visualize separation between lesional and non-lesional samples. PCA results were used to visualize separation between lesional and non-lesional groups and to evaluate subtype-level structure. Microarray-based differential expression analysis was performed separately for the GSE58558 cohort using the limma framework. Raw probe-level intensities were background-corrected and quantile-normalized, after which probes were mapped to gene symbols using platform-specific annotation. A linear modeling approach was applied using limma’s empirical Bayes shrinkage, contrasting lesional versus non-lesional samples within the dataset. The criteria used to define DEGs were identical to those applied in the RNA-seq analysis. Functional Enrichment Analysis Functional enrichment analysis was performed on DEGs identified from non-lesional versus lesional comparisons. Upregulated and downregulated DEGs were analyzed separately. Enrichment was conducted using gseapy.enrichr, which implements the Enrichr statistical framework. To capture both pathway-level and transcriptional program–level signals, KEGG 2021 Human and MSigDB Hallmark 2020 gene set libraries were queried. For each DEG subset (UP or DOWN), the gene symbols were submitted as input, and enrichment scores, adjusted P-values, and leading-edge genes were retrieved following Enrichr’s default statistical procedures (Fisher’s exact test–based overrepresentation analysis with Benjamini–Hochberg correction). Only terms passing an adjusted P-value threshold (e.g., FDR < 0.05) were retained for interpretation. The same enrichment workflow was applied to DEGs defining the major and minor AL subtypes, which were obtained from subtype-specific lesional versus non-lesional contrasts. For each subtype, upregulated and downregulated DEGs were submitted separately to Enrichr , and enriched pathways were summarized to characterize biological programs distinguishing the two transcriptomic subtypes Differential expression analysis of the GSE58558 cyclosporine cohort produced an insufficient number of DEGs for conventional over-representation analysis. To enable pathway-level interpretation despite the limited DEG count, we applied preranked gene set enrichment analysis (GSEA) using the full transcriptome ranked by log2 fold change. Enrichment was performed with gseapy.prerank against the MSigDB Hallmark 2020 and KEGG 2021 Human gene sets, and significance was defined using FDR-adjusted q-values (FDR < 0.05). Molecular Subtyping of Atopic Dermatitis To identify molecular subgroups within atopic lesional skin, we applied the cola framework for consensus partitioning using the VST-normalized expression matrix derived from DEGs of non-lesion vs. lesion of patients. Cola was run using its standard workflow, which combines feature selection, iterative subsampling, and stability assessment across multiple partitioning runs. We selected the skmeans method as the primary clustering algorithm, as recommended for high-dimensional transcriptome data, and evaluated cluster numbers from k = 2 to 6. At each k, cola repeatedly partitions the data (50 iterations in our analysis) and computes consensus matrices that summarize how consistently samples co-cluster across runs. Evaluation metrics provided by cola—including the mean silhouette width, proportion of ambiguous clustering (PAC) score, and consensus heatmap structure. Samples were assigned to subtypes based on their consensus membership, and to ensure high subtype fidelity, only those with silhouette ≥ 0.5 were retained for downstream signature analysis. This filtering step ensured that ambiguous or borderline samples did not confound subtype-specific interpretations. To assess whether each molecular subtype was preferentially associated with lesional versus non-lesional skin, we performed a fisher’s exact test for enrichment of pathology status within each cluster. For a given subtype, we constructed a 2×2 contingency table with rows corresponding to tissue type (AL vs. AN) and columns indicating whether a sample belonged to the subtype of interest (“in”) or to any of the other subtypes (“out”). Fisher’s exact test was applied to each table to obtain an odds ratio and p-value. An odds ratio greater than 1 was interpreted as AL enrichment for that subtype, whereas an odds ratio less than 1 indicated AN enrichment; odds ratios close to 1 were considered to reflect no clear enrichment. Immune Cell Deconvolution To estimate immune-cell composition from bulk transcriptomic data, we applied CIBERSORT (v.0.1.0) using the LM22 leukocyte signature matrix, which infers the relative abundance of 22 immune cell subsets. Normalized expression matrices from RNA-Seq (VST) was formatted according to CIBERSORT requirements and analyzed in relative fractions mode. All samples were processed using 1,000 permutations, and only deconvolution results with p < 0.05 were retained for downstream analysis. Differences in immune-cell composition among non-lesional skin, major AL, and minor AL subtypes were assessed using the CIBERSORT-estimated cell fractions. For each cell type, we performed pairwise two-sided Welch’s t-tests to compare (i) AN vs. major AL, (ii) AN vs. minor AL, and (iii) major vs. minor AL. Cell types with nominal p-values < 0.05 were considered to show statistically significant subtype-associated differences and were subsequently highlighted in box-plot visualizations. Drug Responsiveness Analysis To characterize transcriptional responses to dupilumab and cyclosporine, lesional skin samples from GSE157194 collected before treatment and three months after treatment were used to compute drug-induced expression changes. For each drug, log₂ fold-changes were obtained by contrasting post-treatment versus pre-treatment lesional samples. These drug-response vectors were then compared with the subtype-specific lesional signatures (major and minor AL) by computing cosine similarity. Batch Effect Correction Between Cohorts To harmonize gene expression profiles across the two independent cohorts (GSE157194 and GSE58558), batch effects were mitigated using a rank-based inverse normal transformation (INT). For each gene, expression values were independently replaced by their sample-wise ranks, which were then mapped onto a standard normal distribution using the transformation: Where γ i is the rank of sample i for the given gene, N is the number of samples, and Φ -1 denotes the inverse cumulative distribution function of the standard normal distribution. This procedure produces a rank-preserving, distribution-normalized expression matrix that minimizes batch-driven shifts in mean or variance while maintaining relative ordering within each gene. Following batch correction, principal component analysis (PCA) was applied to the combined expression matrix to assess whether samples from the two cohorts remained separated or were mixed in the reduced-dimensional space. PCA results before and after correction were visually compared to confirm the effective reduction of batch-associated clustering. Cyclosporine Treatment Trajectory Analysis To characterize longitudinal transcriptional dynamics following cyclosporine treatment, we computed trajectory paths for responders and non-responders using the batch-corrected PCA space derived from the integrated GSE157194 and GSE58558 datasets. Each GSE58558 sample was annotated with treatment time point (day 1, week 2, week 12) and response status (responder or non-responder). PCA coordinates (PC1 and PC2) of these samples were then used to estimate temporal movement in transcriptomic space. For each response group, samples were first grouped by treatment time point, and the median PC1 and PC2 values within each time point were calculated to obtain a centroid representing the group’s average transcriptomic state at that stage. Centroids were computed sequentially for day 1, week 2 and week 12. Trajectory paths were visualized by connecting these centroids in temporal order. Results Transcriptomic Differences Between Atopic Lesional and Non-Lesional Skin Atopic dermatitis (AD) presents with diverse clinical manifestations, suggesting that patients may harbor distinct molecular states despite similar phenotypes. We hypothesized that such transcriptomic heterogeneity contributes to the variability in disease severity and treatment response, and that identifying these differences would provide a basis for stratified therapeutic strategies. To evaluate this possibility, we analyzed the GSE157194 dataset, which contains paired lesional and non-lesional skin samples from AD patients, focusing on characterizing the global transcriptional alterations associated with lesional skin. The dataset consisted of 57 atopic lesional (AL) and 54 atopic non-lesional (AN) samples collected prior to treatment. Differential expression analysis was performed using DESeq2, and genes with |log₂fold-change| ≥ 1 and FDR < 0.05 were considered significant. This analysis identified 1,047 upregulated and 355 downregulated genes in AL compared with AN, demonstrating substantial transcriptional remodeling associated with lesional skin (Supplementary Table S1). Clustering analysis based on the identified differentially expressed genes (DEGs) revealed that samples from AL and AN skin did not form completely distinct groups, with several AL and AN samples intermingling within the same clusters (Fig. 1A). This lack of strict separation indicates that the transcriptional profiles of AD skin exist on a continuum rather than as two discrete states. PCA demonstrated incomplete separation between lesional and non-lesional skin samples along the major variance component (PC1, 78.2%) (Fig. 1B). AN samples formed a compact cluster, whereas AL samples displayed a wide dispersion with a broad confidence ellipse, indicating substantial heterogeneity among lesional transcriptomes. These findings suggest that heterogeneous molecular states exist within AD lesions, reflecting differing degrees or types of pathogenic activation across patients. To characterize the biological processes underlying the transcriptional differences between lesional (AL) and non-lesional (AN) skin, we performed pathway enrichment analysis on the differentially expressed genes (Fig. 1C). Upregulated genes in AL were predominantly associated with immune activation pathways, including cytokine–cytokine receptor interaction, IL-17 signaling, inflammatory response, and allograft rejection. These pathways reflect robust activation of type 2 and innate inflammatory programs and are consistent with the elevated immune signaling observed in AD lesions. In contrast, downregulated genes were enriched for a coherent set of metabolic pathways, most notably tyrosine metabolism, retinol metabolism, nicotine addiction, and fatty acid degradation. Suppression of these pathways involves reduced activity of key alcohol dehydrogenases (e.g., ADH1A, ADH1B, ADH6) and cytochrome P450 enzymes (CYP1A1, CYP1A2), suggesting impaired retinoid signaling, diminished oxidative stress regulation, and defective lipid processing in lesional skin. Such metabolic downregulation aligns with the barrier dysfunction characteristic of AD. Collectively, transcriptomic patterns across AD patients reveal that atopic lesional skin is characterized not only by heightened immune activation and broad suppression of metabolic and barrier-associated pathways, but also by substantial transcriptomic heterogeneity among lesions. Identification of Three Transcriptomic Subtypes in AD Skin To determine the number of subtypes that best captures the heterogeneity observed in lesional skin, we evaluated clustering stability across a range of K values using the comprehensive suite of metrics provided by the cola R package (Fig. 2A). The cumulative distribution function (CDF) curves showed the greatest increase in consensus at K = 3, indicating improved cluster separation relative to K = 2. This was further supported by the proportion of ambiguous clustering (PAC), which reached its minimum at K = 3, reflecting the highest cluster robustness. In addition, mean silhouette width peaked at K = 3, indicating optimal within-cluster cohesion and between-cluster separation. Measures of cluster consistency, including concordance, area under the CDF increase, Rand index, and Jaccard index, all showed maximal or near-maximal values at K = 3 compared with higher values of K. Collectively, these metrics converged on K = 3 as the most stable and biologically meaningful solution for resolving transcriptomic variation within the dataset. To further validate the robustness of the three-subtype clusters, we examined the consensus heatmap generated from 50 repeated sK-means partitions (Fig. 2B). The resulting consensus matrix showed three sharply delineated blocks with values approaching 1 within each cluster and near-zero consensus between clusters, demonstrating highly stable and reproducible subtype assignments. Sample-wise silhouette scores were uniformly high across the three groups, indicating strong internal cohesion and clear separation among subtypes. In addition, class probability tracks showed that nearly all samples were assigned to their respective clusters with probability close to 1, reinforcing the reliability of the consensus structure. Together, these results confirm that the three inferred molecular groups represent robust and well-defined transcriptomic subtypes rather than artifacts of clustering variability. In addition to assessing cluster robustness, we next examined the gene signatures that define each of the three subtypes (Fig. 2C). A total of 1,308 signature genes (93.3% with FDR < 0.05) were identified as significantly contributing to subtype separation, and 108 samples were classified as confident based on high silhouette scores and assignment probabilities. The heatmap revealed clear and coherent expression blocks corresponding to the three subgroups, each defined by distinct up- or downregulated gene modules. The key signature genes characterizing each subgroup are provided in Supplementary Table S2. To determine how the three subgroups relate to clinical lesion status, we assessed the enrichment of AL and AN samples within each cluster (Fig. 2D). Subgroup 1 showed a strong over-representation of AN samples, with 43 AN and 6 AL samples (odds ratio = 0.03, P = 4.13 × 10⁻ 14 ), indicating that this cluster corresponds predominantly to non-lesional skin. In contrast, Subgroup 2 exhibited a highly significant enrichment for AL samples (41 AL vs. 7 AN; odds ratio = 17.21, P = 2.31 × 10⁻¹⁰), representing the major lesional subtype. Subgroup 3 also demonstrated AL enrichment (10 AL vs. 4 AN), although with a weaker association (odds ratio = 2.66) and a non-significant p-value (P = 0.153), consistent with this cluster representing a minor lesional subtype with more heterogeneous characteristics. Accordingly, we refer to subgroup 2 as the ‘Major AL subtype’ and subgroup 3 as the ‘Minor AL subtype’ in the following analyses. Subtype -specific transcriptional programs in Major and Minor AL lesions To characterize the major and minor AL subtype-specific transcriptional programs, we first identified DEGs for each subtype relative to its matched non-lesional skin. Major AL (Subgroup 2) exhibited a broad transcriptional shift, with 1,263 upregulated and 674 downregulated genes, whereas minor AL (Subgroup 3) showed a much attenuated response, with only 77 upregulated and 16 downregulated genes, indicating a markedly weaker lesional signature. To contextualize these subtype-specific changes, we next compared these DEG sets with the DEGs obtained from the overall AN vs. AL comparison, irrespective of subgroup membership. A venn diagram revealed that 1,259 genes were shared between the overall AL signature and the major AL subtype, whereas only 58 genes overlapped with the minor AL subtype (Fig. 3A). Major AL subtype also contained 620 subtype-specific DEGs, contrasted with only 35 unique DEGs in minor AL subtype. These results highlight that major AL subtype represents a robust, inflammation-dominant lesional state, whereas minor AL subtype may represent a biologically distinct form of lesional skin with features that differ from major AL subtype and were therefore hypothesized to exhibit unique molecular characteristics Although the major AL and minor AL differed markedly in the magnitude of their transcriptomic alterations, a subset of biological pathways was consistently enriched across both subgroups. Analysis of the genes commonly altered in both AL revealed a compact but coherent