An unbiased tissue transcriptome analysis identifies potential markers for skin phenotypes and therapeutic responses in atopic dermatitis

Abstract

Atopic dermatitis (AD) is a skin disease exhibiting clinical and molecular heterogeneity, thereby jeopardizing the development of personalized treatments. Here we pursue a cross-sectional and longitudinal cohort analysis of 951 whole-skin samples, employing an unsupervised decomposition analysis to link gene expression profiles to disease severity, six distinct skin phenotypes, and blood cytokines representative of given endotypes. Specifically, type 2 and type 17 responses are associated with major skin phenotypes such as erythema and induration, while type 1 response is upregulated in lichen amyloidosis of AD patients. Longitudinal analysis of patients treated with dupilumab finds sustained gene signatures related to type 17 response in lesional skin and upregulated transcription factors in non-lesional skin of patients with poor treatment outcomes. Lastly, several extracellular matrix organization-associated genes are correlated with clinical severity and treatment response to dupilumab. Our findings thus provide potential skin and blood biomarkers for assessing endotypes and therapeutic responses in AD to pave the way for personalized medicine.

Introduction

The skin is a complex organ that covers the entire body, made up of various cells such as keratinocytes, fibroblasts, immune cells, nerve cells, and skin appendage cells. Maintaining healthy skin requires a delicate balance among the skin barrier, immune responses, and sensory perception. Disruptions in these functions can lead to various skin disorders. One common and troublesome condition is atopic dermatitis (AD). While not typically life-threatening, AD causes persistent itching and chronic inflammation, significantly impacting patients’ quality of life. AD involves complex interactions between structural and inflammatory cells in the skin, leading to barrier dysfunction and diverse inflammatory responses^1,2. The symptoms of AD vary widely among patients, with different areas of the skin affected in different ways³. This variability makes it challenging to understand AD’s pathogenesis and develop precise treatments. Therefore, it is crucial to study the relationship between the diverse symptoms of AD and its underlying molecular features.

Past research on other inflammatory diseases has used large-scale human data to uncover complex pathogeneses^4,5. However, similar efforts in AD using blood biomarkers have shown only modest results^2,6. Though potentially more informative, skin tissue studies are limited by the invasiveness and scarring associated with whole-skin sampling. Less invasive methods, such as blistering and tape stripping, have been used to study the outer layer of the skin; however, these methods may fail to capture important information derived from deeper skin layers including signals from dermal fibroblasts and macrophages^1,7,8,9.

AD also has a “waxing and waning” nature, with unpredictable flare-ups even during effective treatments^10,11. In personalized medicine, understanding the skin’s immune state, especially in lesions that do not respond to treatment, is essential. Long-term sequential sampling with detailed clinical data is needed to capture the changes in skin transcriptomic profiles during these flare-ups. Among new treatments, dupilumab, a monoclonal IL-4 receptor antagonist, is the mainstay treatment for severe AD owing to its significant efficacy and safety^12,13,14. Understanding why some patients do not respond well to dupilumab is crucial for improving clinical practices.

This study investigates the diverse molecular pathways of AD through a large cross-sectional and longitudinal cohort study using unsupervised decomposition analysis. We analyze mRNA from full-thickness skin tissue, including both the epidermis and dermis, obtained using minimally invasive 1-mm punch biopsies. This method allows us to collect numerous samples from both affected and unaffected skin areas, as well as tracking changes in patients treated with dupilumab. By applying non-negative matrix factorization (NMF), we identify previously recognized and unrecognized pathological pathways in AD-affected skin linked to clinical symptoms and the impact of dupilumab treatment. These findings offer insights into personalized medicine for AD by improving therapeutic response assessments and disease monitoring.

Results

Keratinocytes and fibroblasts as primary elements of 1-mm punch skin biopsies for RNA sequencing

To understand the diverse molecular profiles of AD, we conducted a large-scale cross-sectional and longitudinal cohort study using RNA-seq data from 1-mm punch biopsy skin samples. An overview of the study design is illustrated in Fig. 1. These samples included those from healthy individuals and patients with psoriasis as disease controls. Out of 1061 skin samples, 1036 met the quality criteria for further analysis. This dataset comprised 529 lesional and 422 non-lesional samples from 156 patients with AD, 18 lesional and 18 non-lesional samples from 18 patients with psoriasis, and 49 samples from 26 healthy individuals (Table 1, Supplementary Fig. 1A, B, Supplementary Tables 1–3). Treatments included topical steroids, dupilumab, and other agents. Cross-sectional analyses were limited to samples from patients not treated with dupilumab, while longitudinal analyses were conducted using serial samples from 24 patients treated with dupilumab.

Fig. 1: Overview of the study design.

The schematic outlines the process from patient skin sampling to transcriptome analysis and computational identification of disease- and response-related signatures. GC gene cluster, ECM extracellular matrix, IEG immediate early genes.