set of immune-related processes (Fig. 3B). These results indicate that, although the major AL exhibits a substantially stronger inflammatory signature overall, both subgroups share a conserved interferon-associated innate immune module. The enrichment of type I interferon signaling genes (e.g., IFI27, OAS2, MX1) suggests that interferon-driven epithelial activation represents a core molecular axis present in all AL lesions, independent of subtype-specific severity or transcriptional amplitude. This shared module likely forms the basal inflammatory architecture of AD lesions, upon which subtype-specific programs are superimposed. The DEGs uniquely associated with the major AL subgroup revealed a striking enrichment of immune and inflammatory pathways. Enrichment analysis highlighted robust activation of biological processes such as cytokine–cytokine receptor interaction, inflammatory response, cytokine-mediated signaling, neutrophil migration, and regulation of immune response (Fig. 3B). These pathways showed exceptionally strong significance (adjusted p-values up to 10⁻²³) and involved a broad set of chemokines and immune mediators, including CXCL6, CXCL9, CXCL10, IL22, and SERPINA3. This extensive inflammatory signaling suggests that the major AL represents a highly immune-activated lesional phenotype, capturing the dominant molecular features characteristic of active AD lesions. In contrast to the strongly immune-activated profile of the major AL, the minor AL subtype displayed a markedly different pattern of pathway enrichment. DEGs unique to the minor AL were highly concentrated in cell-cycle and proliferation–related programs, including G2–M checkpoint, E2F targets, mitotic spindle, cell cycle regulation, and oocyte meiosis (Fig. 3B). These pathways demonstrated substantial enrichment (adjusted p-values as low as 10⁻¹⁵) and were driven by canonical cell-cycle regulators such as CDC20, CCNB2, PLK1, BIRC5, and UBE2C. This proliferative signature suggests that the minor AL represents a distinct lesional phenotype characterized not by heightened immune activation but by enhanced keratinocyte cell-cycle progression and mitotic activity. Such features imply that the minor AL may reflect an epidermal remodeling–dominant state that differs fundamentally from the inflammation-dominated biology of the major AL. Immune Landscape Differences Between Major and Minor AL Subtypes Given the pathway-level distinctions identified in Fig. 3, where the Major AL subtype was characterized by strong enrichment of immune and inflammatory pathways, whereas the Minor AL subtype predominantly exhibited cell-cycle–related programs, we hypothesized that these transcriptomic differences may reflect divergent immune-cell compositions between the two lesional subtypes. To test this hypothesis, we performed immune deconvolution using CIBERSORT to estimate the relative abundance of 22 immune cell types across samples. As shown in Fig. 4A, the resulting heatmap revealed clear subgroup-level differences in immune-cell infiltration. Major AL samples displayed broadly elevated levels of both innate and adaptive immune populations, while Minor AL samples exhibited a markedly muted immune signature. To identify immune populations that most strongly differentiated the two subtypes, we conducted statistical comparisons of cell-type fractions. The major AL subtype exhibited pronounced increases in activated CD4 memory T cells and elevated proportions of both resting and activated dendritic cells, indicating a strongly engaged antigen-presentation landscape and heightened helper T-cell activation (Fig. 4B). M1 macrophages were also significantly increased, consistent with a more inflammatory tissue milieu. In contrast, activated NK cells were substantially reduced in the major AL subtype, whereas resting NK-cell levels were relatively preserved. This pattern aligns with prior reports showing numerical reduction and impaired function of NK cells in atopic dermatitis, suggesting suppressed innate lymphoid activity within lesional skin. Neutrophils displayed a slight upward trend in major AL but did not reach statistical significance, indicating that neutrophil infiltration is not a principal discriminator among subtypes in this dataset. Notably, CD8 T cells were reduced in lesional skin, particularly in major AL. This decrease suggests attenuation of cytotoxic T-cell responses, a phenomenon consistent with chronic AD in which persistent type 2 inflammation and regulatory cytokine signaling can dampen or exhaust CD8⁺ T-cell activity. Together, these findings indicate that major AL represents a highly activated immune environment, dominated by dendritic-cell engagement and CD4-mediated adaptive immunity, accompanied by reduced NK and CD8 activity. In contrast, minor AL displays an immune-cell composition more closely resembling non-lesional skin, with relatively lower activation of both adaptive and innate inflammatory populations. This distinction highlights substantial immunologic heterogeneity across lesional tissue and provides insight into the diverse biological states underlying atopic dermatitis. Drug Responsiveness of AL Subtypes Following our identification of transcriptionally distinct AL subtypes, we observed that the major AL subtype exhibits stronger immune-related signatures, whereas the minor AL subtype shows comparatively greater involvement of proliferative and cell-cycle–associated programs. These molecular differences were further supported by immune-cell deconvolution, in which the major subtype demonstrated pronounced immune activation, while the minor subtype displayed an immune composition more similar to non-lesional skin. Given these subtype-specific biological features, we hypothesized that therapeutic responsiveness would likewise differ between subtypes. The GSE157194 dataset includes gene-expression profiles from AD patients before and three months after treatment with dupilumab or cyclosporine, providing an opportunity to evaluate how each AL subtype aligns with drug-induced transcriptional changes. Therefore, we analyzed the correspondence between subtype-specific AL signatures and drug-response signatures to assess potential differences in therapeutic sensitivity. For the major AL subtype, we next examined whether the genes that distinguish lesional from non-lesional skin tend to shift in the opposite direction after dupilumab treatment. Specifically, we examined whether genes that are abnormally increased or decreased in major AL lesions tend to move back toward a non-lesional expression pattern after three months of dupilumab therapy (Fig. 5A). When we compared the lesional signature of major AL with the transcriptomic changes induced by dupilumab, we observed a clear inverse relationship (cosine similarity = –0.575 ): genes that were highly expressed in major AL lesions tended to be reduced after treatment, while genes that were suppressed in lesions tended to be restored. This pattern indicates that dupilumab counteracts the core transcriptional abnormalities of the major AL subtype, suggesting that the inflammatory programs associated with this subtype are responsive to targeted anti-inflammatory therapy. We next evaluated whether cyclosporine treatment also mitigates the transcriptional abnormalities observed in the major AL subtype. Cyclosporine produced an inverse correspondence with the major AL lesional signature, indicating that its immunosuppressive activity partially counteracts the gene-expression shifts characteristic of this subtype (Fig. 5B). However, the magnitude of reversal was more limited than that observed for dupilumab (cosine similarity = –0.453). This pattern aligns with cyclosporine’s mechanism as a calcineurin inhibitor, which broadly suppresses T-cell activation but does not specifically target the upstream cytokine-driven signaling programs reflected in the major AL signature. Nevertheless, the presence of coherent inverse similarity for both dupilumab and cyclosporine suggests that the major AL subtype harbors an inflammatory transcriptional program that is broadly susceptible to immune-modulating therapies and is therefore expected to respond relatively well to either agent. For the minor AL subtype, we similarly assessed whether the transcriptional differences between lesional and non-lesional skin shift in the opposite direction after dupilumab treatment. Although the overall inflammatory signals of this subtype are less pronounced than in major AL, our earlier analyses showed that both subtypes share a core set of immune-related DEGs. Consistent with this, dupilumab produced a strong inverse correspondence with the minor AL lesional signature (cosine similarity = –0.837), indicating that lesion-associated expression changes in this subtype are also robustly reversed following treatment (Fig. 5C). Genes elevated in minor AL lesions generally decreased after therapy, whereas genes reduced in lesions tended to increase. These results suggest that, despite its comparatively subdued immune activation, the minor AL subtype retains immunologic transcriptional features that are highly responsive to targeted anti-inflammatory intervention. In contrast, the relationship between the minor AL lesional signature and the transcriptional response to cyclosporine was considerably weaker (cosine similarity = –0.287) (Fig. 5D). Cyclosporine broadly suppresses T-cell activity, yet our earlier analyses showed that the minor subtype exhibits only modest T-cell–associated abnormalities relative to major AL. As a result, the expression changes induced by cyclosporine only partially align in the reverse direction of the minor AL lesional signature. While some lesion-associated genes shift toward a non-lesional pattern after treatment, the overall correspondence is limited compared with dupilumab. These findings indicate that the molecular features defining the minor subtype are substantially less sensitive to T-cell–directed immunosuppression, consistent with its relatively muted immune engagement. Together, these analyses indicate that the two AL subtypes are likely to exhibit distinct therapeutic responsiveness, with the major subtype showing transcriptional patterns more broadly counteracted by both dupilumab and cyclosporine, whereas the minor subtype appears selectively aligned with dupilumab-mediated reversal. These subtype-specific response signatures suggest that molecular heterogeneity within AL lesions may translate into meaningful differences in treatment outcomes. Clinical Validation of Predicted Cyclosporine Response Patterns We found that the two AL subtypes exhibit distinct patterns of drug responsiveness, particularly in their predicted responses to cyclosporine. Whereas the major subtype showed a clear inverse relationship with cyclosporine-induced expression changes, the minor subtype displayed only a weak correspondence, suggesting that these groups might differ in their clinical sensitivity to calcineurin inhibition. To evaluate whether these subtype-dependent patterns are reflected in real treatment outcomes, we turned to an independent cohort (GSE58558) in which patients with atopic dermatitis were treated with cyclosporine and followed longitudinally. By examining how responders and non-responders in this dataset relate to the transcriptional profiles of the major and minor AL subtypes, we sought to determine whether the molecular distinctions identified earlier translate into clinically observable differences in therapeutic response. Because GSE58558 was generated using a microarray platform, whereas our subtype definitions were derived from the RNA-Seq–based GSE157194 cohort, substantial batch effects were anticipated when integrating the two datasets. As expected, before any correction, the combined expression matrix showed a pronounced separation by dataset rather than by biological signal, as illustrated in Supplementary Fig. S1. Samples clustered almost exclusively by study origin, indicating that platform-driven variation dominated the principal components and would obscure any meaningful comparison between AL subtypes and cyclosporine treatment responses. This strong batch structure necessitated explicit normalization and correction procedures prior to evaluating whether responders and non-responders align with the transcriptional characteristics of the major and minor AL subtypes. To address the substantial platform-driven variation between the RNA-seq cohort and the microarray-based GSE58558 dataset, we implemented a quantile normalization procedure to align their expression distributions. Following this adjustment, the two studies no longer formed distinct clusters; instead, their samples occupied a shared expression space, confirming that the major source of between-cohort divergence had been effectively minimized (Fig. 6A). With technical variation reduced, we then evaluated how the AL subtype structure inferred from GSE157194 relates to transcriptional trajectories observed in patients treated with cyclosporine in GSE58558. Using PCA as a shared reference frame, we mapped responder and non-responder samples collected after cyclosporine treatment—at early (day 1), intermediate (week 2), and late (week 12) time points—onto the expression landscape defined by the major and minor AL subtype signatures. As a first step, we examined how the AL subtypes identified in GSE157194 are positioned within the PC1–PC2 plane. The two groups showed clear separation along these axes, consistent with the substantial transcriptional differences that define the major and minor subtypes. We next overlaid samples from the cyclosporine-treated GSE58558 cohort onto the same PC1–PC2 coordinates to assess their relationship to these subtype signatures. Strikingly, samples collected one day after cyclosporine administration clustered predominantly within the region occupied by major AL lesional samples from GSE157194 (Fig. 6B). This was especially evident for responder patients: at the earliest time point (day 1), most responder profiles mapped closely to the major-lesional cluster, indicating that their immediate transcriptional state under cyclosporine exposure resembled the inflammatory program that characterizes the major subtype. By two weeks after cyclosporine initiation, the transcriptional profiles of treated samples no longer aligned with the major AL lesional region. Instead, both responders and non-responders shifted toward the area of the PC1–PC2 plane occupied by the minor subtype. This transition suggests that the transcriptional programs associated with heightened inflammation and immune activation in major lesions are substantially dampened by this stage of cyclosporine treatment. Notably, non-responders were consistently positioned closer to the minor-subtype cluster than to the major-subtype region at every time point examined—day 1, week 2, and week 12. This persistent proximity to the minor-subtype signature mirrors the prediction derived from Fig. 5D, where the minor subtype was expected to display comparatively limited transcriptional responsiveness to cyclosporine. The PCA trajectories observed here therefore provide clinical support for this prediction: individuals whose baseline or early treatment profiles resemble the minor subtype tend to show a weaker therapeutic response and eventually fall into the non-responder category. To further evaluate how these clinical response categories relate to the AL subtypes, we examined the transcriptional programs associated with cyclosporine non-response. The DEG profile of non-responders showed marked enrichment for cell-cycle and proliferation-related pathways—such as E2F targets, G2–M checkpoint, Myc targets, and mTORC1 signaling—alongside immune-related modules (Fig. 6C). This enrichment pattern closely parallels the molecular characteristics of the minor AL subtype, which, although not devoid of immune features, is comparatively dominated by proliferative transcriptional programs. The similarity between the non-responder signature and the minor subtype therefore reinforces the idea that lesions governed by a minor-like program exhibit weaker cyclosporine responsiveness. Together, these findings demonstrate that the molecular heterogeneity captured by the major and minor AL subtypes is not only biologically meaningful but also clinically consequential. The subtype-specific patterns of transcriptional drug responsiveness predicted from the discovery cohort were recapitulated in an independent cyclosporine-treated cohort, where patients’ response trajectories aligned with the molecular characteristics of their corresponding subtype. These results highlight the potential of transcriptomic subtyping to elucidate treatment variability in atopic dermatitis and underscore its promise as a framework for guiding more personalized therapeutic strategies. Discussion Chronic inflammatory skin diseases, such as psoriasis and AD, share common pathogenic features, including immune cell infiltration and epidermal hyperplasia. However, comparative transcriptomic studies have highlighted fundamental differences in their molecular architecture [15]. Unlike psoriasis, which presents a relatively uniform transcriptomic signature predominantly driven by the Th17/IL-23 axis [16], AD is characterized by profound heterogeneity. This heterogeneity in AD stems from a complex interplay between barrier dysfunction and variable immune polarization involving Th2, Th22, Th1, and Th17 pathways across different phenotypes, ethnicities, and age groups [16,17]. While the distinct molecular uniformity of psoriasis has facilitated the development of highly effective targeted therapies with consistent patient outcomes, the molecular diversity of AD poses a significant challenge, often leading to variable responses to systemic treatments such as cyclosporine and dupilumab [9–11,18,19]. Therefore, deconvoluting this heterogeneity into biologically distinct subtypes is a critical prerequisite for precision medicine in AD. In this study, we sought to address this challenge by systematically characterizing transcriptomic subtypes of AD and examining their biological and therapeutic relevance. Through integrative analysis of lesional and non-lesional skin transcriptomes, we identified two distinct lesional subtypes with divergent molecular features, alongside a non-lesional–like profile. These lesional subtypes differed in the degree of immune activation, cell-cycle–associated and metabolic programs, and immune cell composition. Importantly, the subtypes also exhibited differential responsiveness to systemic therapies, with one subtype showing strong transcriptional concordance with drug-induced expression changes following dupilumab and cyclosporine treatment, while the other demonstrated attenuated responses for cyclosporine. Validation in an independent cyclosporine-treated cohort further supported the clinical relevance of these molecular distinctions. While this study provides a molecular framework that may inform patient stratification and personalized therapeutic strategies in atopic dermatitis, several limitations warrant consideration. A central challenge in AD research lies in the extraordinary clinical and molecular diversity of the disease. Previous large-scale transcriptomic studies have demonstrated that AD exhibits substantial heterogeneity across disease phenotypes, ethnic backgrounds, age groups, and anatomical sites, with variable engagement of Th2-, Th22-, Th17-, and Th1-associated immune programs rather than a single uniform inflammatory signature [15,20,21]. Longitudinal analyses further indicate that AD lesions comprise both stable, patient-specific transcriptional states and dynamic inflammatory components that fluctuate over time [20], underscoring the complexity of molecular variation within and between patients. In this context, although we analyzed multiple independent cohorts and performed cross-cohort validation, the overall sample size of the datasets analyzed here remains modest relative to the full spectrum of AD heterogeneity described in prior studies. It is therefore likely that additional molecular subtypes or intermediate transcriptional states exist that were not captured in the present analysis. Larger, more diverse cohort studies—ideally integrating longitudinal sampling and multi-omic profiling—will be required to fully resolve the breadth of molecular diversity in AD and to further refine transcriptome-based subtyping frameworks. Another important consideration is that transcriptomic profiling alone may not fully capture the complex, multi-layered biology underlying atopic dermatitis. AD pathogenesis is shaped not only by host gene expression but also by interactions with additional molecular layers, including the skin microbiome, epigenetic regulation, and metabolic processes [22–25]. In particular, multiple studies have demonstrated that alterations in the cutaneous microbiome—most notably Staphylococcus aureus overgrowth—are tightly linked to disease severity, immune activation, and barrier dysfunction in AD [24,25]. Integrative analyses combining host transcriptomics with metagenomic profiling have shown that microbial dysbiosis can modulate inflammatory gene expression programs and influence therapeutic response, underscoring a bidirectional relationship between host and microbiota. These findings indicate that transcriptome-based subtyping, while informative, represents only one dimension of disease stratification. Accumulating evidence further indicates that epigenetic regulation constitutes an additional layer of molecular heterogeneity in atopic dermatitis. Genome-wide DNA methylation studies have demonstrated disease- and severity-associated epigenetic alterations in both epidermal and immune-related genes [22,23]. Moreover, dysregulation of non-coding RNAs, including microRNAs implicated in immune activation and keratinocyte differentiation, has been consistently observed in skin diseases [26,27]. These findings suggest that transcriptional heterogeneity in AD is shaped not only by genetic and immunological factors but also by epigenetic mechanisms that integrate environmental and inflammatory cues. Future studies integrating multi-omic layers across larger and longitudinal AD cohorts will be essential to more fully resolve disease mechanisms, refine molecular subtypes, and improve the prediction of treatment response. Despite these limitations, our study provides evidence that AD comprises biologically distinct transcriptional subtypes at the level of lesional skin. By delineating reproducible molecular patterns linked to immune activation, cellular programs, and differential drug responsiveness, this work moves beyond a uniform view of AD as a single inflammatory entity. The identification of subtype-specific transcriptional states suggests that inter-individual variability in treatment outcomes may, at least in part, be rooted in underlying molecular heterogeneity rather than stochastic or purely clinical factors alone. Within this framework, the major lesional subtype is characterized by pronounced immune activation and shows transcriptional features that are concordant with favorable responses to both dupilumab and cyclosporine. This subtype exhibits elevated expression of immune- and inflammation-associated programs, accompanied by increased abundance of activated immune cell populations, indicating an immunologically dominant disease state. Notably, this molecular profile aligns well with the mechanisms of action of both therapies: dupilumab targets key cytokine signaling pathways central to inflammatory cascades in AD, while cyclosporine broadly suppresses T-cell activation and downstream immune responses. In contrast, the minor lesional subtype displays a molecular profile that is less compatible with the immunosuppressive mechanisms of cyclosporine and is associated with attenuated therapeutic responsiveness. This interpretation is supported by the close resemblance between the transcriptional features of the minor subtype and those observed in cyclosporine non-responders from an independent treatment cohort. A prominent hallmark of the minor subtype is the enrichment of cell cycle– and proliferation-associated pathways, which distinguish it from the immune-dominant major subtype. Similar cell cycle–related transcriptional programs were also observed among genes differentially expressed in cyclosporine non-responders, suggesting a shared molecular state that may be less amenable to therapies primarily targeting immune activation. These findings imply that, in a subset of AD patients, disease activity may be sustained by non-immune–dominant programs, potentially contributing to diminished responsiveness to broad immunosuppressive treatment. Our observation that the cyclosporine-resistant minor subtype is characterized by prominent cell-cycle and proliferative signatures is consistent with pathological features commonly associated with chronic, lichenified atopic dermatitis. Previous studies have shown that as AD lesions progress from acute to chronic stages, disease pathology can shift from predominantly immune-cell infiltration toward epidermal hyperplasia (acanthosis) and tissue remodeling [28,29]. In this chronic state, keratinocyte hyperproliferation—potentially sustained by residual cytokine signaling, barrier dysfunction, or mechanical stress such as scratching—may persist even after overt inflammatory infiltrates have diminished, giving rise to a transcriptional profile that is relatively less immune-dominant yet highly proliferative. Notably, the proliferative and keratinization-associated programs enriched in the minor subtype resemble molecular features reported in the so-called “psoriasis-like” phenotype of atopic dermatitis. In particular, Noda et al . demonstrated that Asian AD patients often exhibit enhanced Th17 polarization and pronounced epidermal hyperplasia compared with European American cohorts, reflecting a hybrid molecular state that shares features of both AD and psoriasis [30]. While the present study does not directly assess ethnic stratification, the enrichment of keratinocyte-intrinsic, cell-cycle–related programs in the minor subtype suggests that this transcriptional state may capture chronic or structurally remodeled lesions in which disease activity is less dependent on acute immune activation. Such a shift toward epidermal-driven pathology may help explain the attenuated responsiveness of this subtype to T-cell–targeted immunosuppressive therapy such as cyclosporine, and highlights the potential need for alternative therapeutic strategies aimed at restoring epidermal differentiation and regulating keratinocyte proliferation. In summary, this study demonstrates that AD comprises biologically distinct transcriptional subtypes that differ in immune activation, epidermal programs, and responsiveness to systemic therapies. By integrating transcriptomic subtyping with independent validation of drug response patterns, our findings highlight molecular heterogeneity as a key determinant of therapeutic outcome in AD. These results underscore the importance of moving beyond a uniform treatment paradigm and toward molecularly informed stratification in AD. As transcriptomic and multi-omic profiling approaches continue to advance, defining and validating molecular subtypes may provide a foundation for more precise therapeutic selection and improved clinical outcomes. Declarations Competing interests The authors declare no competing interests. Acknowledgements The authors gratefully acknowledge the support of Korea Polytechnic College for providing computational resources and analytical infrastructure essential for this study. Funding This research was supported by the GR4A25 project at the Bioconvergence Research Institute of HuGeX Co., Ltd. Author contributions S.H.M. implemented the computational analyses, processed the transcriptomic datasets, and conducted downstream statistical and bioinformatic analyses. J.K., J.H., S.H.H., and B.Y. contributed to study design and interpretation of the clinical and experimental context. S.H., J.H.M., H.H., and K.Y. provided biological interpretation of the transcriptomic results and contributed to the analysis of disease mechanisms and pathway-level findings. K.K. contributed to data visualization and figure preparation. S.-E.H. conceived and supervised the overall project, designed the analytical framework, and performed the integrative transcriptomic analyses. All authors contributed to manuscript writing and revision, and approved the final version of the manuscript. Data availability The datasets analyzed in this study are publicly available. Transcriptomic data for atopic dermatitis skin samples and drug-treated cohorts were obtained from the Gene Expression Omnibus (GEO) under accession numbers GSE157194 and GSE58558. References Silverberg, J. I. Public Health Burden and Epidemiology of Atopic Dermatitis. Dermatol Clin 35, 283–289 (2017). Nutten, S. Atopic dermatitis: global epidemiology and risk factors. Ann Nutr Metab 66 Suppl 1, 8–16 (2015). Guttman-Yassky, E., Nograles, K. E. & Krueger, J. G. Contrasting pathogenesis of atopic dermatitis and psoriasis--part I: clinical and pathologic concepts. J Allergy Clin Immunol 127, 1110–8 (2011). Gittler, J. K. et al. Progressive activation of T(H)2/T(H)22 cytokines and selective epidermal proteins characterizes acute and chronic atopic dermatitis. J Allergy Clin Immunol 130, 1344–54 (2012). Leung, D. Y. M., Boguniewicz, M., Howell, M. D., Nomura, I. & Hamid, Q. A. New insights into atopic dermatitis. J Clin Invest 113, 651–7 (2004). He, H. et al. Single-cell transcriptome analysis of human skin identifies novel fibroblast subpopulation and enrichment of immune subsets in atopic dermatitis. J Allergy Clin Immunol 145, 1615–1628 (2020). Wongvibulsin, S. et al. Transcriptomic analysis of atopic dermatitis in African Americans is characterized by Th2/Th17-centered cutaneous immune activation. Sci Rep 11, 11175 (2021). Möbus, L. et al. Atopic dermatitis displays stable and dynamic skin transcriptome signatures. J Allergy Clin Immunol 147, 213–223 (2021). Blauvelt, A. et al. Long-term management of moderate-to-severe atopic dermatitis with dupilumab and concomitant topical corticosteroids (LIBERTY AD CHRONOS): a 1-year, randomised, double-blinded, placebo-controlled, phase 3 trial. Lancet 389, 2287–2303 (2017). Halling, A.-S., Loft, N., Silverberg, J. I., Guttman-Yassky, E. & Thyssen, J. P. Real-world evidence of dupilumab efficacy and risk of adverse events: A systematic review and meta-analysis. J Am Acad Dermatol 84, 139–147 (2021). van der Schaft, J. et al. Drug survival for ciclosporin A in a long-term daily practice cohort of adult patients with atopic dermatitis. Br J Dermatol 172, 1621–1627 (2015). Garritsen, F. M. et al. Ten years experience with oral immunosuppressive treatment in adult patients with atopic dermatitis in two academic centres. J Eur Acad Dermatol Venereol 29, 1905–12 (2015). Nielsen, T. O. et al. A comparison of PAM50 intrinsic subtyping with immunohistochemistry and clinical prognostic factors in tamoxifen-treated estrogen receptor-positive breast cancer. Clin Cancer Res 16, 5222–32 (2010). Parker, J. S. et al. Supervised risk predictor of breast cancer based on intrinsic subtypes. J Clin Oncol 27, 1160–7 (2009). Tsoi, L. C. et al. Atopic Dermatitis Is an IL-13-Dominant Disease with Greater Molecular Heterogeneity Compared to Psoriasis. J Invest Dermatol 139, 1480–1489 (2019). Guttman-Yassky, E., Nograles, K. E. & Krueger, J. G. Contrasting pathogenesis of atopic dermatitis and psoriasis--part II: immune cell subsets and therapeutic concepts. J Allergy Clin Immunol 127, 1420–32 (2011). Noda, S., Krueger, J. G. & Guttman-Yassky, E. The translational revolution and use of biologics in patients with inflammatory skin diseases. J Allergy Clin Immunol 135, 324–36 (2015). de Bruin-Weller, M. S. et al. Biologics to Treat Atopic Dermatitis: Effectiveness, Safety, and Future Directions. Allergy https://doi.org/10.1111/all.70061 (2025) doi:10.1111/all.70061. Khattri, S. et al. Cyclosporine in patients with atopic dermatitis modulates activated inflammatory pathways and reverses epidermal pathology. J Allergy Clin Immunol 133, 1626–34 (2014). Möbus, L. et al. Atopic dermatitis displays stable and dynamic skin transcriptome signatures. J Allergy Clin Immunol 147, 213–223 (2021). Brunner, P. M. Early immunologic changes during the onset of atopic dermatitis. Ann Allergy Asthma Immunol 123, 152–157 (2019). Olisova, O. Y. et al. Skin DNA methylation profile in atopic dermatitis patients: A case-control study. Exp Dermatol 29, 184–189 (2020). Rodríguez, E. et al. An integrated epigenetic and transcriptomic analysis reveals distinct tissue-specific patterns of DNA methylation associated with atopic dermatitis. J Invest Dermatol 134, 1873–1883 (2014). Byrd, A. L. et al. Staphylococcus aureus and Staphylococcus epidermidis strain diversity underlying pediatric atopic dermatitis. Sci Transl Med 9, (2017). Kong, H. H. et al. Temporal shifts in the skin microbiome associated with disease flares and treatment in children with atopic dermatitis. Genome Res 22, 850–9 (2012). Sonkoly, E. et al. MicroRNAs: novel regulators involved in the pathogenesis of psoriasis? PLoS One 2, e610 (2007). Sonkoly, E. et al. MiR-155 is overexpressed in patients with atopic dermatitis and modulates T-cell proliferative responses by targeting cytotoxic T lymphocyte-associated antigen 4. J Allergy Clin Immunol 126, 581-9.e1-20 (2010). Tsoi, L. C. et al. Progression of acute-to-chronic atopic dermatitis is associated with quantitative rather than qualitative changes in cytokine responses. J Allergy Clin Immunol 145, 1406–1415 (2020). Gittler, J. K. et al. Progressive activation of T(H)2/T(H)22 cytokines and selective epidermal proteins characterizes acute and chronic atopic dermatitis. J Allergy Clin Immunol 130, 1344–54 (2012). Noda, S. et al. The Asian atopic dermatitis phenotype combines features of atopic dermatitis and psoriasis with increased TH17 polarization. J Allergy Clin Immunol 136, 1254–64 (2015). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8477069","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":567780196,"identity":"9a61ba74-7914-4ba6-a083-96a33046b946","order_by":0,"name":"Sang Hyun Moh","email":"","orcid":"","institution":"BIO-FD\u0026C Co., Ltd","correspondingAuthor":false,"prefix":"","firstName":"Sang","middleName":"Hyun","lastName":"Moh","suffix":""},{"id":567780197,"identity":"3699a87c-7e46-4bd1-b930-e0c37f727a21","order_by":1,"name":"Jiyeon Kim","email":"","orcid":"","institution":"BIO-FD\u0026C Co., Ltd","correspondingAuthor":false,"prefix":"","firstName":"Jiyeon","middleName":"","lastName":"Kim","suffix":""},{"id":567780198,"identity":"eb614690-e6ef-4d0d-8f68-94dd4da80c91","order_by":2,"name":"Jisoo Han","email":"","orcid":"","institution":"Sungshin Women’s University","correspondingAuthor":false,"prefix":"","firstName":"Jisoo","middleName":"","lastName":"Han","suffix":""},{"id":567780199,"identity":"aa7f6f53-8a08-4842-a5b4-5e3326409c04","order_by":3,"name":"Sun Hee Hwang","email":"","orcid":"","institution":"Sungshin Women’s University","correspondingAuthor":false,"prefix":"","firstName":"Sun","middleName":"Hee","lastName":"Hwang","suffix":""},{"id":567780200,"identity":"ab3c1a76-74cd-4477-9261-81c930e61cd5","order_by":4,"name":"Bumho Yoo","email":"","orcid":"","institution":"HuGeX Co., Ltd","correspondingAuthor":false,"prefix":"","firstName":"Bumho","middleName":"","lastName":"Yoo","suffix":""},{"id":567780201,"identity":"bd1bce8e-188f-4a74-a33a-37dd23ec70db","order_by":5,"name":"Seonghun Hong","email":"","orcid":"","institution":"HuGeX Co., Ltd","correspondingAuthor":false,"prefix":"","firstName":"Seonghun","middleName":"","lastName":"Hong","suffix":""},{"id":567780202,"identity":"dc1db230-443a-42d2-957d-121023f5a155","order_by":6,"name":"Ji Hong Moh","email":"","orcid":"","institution":"HuGeX Co., Ltd","correspondingAuthor":false,"prefix":"","firstName":"Ji","middleName":"Hong","lastName":"Moh","suffix":""},{"id":567780203,"identity":"40fbe5f2-272e-4c82-9b06-c33885425781","order_by":7,"name":"Hyeshin Hwang","email":"","orcid":"","institution":"HuGeX Co., Ltd","correspondingAuthor":false,"prefix":"","firstName":"Hyeshin","middleName":"","lastName":"Hwang","suffix":""},{"id":567780204,"identity":"34982992-3709-451f-8498-a9cb7717d87a","order_by":8,"name":"Kyunghee Yun","email":"","orcid":"","institution":"BIO-FD\u0026C Co., Ltd","correspondingAuthor":false,"prefix":"","firstName":"Kyunghee","middleName":"","lastName":"Yun","suffix":""},{"id":567780205,"identity":"039ef413-221b-484b-9a19-a5ec8b51c6c2","order_by":9,"name":"Kyungmin Kim","email":"","orcid":"","institution":"HuGeX Co., Ltd","correspondingAuthor":false,"prefix":"","firstName":"Kyungmin","middleName":"","lastName":"Kim","suffix":""},{"id":567780206,"identity":"e4686cdc-246a-449b-9cf6-0a6fbe60bd70","order_by":10,"name":"Seong-Eui Hong","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAxElEQVRIiWNgGAWjYFACHhBhYwDjGuBWiaoljXQth0nQIt/ee/Bxwa/zxvzSzQcYftQwGJs3ENBicOZcsvHMvttmknOOJTD2HGMwkzlASItEjpk0b89tG4MbOQYMvA0MNhIEHTb/DUjLObAWxr/EaGG4wWMmzfPjgBlICzPQFjOCWgzO5Bgb8zYkG4P8cljmmIQxYYe1nzF8zPPHzrBfuvngwzc1NoYzCDoMBBjbgATQ9ANgkjjwh4EExaNgFIyCUTDiAACXIDknaPpjlgAAAABJRU5ErkJggg==","orcid":"","institution":"Korea Polytechnic Colleage","correspondingAuthor":true,"prefix":"","firstName":"Seong-Eui","middleName":"","lastName":"Hong","suffix":""}],"badges":[],"createdAt":"2025-12-30 03:53:24","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8477069/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8477069/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":99681868,"identity":"097a55ce-ba89-4aec-802a-289fc6f13e13","added_by":"auto","created_at":"2026-01-07 08:56:08","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":656204,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTranscriptomic differences, heterogeneity, and pathway alterations between lesional and non-lesional skin. \u003c/strong\u003e(A) Heatmap of the differentially expressed genes between lesional (AL) and non-lesional (AN) skin from the GSE157194 dataset. Gene expression values were z-score normalized. (B) Principal component analysis (PCA) of AL and AN transcriptomes. AL samples displayed extensive spread across PC1 and PC2, whereas AN samples formed a compact cluster. Ellipses represent covariance-based confidence regions for each group. (C) Pathway enrichment analysis of significantly upregulated and downregulated genes in AL versus AN. Circle size denotes enrichment odds ratio, and the x-axis shows −log\u003csub\u003e10\u003c/sub\u003e(FDR).\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8477069/v1/dd068c015027d38a24eb018c.png"},{"id":99681863,"identity":"5d4c3516-96e7-463b-b9bd-7d3de6c2fed2","added_by":"auto","created_at":"2026-01-07 08:56:05","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":278374,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eIdentification and characterization of three molecular subgroups of atopic dermatitis using consensus clustering. \u003c/strong\u003e(A) Selection of the optimal number of clusters (k) using consensus clustering.\u003cstrong\u003e \u003c/strong\u003eMultiple Cola-provided metrics—including empirical cumulative distribution function (CDF) curves, ΔPAC, mean silhouette width, concordance, area under the CDF curve, and Jaccard stability.\u003cstrong\u003e \u003c/strong\u003e(B) Consensus matrix heatmap for k = 3. (C) Subtype-defining transcriptomic signatures. (D) Distribution of clinical lesion status (AL vs. AN) across the three subgroups.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8477069/v1/17f557c821cfd9939b2e8711.png"},{"id":99681856,"identity":"c2fe310c-7230-4475-b169-88c10eab16e2","added_by":"auto","created_at":"2026-01-07 08:56:01","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":112905,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDEGs and pathway characteristics of Major and Minor AL subtypes \u003c/strong\u003e(A) Venn diagram illustrating the overlap of DEGs among three comparisons: Overall AN vs. AL, Major AL vs. AN, and Minor AL vs. AN.\u003cstrong\u003e \u003c/strong\u003e(B) Bubble plots summarizing pathway enrichment analyses for (left) common pathways shared by Major and Minor AL, (middle) Major AL–specific pathways, and (right) Minor AL–specific pathways.\u003cstrong\u003e \u003c/strong\u003eThe x-axis indicates –log\u003csub\u003e10\u003c/sub\u003e(FDR) for each enriched pathway.\u003cstrong\u003e \u003c/strong\u003eThe size of each bubble represents the odds ratio, reflecting the strength of enrichment.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8477069/v1/bf717c8e8e07f6f1455619e3.png"},{"id":99681858,"identity":"6d5d18d9-1f10-4c06-a5d7-7e4370c56ef5","added_by":"auto","created_at":"2026-01-07 08:56:01","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":198542,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eImmune cell profiling of atopic dermatitis subtypes. \u003c/strong\u003e(A) CIBERSORT-estimated immune-cell fractions (0–1 scale) visualized as a hierarchical clustered heatmap across non-lesional skin and the two AL subtypes (major AL and minor AL). (B) Boxplots showing immune-cell types with statistically significant differences across the three groups (non-lesional, major AL, minor AL).\u003cstrong\u003e \u003c/strong\u003eSignificance was assessed using pairwise student t-test, and significant comparisons are annotated as: \u003csup\u003e*\u003c/sup\u003ep \u0026lt; 0.05, \u003csup\u003e**\u003c/sup\u003ep \u0026lt; 0.01, \u003csup\u003e***\u003c/sup\u003ep \u0026lt; 0.001, and “n.s.” for non-significant differences.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8477069/v1/74173850b83333dde035df9c.png"},{"id":99681832,"identity":"b2b127c8-a888-42d3-829c-847249622e5f","added_by":"auto","created_at":"2026-01-07 08:55:58","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":242205,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTranscriptional drug responsiveness of AL subtypes \u003c/strong\u003e(A–B) Scatter plots illustrating the transcriptional drug responsiveness of the major AL subtype to\u003cstrong\u003e \u003c/strong\u003e(A) dupilumab and (B) cyclosporine.\u003cstrong\u003e \u003c/strong\u003eThe x-axis represents subtype-specific lesional log₂fold-change (lesional vs. non-lesional), and the y-axis shows drug-induced log₂ fold-change (post- vs. pre-treatment).\u003cstrong\u003e \u003c/strong\u003eGenes defining the major AL subtype (DEGs) are highlighted in color, with all other genes shown in gray.\u003cstrong\u003e \u003c/strong\u003e(C–D) Transcriptional drug responsiveness of the minor AL subtype to\u003cstrong\u003e \u003c/strong\u003e(C) dupilumab and (D) cyclosporine, presented using the same visualization scheme.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-8477069/v1/69900a1306bd0c987a97e4f6.png"},{"id":99681825,"identity":"cd276330-1f68-4265-b404-0565b3fc88b6","added_by":"auto","created_at":"2026-01-07 08:55:57","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":156202,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTranscriptomic alignment of AL subtypes with clinical cyclosporine response patterns. \u003c/strong\u003e(A) PCA of the combined GSE157194 and GSE58558 datasets following rank-based inverse normal transformation, demonstrating effective mitigation of batch effects between the two cohorts.\u003cstrong\u003e \u003c/strong\u003e(B) Projection of cyclosporine-treated samples (GSE58558) collected at day 1, week 2, and week 12 onto the AL subtype PCA space derived from GSE157194. Major and minor AL subtype centroids are shown for reference (★), and responder and non-responder trajectories are visualized as average movement across treatment time points.\u003cstrong\u003e \u003c/strong\u003e(C) Pathway enrichment analysis of genes differentially expressed between non-lesional and lesional skin in cyclosporine non-responders, highlighting prominent enrichment of cell-cycle–associated programs.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-8477069/v1/3549c7ec245ff7d2bed65a68.png"},{"id":99681933,"identity":"59d7cd10-28f2-43e3-8cf6-daef17ec3669","added_by":"auto","created_at":"2026-01-07 08:56:31","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2532272,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8477069/v1/771b1c6f-8649-4855-a6e5-4b6fdf03541d.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Transcriptomic Subtyping of Atopic Dermatitis Reveals Distinct Drug Response Signatures","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAtopic dermatitis (AD) is a common chronic inflammatory skin disease affecting up to 10–20% of children and 2–10% of adults worldwide [1]. The disease is characterized by recurrent eczematous lesions, intense pruritus, and a relapsing clinical course, leading to substantial impairment in quality of life and increased psychosocial burden [2]. In moderate-to-severe cases, AD often requires long-term systemic therapy, underscoring the clinical need for effective and durable treatment strategies.\u003c/p\u003e\n\u003cp\u003eFrom a pathogenic perspective, AD has long been regarded as an immune-mediated disorder, with type 2–driven inflammation playing a central role in disease development and progression [3–5]. However, recent transcriptomic studies have revealed considerable inter-patient variability in the degree and composition of immune involvement [6,7]. Single-cell transcriptomic analyses of human AD skin have demonstrated heterogeneous enrichment of immune cell subsets together with distinct stromal and fibroblast populations, highlighting marked cellular diversity within lesional tissue [6]. Consistent with these observations, bulk transcriptomic profiling across patient cohorts has shown that AD lesions can exhibit variable combinations of Th2- and Th17-associated immune programs rather than a uniform type 2–dominant signature [7]. Moreover, longitudinal transcriptomic analyses suggest that AD lesions comprise both stable, patient-specific molecular signatures and dynamic inflammatory programs that fluctuate with disease activity, underscoring intrinsic molecular heterogeneity beyond transient inflammation [8].\u003c/p\u003e\n\u003cp\u003eThis intrinsic molecular heterogeneity has important clinical implications, particularly with respect to therapeutic response. Despite the availability of targeted and systemic treatments, clinical outcomes in AD remain highly variable, with substantial differences in treatment efficacy observed across patients receiving the same therapy [9,10]. This variability in treatment response has been well documented in both controlled clinical trials and real-world settings. In the pivotal phase 3 LIBERTY AD CHRONOS trial, dupilumab treatment resulted in marked overall clinical improvement at the population level; however, only a subset of patients achieved stringent response criteria. At week 16, Investigator’s Global Assessment (IGA) 0/1 with a\u0026nbsp;≥2-point improvement was observed in approximately 39% of dupilumab-treated patients, while EASI-75 responses were achieved in 64–69%, indicating that a considerable proportion of patients exhibited partial or suboptimal responses despite effective type 2 pathway blockade. [9]. Similar patterns persisted at week 52, underscoring sustained inter-individual variability in treatment outcomes. A large systematic review and meta-analysis encompassing over 3,300 AD patients across 22 observational studies reported pooled response rates of 85.1% for EASI-50, 59.8% for EASI-75, and only 26.8% for EASI-90 after 16 weeks of dupilumab therapy, indicating that a substantial fraction of patients experience partial or incomplete clinical improvement despite treatment [10]. These findings suggest that, beyond overall efficacy, inter-individual differences in underlying molecular and immunological programs likely contribute to differential therapeutic responsiveness in AD.\u003c/p\u003e\n\u003cp\u003eIn addition to biologic therapies, similar heterogeneity in treatment response has been reported for conventional systemic immunosuppressants such as cyclosporine. Although cyclosporine is widely used for moderate-to-severe AD and can induce rapid clinical improvement, both clinical trials and real-world studies have demonstrated substantial inter-individual variability in treatment efficacy and durability. Substantial inter-patient heterogeneity in real-world effectiveness can be quantified by discontinuation patterns (“drug survival”) and the proportion of patients stopping therapy due to lack of benefit. In a large daily-practice cohort of adults with severe AD (n=356), the overall cyclosporine drug survival declined steeply over time (34%, 18%, 12%, and 4% remaining on cyclosporine \u0026nbsp;at 1, 2, 3, and 6 years, respectively; median overall drug survival 256 days), and discontinuations reflected multiple clinical trajectories—including disease control (26.4%), treatment-limiting side effects (22.2%), and ineffectiveness (16.3%), with an additional 6.2% stopping due to combined side effects plus ineffectiveness—indicating that a meaningful subset of patients either cannot tolerate cyclosporine or do not achieve adequate clinical response despite treatment [11]. Consistent estimates were reported in a 10-year experience from two academic centers: among cyclosporine courses that were discontinued (234/267), approximately 15% were stopped as “not effective,” while 24% were discontinued due to adverse events and 29% due to controlled disease (median treatment duration 258 days), reinforcing that cyclosporine response is heterogeneous—ranging from clear clinical control to early termination for either insufficient efficacy or toxicity [12].\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe pronounced inter-individual variability in clinical responses to both dupilumab and cyclosporine observed across previous studies suggests that patients diagnosed with AD may harbor distinct underlying molecular states despite shared clinical features. Such heterogeneity in treatment outcomes is unlikely to be fully explained by differences in drug exposure or adherence alone, and instead points to intrinsic molecular differences that may shape therapeutic responsiveness. These observations highlight the need to move beyond a uniform view of AD as a single inflammatory entity and to systematically characterize patient-level molecular heterogeneity. In this context, defining transcriptional subtypes of lesional skin and linking them to differential drug responses represents a critical step toward improving treatment stratification and advancing precision medicine approaches in AD.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn several complex and heterogeneous diseases, systematic molecular stratification has provided a powerful framework for capturing biological diversity and informing therapeutic decision-making beyond conventional clinical classification. In breast cancer, for example, transcriptome-based subtyping schemes such as the PAM50 classifier have established biologically distinct molecular subtypes with reproducible differences in prognosis and treatment response, and are now routinely integrated into clinical practice [13,14]. By contrast, although atopic dermatitis exhibits marked molecular heterogeneity and variable responses to systemic therapies, a comparably standardized and clinically actionable transcriptomic subtyping framework has not yet been established. This gap highlights the need for molecularly informed approaches to stratify AD patients and to better understand heterogeneity in therapeutic response.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eOn the basis of these considerations, we sought to systematically delineate molecular subtypes of AD using lesional skin transcriptomic profiles and to investigate whether such subtypes are associated with distinct biological characteristics and therapeutic responses. Using an integrative transcriptomic approach, we identified two lesional subgroups with divergent molecular features. These lesional subtypes differed in the extent of immune activation, metabolic and cell-cycle–associated programs, immune cell composition, and responsiveness to systemic therapies. Together, these findings support the existence of biologically distinct AD subtypes at the transcriptomic level and provide a molecular framework for understanding heterogeneity in disease mechanisms and treatment response.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003eData Sources and Preprocessing\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTranscriptomic profiling was performed using the RNA-seq dataset GSE157194, which contains paired lesional (AL) and non-lesional (AN) skin biopsies collected prior to any systemic treatment. After quality filtering and metadata verification, 57 AL samples and 54 AN samples at baseline were included for subtype discovery and differential expression analysis. The dataset also includes 55 follow-up lesional samples collected three months after treatment with either dupilumab or cyclosporine, which were used to assess drug-induced transcriptional changes. Raw count matrices were processed using DESeq2 (v1.46.0), and both variance-stabilizing transformation (VST) and counts per million (CPM) values were generated to obtain stabilized and comparable expression measures for downstream analyses.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe microarray cohort GSE58558, which profiles skin biopsies from atopic dermatitis patients treated with cyclosporine, was downloaded using the GEOquery package. The dataset provides log₂-normalized probe-level intensities together with metadata specifying lesional versus non-lesional sampling and treatment response categories. In total, the cohort includes 50 responder and 6 non-responder lesional samples, as well as 48 responder and 5 non-responder non-lesional samples, spanning multiple treatment time points (day 1, week 2, week 12). Probe identifiers were mapped to HGNC gene symbols using the hgu133plus2.db annotation database. When multiple probes mapped to the same gene symbol, the first matching probe listed in the annotation database was retained.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDifferential Expression Analysis\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eand\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eClustering\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDifferentially expressed genes (DEGs) were identified using DESeq2 applied to raw count matrices for GSE157194. Differential expression was assessed by contrasting non-lesional (AN) and lesional (AL) samples, and genes with |log₂FC|\u0026nbsp;\u0026ge;\u0026nbsp;1 and an adjusted p-value \u0026lt; 0.05 were considered significant. The same criteria were applied to define subtype-specific DEGs for the major and minor AL subtypes by performing non-lesional versus lesional comparisons within each subtype.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eHierarchical clustering was performed using\u0026nbsp;seaborn.clustermap applied to CPM-normalized DEG expression matrices. Before clustering, gene-wise values were standardized (standard_scale=0), and robust scaling was applied (robust=True) to reduce the influence of outliers. Clustering was computed using the default Euclidean distance metric and average linkage method. Principal component analysis (PCA) was applied to the CPM-normalized DEG expression matrix to assess global transcriptional structure. Dimensionality reduction was performed using the standard PCA implementation in\u0026nbsp;sklearn\u0026nbsp;(n_components=2), and the resulting PC1\u0026ndash;PC2 coordinates were used to visualize separation between lesional and non-lesional samples. PCA results were used to visualize separation between lesional and non-lesional groups and to evaluate subtype-level structure.\u003c/p\u003e\n\u003cp\u003eMicroarray-based differential expression analysis was performed separately for the GSE58558 cohort using the limma framework. Raw probe-level intensities were background-corrected and quantile-normalized, after which probes were mapped to gene symbols using platform-specific annotation. A linear modeling approach was applied using limma\u0026rsquo;s empirical Bayes shrinkage, contrasting lesional versus non-lesional samples within the dataset. The criteria used to define DEGs were identical to those applied in the RNA-seq analysis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunctional Enrichment Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFunctional enrichment analysis was performed on DEGs identified from non-lesional versus lesional comparisons. Upregulated and downregulated DEGs were analyzed separately. Enrichment was conducted using\u0026nbsp;gseapy.enrichr, which implements the Enrichr statistical framework. To capture both pathway-level and transcriptional program\u0026ndash;level signals, KEGG 2021 Human and MSigDB Hallmark 2020 gene set libraries were queried. For each DEG subset (UP or DOWN), the gene symbols were submitted as input, and enrichment scores, adjusted P-values, and leading-edge genes were retrieved following Enrichr\u0026rsquo;s default statistical procedures (Fisher\u0026rsquo;s exact test\u0026ndash;based overrepresentation analysis with Benjamini\u0026ndash;Hochberg correction). Only terms passing an adjusted P-value threshold (e.g., FDR \u0026lt; 0.05) were retained for interpretation.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe same enrichment workflow was applied to DEGs defining the major and minor AL subtypes, which were obtained from subtype-specific lesional versus non-lesional contrasts. For each subtype, upregulated and downregulated DEGs were submitted separately to Enrichr , and enriched pathways were summarized to characterize biological programs distinguishing the two transcriptomic subtypes\u003c/p\u003e\n\u003cp\u003eDifferential expression analysis of the GSE58558 cyclosporine cohort produced an insufficient number of DEGs for conventional over-representation analysis. To enable pathway-level interpretation despite the limited DEG count, we applied preranked gene set enrichment analysis (GSEA) using the full transcriptome ranked by log2 fold change. Enrichment was performed with gseapy.prerank against the MSigDB Hallmark 2020 and KEGG 2021 Human gene sets, and significance was defined using FDR-adjusted q-values (FDR \u0026lt; 0.05).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMolecular Subtyping of Atopic Dermatitis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo identify molecular subgroups within atopic lesional skin, we applied the cola framework for consensus partitioning using the VST-normalized expression matrix derived from DEGs of non-lesion vs. lesion of patients. Cola was run using its standard workflow, which combines feature selection, iterative subsampling, and stability assessment across multiple partitioning runs. We selected the skmeans method as the primary clustering algorithm, as recommended for high-dimensional transcriptome data, and evaluated cluster numbers from k = 2 to 6. At each k, cola repeatedly partitions the data (50 iterations in our analysis) and computes consensus matrices that summarize how consistently samples co-cluster across runs.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eEvaluation metrics provided by cola\u0026mdash;including the mean silhouette width, proportion of ambiguous clustering (PAC) score, and consensus heatmap structure. Samples were assigned to subtypes based on their consensus membership, and to ensure high subtype fidelity, only those with silhouette \u0026ge; 0.5 were retained for downstream signature analysis. This filtering step ensured that ambiguous or borderline samples did not confound subtype-specific interpretations.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo assess whether each molecular subtype was preferentially associated with lesional versus non-lesional skin, we performed a fisher\u0026rsquo;s exact test for enrichment of pathology status within each cluster. For a given subtype, we constructed a 2\u0026times;2 contingency table with rows corresponding to tissue type (AL vs. AN) and columns indicating whether a sample belonged to the subtype of interest (\u0026ldquo;in\u0026rdquo;) or to any of the other subtypes (\u0026ldquo;out\u0026rdquo;). Fisher\u0026rsquo;s exact test was applied to each table to obtain an odds ratio and p-value. An odds ratio greater than 1 was interpreted as AL enrichment for that subtype, whereas an odds ratio less than 1 indicated AN enrichment; odds ratios close to 1 were considered to reflect no clear enrichment.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eImmune Cell Deconvolution\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo estimate immune-cell composition from bulk transcriptomic data, we applied CIBERSORT (v.0.1.0) using the LM22 leukocyte signature matrix, which infers the relative abundance of 22 immune cell subsets. Normalized expression matrices from RNA-Seq (VST) was formatted according to CIBERSORT requirements and analyzed in relative fractions mode. All samples were processed using 1,000 permutations, and only deconvolution results with p \u0026lt; 0.05 were retained for downstream analysis.\u003c/p\u003e\n\u003cp\u003eDifferences in immune-cell composition among non-lesional skin, major AL, and minor AL subtypes were assessed using the CIBERSORT-estimated cell fractions. For each cell type, we performed pairwise two-sided Welch\u0026rsquo;s t-tests to compare (i) AN vs. major AL, (ii) AN vs. minor AL, and (iii) major vs. minor AL. Cell types with nominal p-values \u0026lt; 0.05 were considered to show statistically significant subtype-associated differences and were subsequently highlighted in box-plot visualizations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDrug Responsiveness Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo characterize transcriptional responses to dupilumab and cyclosporine, lesional skin samples from GSE157194 collected before treatment and three months after treatment were used to compute drug-induced expression changes. For each drug, log₂ fold-changes were obtained by contrasting post-treatment versus pre-treatment lesional samples. These drug-response vectors were then compared with the subtype-specific lesional signatures (major and minor AL) by computing cosine similarity.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eBatch Effect Correction Between Cohorts\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo harmonize gene expression profiles across the two independent cohorts (GSE157194 and GSE58558), batch effects were mitigated using a rank-based inverse normal transformation (INT). For each gene, expression values were independently replaced by their sample-wise ranks, which were then mapped onto a standard normal distribution using the transformation:\u003c/p\u003e\n\u003cp\u003eWhere \u0026gamma;\u003csub\u003ei\u003c/sub\u003e is the rank of sample i for the given gene, N is the number of samples, and \u0026Phi;\u003csup\u003e-1\u003c/sup\u003e denotes the inverse cumulative distribution function of the standard normal distribution. This procedure produces a rank-preserving, distribution-normalized expression matrix that minimizes batch-driven shifts in mean or variance while maintaining relative ordering within each gene.\u003c/p\u003e\n\u003cp\u003eFollowing batch correction, principal component analysis (PCA) was applied to the combined expression matrix to assess whether samples from the two cohorts remained separated or were mixed in the reduced-dimensional space. PCA results before and after correction were visually compared to confirm the effective reduction of batch-associated clustering.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCyclosporine Treatment Trajectory Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo characterize longitudinal transcriptional dynamics following cyclosporine treatment, we computed trajectory paths for responders and non-responders using the batch-corrected PCA space derived from the integrated GSE157194 and GSE58558 datasets. Each GSE58558 sample was annotated with treatment time point (day 1, week 2, week 12) and response status (responder or non-responder). PCA coordinates (PC1 and PC2) of these samples were then used to estimate temporal movement in transcriptomic space. For each response group, samples were first grouped by treatment time point, and the median PC1 and PC2 values within each time point were calculated to obtain a centroid representing the group\u0026rsquo;s average transcriptomic state at that stage. Centroids were computed sequentially for day 1, week 2 and week 12. Trajectory paths were visualized by connecting these centroids in temporal order.\u003c/p\u003e"},{"header":"Results ","content":"\u003cp\u003e\u003cstrong\u003eTranscriptomic Differences Between Atopic Lesional and Non-Lesional Skin\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAtopic dermatitis (AD) presents with diverse clinical manifestations, suggesting that patients may harbor distinct molecular states despite similar phenotypes. We hypothesized that such transcriptomic heterogeneity contributes to the variability in disease severity and treatment response, and that identifying these differences would provide a basis for stratified therapeutic strategies. To evaluate this possibility, we analyzed the GSE157194 dataset, which contains paired lesional and non-lesional skin samples from AD patients, focusing on characterizing the global transcriptional alterations associated with lesional skin. The dataset consisted of 57 atopic lesional (AL) and 54 atopic non-lesional (AN) samples collected prior to treatment. Differential expression analysis was performed using DESeq2, and genes with |log₂fold-change| \u0026ge; 1 and FDR \u0026lt; 0.05 were considered significant. This analysis identified 1,047 upregulated and 355 downregulated genes in AL compared with AN, demonstrating substantial transcriptional remodeling associated with lesional skin (Supplementary Table S1). Clustering analysis based on the identified differentially expressed genes (DEGs) revealed that samples from AL and AN skin did not form completely distinct groups, with several AL and AN samples intermingling within the same clusters (Fig. 1A). This lack of strict separation indicates that the transcriptional profiles of AD skin exist on a continuum rather than as two discrete states. PCA demonstrated incomplete separation between lesional and non-lesional skin samples along the major variance component (PC1, 78.2%) (Fig. 1B). AN samples formed a compact cluster, whereas AL samples displayed a wide dispersion with a broad confidence ellipse, indicating substantial heterogeneity among lesional transcriptomes. These findings suggest that heterogeneous molecular states exist within AD lesions, reflecting differing degrees or types of pathogenic activation across patients.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo characterize the biological processes underlying the transcriptional differences between lesional (AL) and non-lesional (AN) skin, we performed pathway enrichment analysis on the differentially expressed genes (Fig. 1C). Upregulated genes in AL were predominantly associated with immune activation pathways, including cytokine\u0026ndash;cytokine receptor interaction, IL-17 signaling, inflammatory response, and allograft rejection. These pathways reflect robust activation of type 2 and innate inflammatory programs and are consistent with the elevated immune signaling observed in AD lesions. In contrast, downregulated genes were enriched for a coherent set of metabolic pathways, most notably tyrosine metabolism, retinol metabolism, nicotine addiction, and fatty acid degradation. Suppression of these pathways involves reduced activity of key alcohol dehydrogenases (e.g., ADH1A, ADH1B, ADH6) and cytochrome P450 enzymes (CYP1A1, CYP1A2), suggesting impaired retinoid signaling, diminished oxidative stress regulation, and defective lipid processing in lesional skin. Such metabolic downregulation aligns with the barrier dysfunction characteristic of AD.\u003c/p\u003e\n\u003cp\u003eCollectively, transcriptomic patterns across AD patients reveal that atopic lesional skin is characterized not only by heightened immune activation and broad suppression of metabolic and barrier-associated pathways, but also by substantial transcriptomic heterogeneity among lesions.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eIdentification of Three Transcriptomic Subtypes in AD Skin\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo determine the number of subtypes that best captures the heterogeneity observed in lesional skin, we evaluated clustering stability across a range of K values using the comprehensive suite of metrics provided by the cola R package (Fig. 2A). The cumulative distribution function (CDF) curves showed the greatest increase in consensus at K = 3, indicating improved cluster separation relative to K = 2. This was further supported by the proportion of ambiguous clustering (PAC), which reached its minimum at K = 3, reflecting the highest cluster robustness. In addition, mean silhouette width peaked at K = 3, indicating optimal within-cluster cohesion and between-cluster separation. Measures of cluster consistency, including concordance, area under the CDF increase, Rand index, and Jaccard index, all showed maximal or near-maximal values at K = 3 compared with higher values of K. Collectively, these metrics converged on K = 3 as the most stable and biologically meaningful solution for resolving transcriptomic variation within the dataset.\u003c/p\u003e\n\u003cp\u003eTo further validate the robustness of the three-subtype clusters, we examined the consensus heatmap generated from 50 repeated sK-means partitions (Fig. 2B). The resulting consensus matrix showed three sharply delineated blocks with values approaching 1 within each cluster and near-zero consensus between clusters, demonstrating highly stable and reproducible subtype assignments. Sample-wise silhouette scores were uniformly high across the three groups, indicating strong internal cohesion and clear separation among subtypes. In addition, class probability tracks showed that nearly all samples were assigned to their respective clusters with probability close to 1, reinforcing the reliability of the consensus structure. Together, these results confirm that the three inferred molecular groups represent robust and well-defined transcriptomic subtypes rather than artifacts of clustering variability.\u003c/p\u003e\n\u003cp\u003eIn addition to assessing cluster robustness, we next examined the gene signatures that define each of the three subtypes (Fig. 2C). A total of 1,308 signature genes (93.3% with FDR \u0026lt; 0.05) were identified as significantly contributing to subtype separation, and 108 samples were classified as confident based on high silhouette scores and assignment probabilities. The heatmap revealed clear and coherent expression blocks corresponding to the three subgroups, each defined by distinct up- or downregulated gene modules. The key signature genes characterizing each subgroup are provided in Supplementary Table S2.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo determine how the three subgroups relate to clinical lesion status, we assessed the enrichment of AL and AN samples within each cluster (Fig. 2D). Subgroup 1 showed a strong over-representation of AN samples, with 43 AN and 6 AL samples (odds ratio = 0.03, P = 4.13 \u0026times; 10⁻\u003csup\u003e14\u003c/sup\u003e), indicating that this cluster corresponds predominantly to non-lesional skin. In contrast, Subgroup 2 exhibited a highly significant enrichment for AL samples (41 AL vs. 7 AN; odds ratio = 17.21, P = 2.31 \u0026times; 10⁻\u0026sup1;⁰), representing the major lesional subtype. Subgroup 3 also demonstrated AL enrichment (10 AL vs. 4 AN), although with a weaker association (odds ratio = 2.66) and a non-significant p-value (P = 0.153), consistent with this cluster representing a minor lesional subtype with more heterogeneous characteristics. Accordingly, we refer to subgroup 2 as the \u0026lsquo;Major AL subtype\u0026rsquo; and subgroup 3 as the \u0026lsquo;Minor AL subtype\u0026rsquo; in the following analyses.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSubtype\u003c/strong\u003e\u003cstrong\u003e-specific transcriptional programs in Major and Minor AL lesions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo characterize the major and minor AL subtype-specific transcriptional programs, we first identified DEGs for each subtype relative to its matched non-lesional skin. Major AL (Subgroup 2) exhibited a broad transcriptional shift, with 1,263 upregulated and 674 downregulated genes, whereas minor AL (Subgroup 3) showed a much attenuated response, with only 77 upregulated and 16 downregulated genes, indicating a markedly weaker lesional signature. To contextualize these subtype-specific changes, we next compared these DEG sets with the DEGs obtained from the overall AN vs. AL comparison, irrespective of subgroup membership. A venn diagram revealed that 1,259 genes were shared between the overall AL signature and the major AL subtype, whereas only 58 genes overlapped with the minor AL subtype (Fig. 3A). Major AL subtype also contained 620 subtype-specific DEGs, contrasted with only 35 unique DEGs in minor AL subtype. These results highlight that major AL subtype represents a robust, inflammation-dominant lesional state, whereas minor AL subtype may represent a biologically distinct form of lesional skin with features that differ from major AL subtype and were therefore hypothesized to exhibit unique molecular characteristics\u003c/p\u003e\n\u003cp\u003eAlthough the major AL and minor AL differed markedly in the magnitude of their transcriptomic alterations, a subset of biological pathways was consistently enriched across both subgroups. Analysis of the genes commonly altered in both AL revealed a compact but coherent set of immune-related processes (Fig. 3B). These results indicate that, although the major AL exhibits a substantially stronger inflammatory signature overall, both subgroups share a conserved interferon-associated innate immune module. The enrichment of type I interferon signaling genes (e.g., IFI27, OAS2, MX1) suggests that interferon-driven epithelial activation represents a core molecular axis present in all AL lesions, independent of subtype-specific severity or transcriptional amplitude. This shared module likely forms the basal inflammatory architecture of AD lesions, upon which subtype-specific programs are superimposed.\u003c/p\u003e\n\u003cp\u003eThe DEGs uniquely associated with the major AL subgroup revealed a striking enrichment of immune and inflammatory pathways. Enrichment analysis highlighted robust activation of biological processes such as cytokine\u0026ndash;cytokine receptor interaction, inflammatory response, cytokine-mediated signaling, neutrophil migration, and regulation of immune response (Fig. 3B). These pathways showed exceptionally strong significance (adjusted p-values up to 10⁻\u0026sup2;\u0026sup3;) and involved a broad set of chemokines and immune mediators, including CXCL6, CXCL9, CXCL10, IL22, and SERPINA3. This extensive inflammatory signaling suggests that the major AL represents a highly immune-activated lesional phenotype, capturing the dominant molecular features characteristic of active AD lesions.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn contrast to the strongly immune-activated profile of the major AL, the minor AL subtype displayed a markedly different pattern of pathway enrichment. DEGs unique to the minor AL were highly concentrated in cell-cycle and proliferation\u0026ndash;related programs, including G2\u0026ndash;M checkpoint, E2F targets, mitotic spindle, cell cycle regulation, and oocyte meiosis (Fig. 3B). These pathways demonstrated substantial enrichment (adjusted p-values as low as 10⁻\u0026sup1;⁵) and were driven by canonical cell-cycle regulators such as CDC20, CCNB2, PLK1, BIRC5, and UBE2C. This proliferative signature suggests that the minor AL represents a distinct lesional phenotype characterized not by heightened immune activation but by enhanced keratinocyte cell-cycle progression and mitotic activity. Such features imply that the minor AL may reflect an epidermal remodeling\u0026ndash;dominant state that differs fundamentally from the inflammation-dominated biology of the major AL.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eImmune Landscape Differences Between Major and Minor AL\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eSubtypes\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eGiven the pathway-level distinctions identified in Fig.\u0026nbsp;3, where the Major AL subtype\u0026nbsp;was characterized by strong enrichment of immune and inflammatory pathways, whereas the Minor AL subtype\u0026nbsp;predominantly exhibited cell-cycle\u0026ndash;related programs, we hypothesized that these transcriptomic differences may reflect divergent immune-cell compositions between the two lesional subtypes.\u0026nbsp;To test this hypothesis, we performed immune deconvolution using CIBERSORT to estimate the relative abundance of 22 immune cell types across samples. As shown in Fig.\u0026nbsp;4A, the resulting heatmap revealed clear subgroup-level differences in immune-cell infiltration. Major AL samples displayed broadly elevated levels of both innate and adaptive immune populations, while Minor AL samples exhibited a markedly muted immune signature.\u0026nbsp;To identify immune populations that most strongly differentiated the two subtypes, we conducted statistical comparisons of cell-type fractions.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe major AL subtype exhibited pronounced increases in activated CD4 memory T cells and elevated proportions of both resting and activated dendritic cells, indicating a strongly engaged antigen-presentation landscape and heightened helper T-cell activation\u0026nbsp;(Fig.\u0026nbsp;4B). M1 macrophages were also significantly increased, consistent with a more inflammatory tissue milieu. In contrast, activated NK cells were substantially reduced in the major AL subtype, whereas resting NK-cell levels were relatively preserved. This pattern aligns with prior reports showing numerical reduction and impaired function of NK cells in atopic dermatitis, suggesting suppressed innate lymphoid activity within lesional skin.\u0026nbsp;Neutrophils displayed a slight upward trend in major AL but did not reach statistical significance, indicating that neutrophil infiltration is not a principal discriminator among subtypes in this dataset. Notably, CD8 T cells were reduced in lesional skin, particularly in major AL. This decrease suggests attenuation of cytotoxic T-cell responses, a phenomenon consistent with chronic AD in which persistent type 2 inflammation and regulatory cytokine signaling can dampen or exhaust CD8⁺ T-cell activity.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTogether, these findings indicate that major AL represents a highly activated immune environment, dominated by dendritic-cell engagement and CD4-mediated adaptive immunity, accompanied by reduced NK and CD8 activity. In contrast, minor AL displays an immune-cell composition more closely resembling non-lesional skin, with relatively lower activation of both adaptive and innate inflammatory populations. This distinction highlights substantial immunologic heterogeneity across lesional tissue and provides insight into the diverse biological states underlying atopic dermatitis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDrug Responsiveness of AL Subtypes\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFollowing our identification of transcriptionally distinct AL subtypes, we observed that the major AL subtype exhibits stronger immune-related signatures, whereas the minor AL subtype shows comparatively greater involvement of proliferative and cell-cycle\u0026ndash;associated programs. These molecular differences were further supported by immune-cell deconvolution, in which the major subtype demonstrated pronounced immune activation, while the minor subtype displayed an immune composition more similar to non-lesional skin. Given these subtype-specific biological features, we hypothesized that therapeutic responsiveness would likewise differ between subtypes. The GSE157194 dataset includes gene-expression profiles from AD patients before and three months after treatment with dupilumab or cyclosporine, providing an opportunity to evaluate how each AL subtype aligns with drug-induced transcriptional changes. Therefore, we analyzed the correspondence between subtype-specific AL signatures and drug-response signatures to assess potential differences in therapeutic sensitivity.\u003c/p\u003e\n\u003cp\u003eFor the major AL subtype, we next examined whether the genes that distinguish lesional from non-lesional skin tend to shift in the opposite direction after dupilumab treatment. Specifically, we examined whether genes that are abnormally increased or decreased in major AL lesions tend to move back toward a non-lesional expression pattern after three months of dupilumab therapy\u0026nbsp;\u003cbr\u003e(Fig. 5A). When we compared the lesional signature of major AL with the transcriptomic changes induced by dupilumab, we observed a clear inverse relationship (cosine similarity = \u003cstrong\u003e\u0026ndash;0.575\u003c/strong\u003e): genes that were highly expressed in major AL lesions tended to be reduced after treatment, while genes that were suppressed in lesions tended to be restored. This pattern indicates that dupilumab counteracts the core transcriptional abnormalities of the major AL subtype, suggesting that the inflammatory programs associated with this subtype are responsive to targeted anti-inflammatory therapy.\u003c/p\u003e\n\u003cp\u003eWe next evaluated whether cyclosporine treatment also mitigates the transcriptional abnormalities observed in the major AL subtype. Cyclosporine produced an inverse correspondence with the major AL lesional signature, indicating that its immunosuppressive activity partially counteracts the gene-expression shifts characteristic of this subtype (Fig. 5B). However, the magnitude of reversal was more limited than that observed for dupilumab (cosine similarity = \u0026ndash;0.453). This pattern aligns with cyclosporine\u0026rsquo;s mechanism as a calcineurin inhibitor, which broadly suppresses T-cell activation but does not specifically target the upstream cytokine-driven signaling programs reflected in the major AL signature. Nevertheless, the presence of coherent inverse similarity for both dupilumab and cyclosporine suggests that the major AL subtype harbors an inflammatory transcriptional program that is broadly susceptible to immune-modulating therapies and is therefore expected to respond relatively well to either agent.\u003c/p\u003e\n\u003cp\u003eFor the minor AL subtype, we similarly assessed whether the transcriptional differences between lesional and non-lesional skin shift in the opposite direction after dupilumab treatment. Although the overall inflammatory signals of this subtype are less pronounced than in major AL, our earlier analyses showed that both subtypes share a core set of immune-related DEGs. Consistent with this, dupilumab produced a strong inverse correspondence with the minor AL lesional signature (cosine similarity = \u0026ndash;0.837), indicating that lesion-associated expression changes in this subtype are also robustly reversed following treatment (Fig. 5C). Genes elevated in minor AL lesions generally decreased after therapy, whereas genes reduced in lesions tended to increase. These results suggest that, despite its comparatively subdued immune activation, the minor AL subtype retains immunologic transcriptional features that are highly responsive to targeted anti-inflammatory intervention.\u003c/p\u003e\n\u003cp\u003eIn contrast, the relationship between the minor AL lesional signature and the transcriptional response to cyclosporine was considerably weaker (cosine similarity = \u0026ndash;0.287) (Fig. 5D). Cyclosporine broadly suppresses T-cell activity, yet our earlier analyses showed that the minor subtype exhibits only modest T-cell\u0026ndash;associated abnormalities relative to major AL. As a result, the expression changes induced by cyclosporine only partially align in the reverse direction of the minor AL lesional signature. While some lesion-associated genes shift toward a non-lesional pattern after treatment, the overall correspondence is limited compared with dupilumab. These findings indicate that the molecular features defining the minor subtype are substantially less sensitive to T-cell\u0026ndash;directed immunosuppression, consistent with its relatively muted immune engagement.\u003c/p\u003e\n\u003cp\u003eTogether, these analyses indicate that the two AL subtypes are likely to exhibit distinct therapeutic responsiveness, with the major subtype showing transcriptional patterns more broadly counteracted by both dupilumab and cyclosporine, whereas the minor subtype appears selectively aligned with dupilumab-mediated reversal. These subtype-specific response signatures suggest that molecular heterogeneity within AL lesions may translate into meaningful differences in treatment outcomes.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical Validation of Predicted Cyclosporine Response Patterns\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe found that the two AL subtypes exhibit distinct patterns of drug responsiveness, particularly in their predicted responses to cyclosporine. Whereas the major subtype showed a clear inverse relationship with cyclosporine-induced expression changes, the minor subtype displayed only a weak correspondence, suggesting that these groups might differ in their clinical sensitivity to calcineurin inhibition. To evaluate whether these subtype-dependent patterns are reflected in real treatment outcomes, we turned to an independent cohort (GSE58558) in which patients with atopic dermatitis were treated with cyclosporine and followed longitudinally. By examining how responders and non-responders in this dataset relate to the transcriptional profiles of the major and minor AL subtypes, we sought to determine whether the molecular distinctions identified earlier translate into clinically observable differences in therapeutic response.\u003c/p\u003e\n\u003cp\u003eBecause GSE58558 was generated using a microarray platform, whereas our subtype definitions were derived from the RNA-Seq\u0026ndash;based GSE157194 cohort, substantial batch effects were anticipated when integrating the two datasets. As expected, before any correction, the combined expression matrix showed a pronounced separation by dataset rather than by biological signal, as illustrated in Supplementary Fig. S1. Samples clustered almost exclusively by study origin, indicating that platform-driven variation dominated the principal components and would obscure any meaningful comparison between AL subtypes and cyclosporine treatment responses. This strong batch structure necessitated explicit normalization and correction procedures prior to evaluating whether responders and non-responders align with the transcriptional characteristics of the major and minor AL subtypes.\u003c/p\u003e\n\u003cp\u003eTo address the substantial platform-driven variation between the RNA-seq cohort and the microarray-based GSE58558 dataset, we implemented a quantile normalization procedure to align their expression distributions. Following this adjustment, the two studies no longer formed distinct clusters; instead, their samples occupied a shared expression space, confirming that the major source of between-cohort divergence had been effectively minimized (Fig. 6A). With technical variation reduced, we then evaluated how the AL subtype structure inferred from GSE157194 relates to transcriptional trajectories observed in patients treated with cyclosporine in GSE58558. Using PCA as a shared reference frame, we mapped responder and non-responder samples collected after cyclosporine treatment\u0026mdash;at early (day 1), intermediate (week 2), and late (week 12) time points\u0026mdash;onto the expression landscape defined by the major and minor AL subtype signatures.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAs a first step, we examined how the AL subtypes identified in GSE157194 are positioned within the PC1\u0026ndash;PC2 plane. The two groups showed clear separation along these axes, consistent with the substantial transcriptional differences that define the major and minor subtypes. We next overlaid samples from the cyclosporine-treated GSE58558 cohort onto the same PC1\u0026ndash;PC2 coordinates to assess their relationship to these subtype signatures. Strikingly, samples collected one day after cyclosporine administration clustered predominantly within the region occupied by major AL lesional samples from GSE157194 (Fig. 6B). This was especially evident for responder patients: at the earliest time point (day 1), most responder profiles mapped closely to the major-lesional cluster, indicating that their immediate transcriptional state under cyclosporine exposure resembled the inflammatory program that characterizes the major subtype.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eBy two weeks after cyclosporine initiation, the transcriptional profiles of treated samples no longer aligned with the major AL lesional region. Instead, both responders and non-responders shifted toward the area of the PC1\u0026ndash;PC2 plane occupied by the minor subtype. This transition suggests that the transcriptional programs associated with heightened inflammation and immune activation in major lesions are substantially dampened by this stage of cyclosporine treatment. Notably, non-responders were consistently positioned closer to the minor-subtype cluster than to the major-subtype region at every time point examined\u0026mdash;day 1, week 2, and week 12. This persistent proximity to the minor-subtype signature mirrors the prediction derived from Fig. 5D, where the minor subtype was expected to display comparatively limited transcriptional responsiveness to cyclosporine. The PCA trajectories observed here therefore provide clinical support for this prediction: individuals whose baseline or early treatment profiles resemble the minor subtype tend to show a weaker therapeutic response and eventually fall into the non-responder category. To further evaluate how these clinical response categories relate to the AL subtypes, we examined the transcriptional programs associated with cyclosporine non-response. The DEG profile of non-responders showed marked enrichment for cell-cycle and proliferation-related pathways\u0026mdash;such as E2F targets, G2\u0026ndash;M checkpoint, Myc targets, and mTORC1 signaling\u0026mdash;alongside immune-related modules (Fig. 6C). This enrichment pattern closely parallels the molecular characteristics of the minor AL subtype, which, although not devoid of immune features, is comparatively dominated by proliferative transcriptional programs. The similarity between the non-responder signature and the minor subtype therefore reinforces the idea that lesions governed by a minor-like program exhibit weaker cyclosporine responsiveness.\u003c/p\u003e\n\u003cp\u003eTogether, these findings demonstrate that the molecular heterogeneity captured by the major and minor AL subtypes is not only biologically meaningful but also clinically consequential. The subtype-specific patterns of transcriptional drug responsiveness predicted from the discovery cohort were recapitulated in an independent cyclosporine-treated cohort, where patients\u0026rsquo; response trajectories aligned with the molecular characteristics of their corresponding subtype. These results highlight the potential of transcriptomic subtyping to elucidate treatment variability in atopic dermatitis and underscore its promise as a framework for guiding more personalized therapeutic strategies.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eChronic inflammatory skin diseases, such as psoriasis and AD, share common pathogenic features, including immune cell infiltration and epidermal hyperplasia. However, comparative transcriptomic studies have highlighted fundamental differences in their molecular architecture [15]. Unlike psoriasis, which presents a relatively uniform transcriptomic signature predominantly driven by the Th17/IL-23 axis [16], AD is characterized by profound heterogeneity. This heterogeneity in AD stems from a complex interplay between barrier dysfunction and variable immune polarization involving Th2, Th22, Th1, and Th17 pathways across different phenotypes, ethnicities, and age groups [16,17]. While the distinct molecular uniformity of psoriasis has facilitated the development of highly effective targeted therapies with consistent patient outcomes, the molecular diversity of AD poses a significant challenge, often leading to variable responses to systemic treatments such as cyclosporine and dupilumab [9–11,18,19]. Therefore, deconvoluting this heterogeneity into biologically distinct subtypes is a critical prerequisite for precision medicine in AD.\u003c/p\u003e\n\u003cp\u003eIn this study, we sought to address this challenge by systematically characterizing transcriptomic subtypes of AD and examining their biological and therapeutic relevance. Through integrative analysis of lesional and non-lesional skin transcriptomes, we identified two distinct lesional subtypes with divergent molecular features, alongside a non-lesional–like profile. These lesional subtypes differed in the degree of immune activation, cell-cycle–associated and metabolic programs, and immune cell composition. Importantly, the subtypes also exhibited differential responsiveness to systemic therapies, with one subtype showing strong transcriptional concordance with drug-induced expression changes following dupilumab and cyclosporine treatment, while the other demonstrated attenuated responses for cyclosporine. Validation in an independent cyclosporine-treated cohort further supported the clinical relevance of these molecular distinctions.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWhile this study provides a molecular framework that may inform patient stratification and personalized therapeutic strategies in atopic dermatitis, several limitations warrant consideration. A central challenge in AD research lies in the extraordinary clinical and molecular diversity of the disease. Previous large-scale transcriptomic studies have demonstrated that AD exhibits substantial heterogeneity across disease phenotypes, ethnic backgrounds, age groups, and anatomical sites, with variable engagement of Th2-, Th22-, Th17-, and Th1-associated immune programs rather than a single uniform inflammatory signature [15,20,21]. Longitudinal analyses further indicate that AD lesions comprise both stable, patient-specific transcriptional states and dynamic inflammatory components that fluctuate over time [20], underscoring the complexity of molecular variation within and between patients. In this context, although we analyzed multiple independent cohorts and performed cross-cohort validation, the overall sample size of the datasets analyzed here remains modest relative to the full spectrum of AD heterogeneity described in prior studies. It is therefore likely that additional molecular subtypes or intermediate transcriptional states exist that were not captured in the present analysis. Larger, more diverse cohort studies—ideally integrating longitudinal sampling and multi-omic profiling—will be required to fully resolve the breadth of molecular diversity in AD and to further refine transcriptome-based subtyping frameworks.\u003c/p\u003e\n\u003cp\u003eAnother important consideration is that transcriptomic profiling alone may not fully capture the complex, multi-layered biology underlying atopic dermatitis. AD pathogenesis is shaped not only by host gene expression but also by interactions with additional molecular layers, including the skin microbiome, epigenetic regulation, and metabolic processes [22–25]. In particular, multiple studies have demonstrated that alterations in the cutaneous microbiome—most notably Staphylococcus aureus overgrowth—are tightly linked to disease severity, immune activation, and barrier dysfunction in AD [24,25]. Integrative analyses combining host transcriptomics with metagenomic profiling have shown that microbial dysbiosis can modulate inflammatory gene expression programs and influence therapeutic response, underscoring a bidirectional relationship between host and microbiota. These findings indicate that transcriptome-based subtyping, while informative, represents only one dimension of disease stratification.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAccumulating evidence further indicates that epigenetic regulation constitutes an additional layer of molecular heterogeneity in atopic dermatitis. Genome-wide DNA methylation studies have demonstrated disease- and severity-associated epigenetic alterations in both epidermal and immune-related genes [22,23]. Moreover, dysregulation of non-coding RNAs, including microRNAs implicated in immune activation and keratinocyte differentiation, has been consistently observed in skin diseases [26,27]. These findings suggest that transcriptional heterogeneity in AD is shaped not only by genetic and immunological factors but also by epigenetic mechanisms that integrate environmental and inflammatory cues. Future studies integrating multi-omic layers across larger and longitudinal AD cohorts will be essential to more fully resolve disease mechanisms, refine molecular subtypes, and improve the prediction of treatment response.\u003c/p\u003e\n\u003cp\u003eDespite these limitations, our study provides evidence that AD comprises biologically distinct transcriptional subtypes at the level of lesional skin. By delineating reproducible molecular patterns linked to immune activation, cellular programs, and differential drug responsiveness, this work moves beyond a uniform view of AD as a single inflammatory entity. The identification of subtype-specific transcriptional states suggests that inter-individual variability in treatment outcomes may, at least in part, be rooted in underlying molecular heterogeneity rather than stochastic or purely clinical factors alone. Within this framework, the major lesional subtype is characterized by pronounced immune activation and shows transcriptional features that are concordant with favorable responses to both dupilumab and cyclosporine. This subtype exhibits elevated expression of immune- and inflammation-associated programs, accompanied by increased abundance of activated immune cell populations, indicating an immunologically dominant disease state. Notably, this molecular profile aligns well with the mechanisms of action of both therapies: dupilumab targets key cytokine signaling pathways central to inflammatory cascades in AD, while cyclosporine broadly suppresses T-cell activation and downstream immune responses.\u003c/p\u003e\n\u003cp\u003eIn contrast, the minor lesional subtype displays a molecular profile that is less compatible with the immunosuppressive mechanisms of cyclosporine and is associated with attenuated therapeutic responsiveness. This interpretation is supported by the close resemblance between the transcriptional features of the minor subtype and those observed in cyclosporine non-responders from an independent treatment cohort. A prominent hallmark of the minor subtype is the enrichment of cell cycle– and proliferation-associated pathways, which distinguish it from the immune-dominant major subtype. Similar cell cycle–related transcriptional programs were also observed among genes differentially expressed in cyclosporine non-responders, suggesting a shared molecular state that may be less amenable to therapies primarily targeting immune activation. These findings imply that, in a subset of AD patients, disease activity may be sustained by non-immune–dominant programs, potentially contributing to diminished responsiveness to broad immunosuppressive treatment.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eOur observation that the cyclosporine-resistant minor subtype is characterized by prominent cell-cycle and proliferative signatures is consistent with pathological features commonly associated with chronic, lichenified atopic dermatitis. Previous studies have shown that as AD lesions progress from acute to chronic stages, disease pathology can shift from predominantly immune-cell infiltration toward epidermal hyperplasia (acanthosis) and tissue remodeling [28,29]. In this chronic state, keratinocyte hyperproliferation—potentially sustained by residual cytokine signaling, barrier dysfunction, or mechanical stress such as scratching—may persist even after overt inflammatory infiltrates have diminished, giving rise to a transcriptional profile that is relatively less immune-dominant yet highly proliferative. Notably, the proliferative and keratinization-associated programs enriched in the minor subtype resemble molecular features reported in the so-called “psoriasis-like” phenotype of atopic dermatitis. In particular, Noda \u003cem\u003eet al\u003c/em\u003e. demonstrated that Asian AD patients often exhibit enhanced Th17 polarization and pronounced epidermal hyperplasia compared with European American cohorts, reflecting a hybrid molecular state that shares features of both AD and psoriasis [30]. While the present study does not directly assess ethnic stratification, the enrichment of keratinocyte-intrinsic, cell-cycle–related programs in the minor subtype suggests that this transcriptional state may capture chronic or structurally remodeled lesions in which disease activity is less dependent on acute immune activation. Such a shift toward epidermal-driven pathology may help explain the attenuated responsiveness of this subtype to T-cell–targeted immunosuppressive therapy such as cyclosporine, and highlights the potential need for alternative therapeutic strategies aimed at restoring epidermal differentiation and regulating keratinocyte proliferation.\u003c/p\u003e\n\u003cp\u003eIn summary, this study demonstrates that AD comprises biologically distinct transcriptional subtypes that differ in immune activation, epidermal programs, and responsiveness to systemic therapies. By integrating transcriptomic subtyping with independent validation of drug response patterns, our findings highlight molecular heterogeneity as a key determinant of therapeutic outcome in AD. These results underscore the importance of moving beyond a uniform treatment paradigm and toward molecularly informed stratification in AD. As transcriptomic and multi-omic profiling approaches continue to advance, defining and validating molecular subtypes may provide a foundation for more precise therapeutic selection and improved clinical outcomes.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors gratefully acknowledge the support of \u003cstrong\u003eKorea Polytechnic College\u003c/strong\u003e for providing computational resources and analytical infrastructure essential for this study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research was supported by the GR4A25 project at the Bioconvergence Research Institute of HuGeX Co., Ltd.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eS.H.M. implemented the computational analyses, processed the transcriptomic datasets, and conducted downstream statistical and bioinformatic analyses. J.K., J.H., S.H.H., and B.Y. contributed to study design and interpretation of the clinical and experimental context. S.H., J.H.M., H.H., and K.Y. provided biological interpretation of the transcriptomic results and contributed to the analysis of disease mechanisms and pathway-level findings. \u0026nbsp; K.K. contributed to data visualization and figure preparation. S.-E.H. conceived and supervised the overall project, designed the analytical framework, and performed the integrative transcriptomic analyses. All authors contributed to manuscript writing and revision, and approved the final version of the manuscript.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets analyzed in this study are publicly available. Transcriptomic data for atopic dermatitis skin samples and drug-treated cohorts were obtained from the Gene Expression Omnibus (GEO) under accession numbers GSE157194 and GSE58558.\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eSilverberg, J. I. Public Health Burden and Epidemiology of Atopic Dermatitis. \u003cem\u003eDermatol Clin\u003c/em\u003e 35, 283\u0026ndash;289 (2017).\u003c/li\u003e\n\u003cli\u003eNutten, S. Atopic dermatitis: global epidemiology and risk factors. \u003cem\u003eAnn Nutr Metab\u003c/em\u003e 66 Suppl 1, 8\u0026ndash;16 (2015).\u003c/li\u003e\n\u003cli\u003eGuttman-Yassky, E., Nograles, K. E. \u0026amp; Krueger, J. G. Contrasting pathogenesis of atopic dermatitis and psoriasis--part I: clinical and pathologic concepts. \u003cem\u003eJ Allergy Clin Immunol\u003c/em\u003e 127, 1110\u0026ndash;8 (2011).\u003c/li\u003e\n\u003cli\u003eGittler, J. K. \u003cem\u003eet al.\u003c/em\u003e Progressive activation of T(H)2/T(H)22 cytokines and selective epidermal proteins characterizes acute and chronic atopic dermatitis. \u003cem\u003eJ Allergy Clin Immunol\u003c/em\u003e 130, 1344\u0026ndash;54 (2012).\u003c/li\u003e\n\u003cli\u003eLeung, D. Y. M., Boguniewicz, M., Howell, M. D., Nomura, I. \u0026amp; Hamid, Q. A. New insights into atopic dermatitis. \u003cem\u003eJ Clin Invest\u003c/em\u003e 113, 651\u0026ndash;7 (2004).\u003c/li\u003e\n\u003cli\u003eHe, H. \u003cem\u003eet al.\u003c/em\u003e Single-cell transcriptome analysis of human skin identifies novel fibroblast subpopulation and enrichment of immune subsets in atopic dermatitis. \u003cem\u003eJ Allergy Clin Immunol\u003c/em\u003e 145, 1615\u0026ndash;1628 (2020).\u003c/li\u003e\n\u003cli\u003eWongvibulsin, S. \u003cem\u003eet al.\u003c/em\u003e Transcriptomic analysis of atopic dermatitis in African Americans is characterized by Th2/Th17-centered cutaneous immune activation. \u003cem\u003eSci Rep\u003c/em\u003e 11, 11175 (2021).\u003c/li\u003e\n\u003cli\u003eM\u0026ouml;bus, L. \u003cem\u003eet al.\u003c/em\u003e Atopic dermatitis displays stable and dynamic skin transcriptome signatures. \u003cem\u003eJ Allergy Clin Immunol\u003c/em\u003e 147, 213\u0026ndash;223 (2021).\u003c/li\u003e\n\u003cli\u003eBlauvelt, A. \u003cem\u003eet al.\u003c/em\u003e Long-term management of moderate-to-severe atopic dermatitis with dupilumab and concomitant topical corticosteroids (LIBERTY AD CHRONOS): a 1-year, randomised, double-blinded, placebo-controlled, phase 3 trial. \u003cem\u003eLancet\u003c/em\u003e 389, 2287\u0026ndash;2303 (2017).\u003c/li\u003e\n\u003cli\u003eHalling, A.-S., Loft, N., Silverberg, J. I., Guttman-Yassky, E. \u0026amp; Thyssen, J. P. Real-world evidence of dupilumab efficacy and risk of adverse events: A systematic review and meta-analysis. \u003cem\u003eJ Am Acad Dermatol\u003c/em\u003e 84, 139\u0026ndash;147 (2021).\u003c/li\u003e\n\u003cli\u003evan der Schaft, J. \u003cem\u003eet al.\u003c/em\u003e Drug survival for ciclosporin A in a long-term daily practice cohort of adult patients with atopic dermatitis. \u003cem\u003eBr J Dermatol\u003c/em\u003e 172, 1621\u0026ndash;1627 (2015).\u003c/li\u003e\n\u003cli\u003eGarritsen, F. M. \u003cem\u003eet al.\u003c/em\u003e Ten years experience with oral immunosuppressive treatment in adult patients with atopic dermatitis in two academic centres. \u003cem\u003eJ Eur Acad Dermatol Venereol\u003c/em\u003e 29, 1905\u0026ndash;12 (2015).\u003c/li\u003e\n\u003cli\u003eNielsen, T. O. \u003cem\u003eet al.\u003c/em\u003e A comparison of PAM50 intrinsic subtyping with immunohistochemistry and clinical prognostic factors in tamoxifen-treated estrogen receptor-positive breast cancer. \u003cem\u003eClin Cancer Res\u003c/em\u003e 16, 5222\u0026ndash;32 (2010).\u003c/li\u003e\n\u003cli\u003eParker, J. S. \u003cem\u003eet al.\u003c/em\u003e Supervised risk predictor of breast cancer based on intrinsic subtypes. \u003cem\u003eJ Clin Oncol\u003c/em\u003e 27, 1160\u0026ndash;7 (2009).\u003c/li\u003e\n\u003cli\u003eTsoi, L. C. \u003cem\u003eet al.\u003c/em\u003e Atopic Dermatitis Is an IL-13-Dominant Disease with Greater Molecular Heterogeneity Compared to Psoriasis. \u003cem\u003eJ Invest Dermatol\u003c/em\u003e 139, 1480\u0026ndash;1489 (2019).\u003c/li\u003e\n\u003cli\u003eGuttman-Yassky, E., Nograles, K. E. \u0026amp; Krueger, J. G. Contrasting pathogenesis of atopic dermatitis and psoriasis--part II: immune cell subsets and therapeutic concepts. \u003cem\u003eJ Allergy Clin Immunol\u003c/em\u003e 127, 1420\u0026ndash;32 (2011).\u003c/li\u003e\n\u003cli\u003eNoda, S., Krueger, J. G. \u0026amp; Guttman-Yassky, E. The translational revolution and use of biologics in patients with inflammatory skin diseases. \u003cem\u003eJ Allergy Clin Immunol\u003c/em\u003e 135, 324\u0026ndash;36 (2015).\u003c/li\u003e\n\u003cli\u003ede Bruin-Weller, M. S. \u003cem\u003eet al.\u003c/em\u003e Biologics to Treat Atopic Dermatitis: Effectiveness, Safety, and Future Directions. \u003cem\u003eAllergy\u003c/em\u003e https://doi.org/10.1111/all.70061 (2025) doi:10.1111/all.70061.\u003c/li\u003e\n\u003cli\u003eKhattri, S. \u003cem\u003eet al.\u003c/em\u003e Cyclosporine in patients with atopic dermatitis modulates activated inflammatory pathways and reverses epidermal pathology. \u003cem\u003eJ Allergy Clin Immunol\u003c/em\u003e 133, 1626\u0026ndash;34 (2014).\u003c/li\u003e\n\u003cli\u003eM\u0026ouml;bus, L. \u003cem\u003eet al.\u003c/em\u003e Atopic dermatitis displays stable and dynamic skin transcriptome signatures. \u003cem\u003eJ Allergy Clin Immunol\u003c/em\u003e 147, 213\u0026ndash;223 (2021).\u003c/li\u003e\n\u003cli\u003eBrunner, P. M. Early immunologic changes during the onset of atopic dermatitis. \u003cem\u003eAnn Allergy Asthma Immunol\u003c/em\u003e 123, 152\u0026ndash;157 (2019).\u003c/li\u003e\n\u003cli\u003eOlisova, O. Y. \u003cem\u003eet al.\u003c/em\u003e Skin DNA methylation profile in atopic dermatitis patients: A case-control study. \u003cem\u003eExp Dermatol\u003c/em\u003e 29, 184\u0026ndash;189 (2020).\u003c/li\u003e\n\u003cli\u003eRodr\u0026iacute;guez, E. \u003cem\u003eet al.\u003c/em\u003e An integrated epigenetic and transcriptomic analysis reveals distinct tissue-specific patterns of DNA methylation associated with atopic dermatitis. \u003cem\u003eJ Invest Dermatol\u003c/em\u003e 134, 1873\u0026ndash;1883 (2014).\u003c/li\u003e\n\u003cli\u003eByrd, A. L. \u003cem\u003eet al.\u003c/em\u003e Staphylococcus aureus and Staphylococcus epidermidis strain diversity underlying pediatric atopic dermatitis. \u003cem\u003eSci Transl Med\u003c/em\u003e 9, (2017).\u003c/li\u003e\n\u003cli\u003eKong, H. H. \u003cem\u003eet al.\u003c/em\u003e Temporal shifts in the skin microbiome associated with disease flares and treatment in children with atopic dermatitis. \u003cem\u003eGenome Res\u003c/em\u003e 22, 850\u0026ndash;9 (2012).\u003c/li\u003e\n\u003cli\u003eSonkoly, E. \u003cem\u003eet al.\u003c/em\u003e MicroRNAs: novel regulators involved in the pathogenesis of psoriasis? \u003cem\u003ePLoS One\u003c/em\u003e 2, e610 (2007).\u003c/li\u003e\n\u003cli\u003eSonkoly, E. \u003cem\u003eet al.\u003c/em\u003e MiR-155 is overexpressed in patients with atopic dermatitis and modulates T-cell proliferative responses by targeting cytotoxic T lymphocyte-associated antigen 4. \u003cem\u003eJ Allergy Clin Immunol\u003c/em\u003e 126, 581-9.e1-20 (2010).\u003c/li\u003e\n\u003cli\u003eTsoi, L. C. \u003cem\u003eet al.\u003c/em\u003e Progression of acute-to-chronic atopic dermatitis is associated with quantitative rather than qualitative changes in cytokine responses. \u003cem\u003eJ Allergy Clin Immunol\u003c/em\u003e 145, 1406\u0026ndash;1415 (2020).\u003c/li\u003e\n\u003cli\u003eGittler, J. K. \u003cem\u003eet al.\u003c/em\u003e Progressive activation of T(H)2/T(H)22 cytokines and selective epidermal proteins characterizes acute and chronic atopic dermatitis. \u003cem\u003eJ Allergy Clin Immunol\u003c/em\u003e 130, 1344\u0026ndash;54 (2012).\u003c/li\u003e\n\u003cli\u003eNoda, S. \u003cem\u003eet al.\u003c/em\u003e The Asian atopic dermatitis phenotype combines features of atopic dermatitis and psoriasis with increased TH17 polarization. \u003cem\u003eJ Allergy Clin Immunol\u003c/em\u003e 136, 1254\u0026ndash;64 (2015).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-8477069/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8477069/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"Atopic dermatitis (AD) is a chronic inflammatory skin disease characterized by marked clinical heterogeneity and variable responses to systemic therapies. Although transcriptomic studies have revealed substantial molecular diversity within AD lesions, how this heterogeneity relates to therapeutic responsiveness remains incompletely understood. Here, we performed integrative transcriptomic analyses of lesional and non-lesional skin from patients with AD to define molecular subtypes and examine their biological and therapeutic relevance. Unsupervised clustering of lesional skin transcriptomes identified two distinct lesional subtypes with divergent molecular features. These subtypes differed markedly in immune activation, cell-cycle–associated programs, and immune cell composition. We next assessed subtype-specific drug responsiveness by comparing lesional gene expression signatures with transcriptomic changes induced by dupilumab and cyclosporine treatment. The major lesional subtype exhibited strong inverse transcriptional concordance with drug-induced expression changes, whereas the minor lesional subtype showed attenuated responses, particularly to cyclosporine. Analysis of an independent cyclosporine-treated cohort further demonstrated that clinical non-responders displayed transcriptomic features resembling the minor lesional subtype, including enrichment of cell-cycle–associated programs. Together, these findings demonstrate that AD lesions comprise biologically distinct transcriptomic subtypes with differential immune composition and systemic drug responsiveness, providing a molecular framework for understanding heterogeneity in AD and supporting transcriptome-based stratification for precision treatment strategies.","manuscriptTitle":"Transcriptomic Subtyping of Atopic Dermatitis Reveals Distinct Drug Response Signatures","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-07 08:53:56","doi":"10.21203/rs.3.rs-8477069/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"0a8b56d2-8cdf-4c11-ba71-83f43885a982","owner":[],"postedDate":"January 7th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":60423006,"name":"Health sciences/Biomarkers"},{"id":60423007,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":60423008,"name":"Health sciences/Diseases"},{"id":60423009,"name":"Biological sciences/Immunology"}],"tags":[],"updatedAt":"2026-01-07T08:53:56+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-07 08:53:56","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8477069","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8477069","identity":"rs-8477069","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}