Table 1 Patient demographics

Before conducting the analyses, we validated the cellular composition of the 1-mm skin samples. The levels of COL1A1 and KRT10 were substantially higher than those of CD3E and CD207 (Fig. 2A), implying that the transcripts from these biopsies predominantly originated from keratinocytes and fibroblasts. In the skin, keratinocytes and fibroblasts function as immune sentinels, producing various immune mediators in response to distinct inflammatory signals^1,2,15. Thus, we inferred that our dataset indirectly reflected the immune status of the samples through the gene expression features of keratinocytes and fibroblasts.

Fig. 2: Metagenes obtained using non-negative matrix factorization (NMF) and comparison of metagene expression across diseases.

A Density plot depicting representative genes associated with cellular components in the skin. The 1-mm punch whole-skin sample used for mRNA-seq included abundant keratinocyte- and fibroblast-derived transcriptomes. 1036 samples were used for analysis. B Heatmap illustrating 2368 highly-weighted genes in metagenes obtained from the decomposition analysis using NMF. Each value represents the gene weights, and the representative genes of each metagene are indicated in the right column. The bubble plot below the heatmap depicts the scaled gene weights (z-score_g) of the genes that are representative of cellular components. 1036 samples were used for analysis. C The bottom heatmap represents the statistical significance of the comparison between normal samples and lesional AD, non-lesional AD, lesional psoriasis, and non-lesional psoriasis samples, tested by Wilcoxon test. The p values were corrected using the Bonferroni method for multiple comparisons. Blue indicates lower expression, and red indicates significantly higher expression than that in normal samples. Statistically insignificant results are highlighted in gray. All tests were two-sided. n (samples) = 49 (normal), 246 (lesional AD), 170 (non-lesional AD), 18 (lesional psoriasis), and 18 (non-lesional psoriasis). Source data are provided as a Source Data file. AD atopic dermatitis, Pso psoriasis.

Decomposition using NMF depicts disease-related metagenes and sample batch effects

A differentially expressed gene analysis comparing AD and psoriasis skin with healthy skin revealed genes associated with epidermal barrier and immunological dysfunction, consistent with previous studies (Supplementary Fig. 2A, Supplementary Data 1)^{2,8,16,17,18,19}. Unsupervised hierarchical clustering identified sample biases in gene expression due to the inclusion of skin appendages, such as sebaceous and sweat glands (Supplementary Fig. 2B).

Therefore, to exclude sample biases and unravel the molecular heterogeneity in AD skin, we used NMF, an unsupervised decomposition method. NMF decomposes high-dimensional gene expression data into a smaller number of gene sets in an unbiased manner²⁰. This method uncovers hidden co-expression profiles and reveals novel genes that correlate with recognized genes, providing insights into previously unrecognized gene interactions underlying the biology of complex diseases. Applying NMF to our data, we identified 29 gene sets, which we henceforth refer to as “SKIn-Tissue derived Metagenes” (SKITm), hereafter referred to as SKITm1 through SKITm29 (Fig. 2B and Supplementary Fig. 2C).

Genes with high weights in each metagene, defined by a z-score >3 (referred to as z-score_g), were considered key representative gene signatures for their respective metagenes (Fig. 2B, Supplementary Fig. 3, Supplementary Data 2 and 3). Genes highly weighted specifically in only one metagene were extracted as “marker” genes, serving as candidate biomarkers of each metagene (see “Methods” for details). To understand the biological implications of the metagenes, we verified the highly weighted genes using knowledge-based analytical approaches in reference to previous studies (descriptions of the top 10 highly-weighted genes and marker genes for all metagenes are presented in Supplementary Data 4. An example of enrichment analysis for SKITm10 is shown in Supplementary Fig. 4, Supplementary Tables 4 and 5)^{17,19,21,22,23,24,25,26,27,28}. Additionally, the expression levels of key cytokines representing type 2 and type 17 inflammation (IL13, CCL13, CCL18, CCL26, IL17A, IL17C, IL17F, IL23A, CCL20) were quantified by quantitative real-time PCR in 33 AD lesional samples and their correlation with the expression levels of key metagenes was analyzed. As a result, IL13, CCL13, and CCL26 expression levels showed significant correlations with SKITm17, IL17A, IL23A, and CCL20 were significantly correlated with SKITm10, and IL17C was significantly correlated with SKITm15, supporting the biological relevance of the metagenes (Fig. 3).

Fig. 3: Correlation between qRT-PCR–quantified cytokines expression levels and RNA-seq-derived metagenes.

The expression levels of cytokines, quantified by qRT-PCR, were correlated with metagene expression levels derived from RNA-seq data, both obtained from the same set of skin samples (n = 33). The y-axis represents the cytokine levels quantified by qRT-PCR, and the x-axis represents the metagene expression levels derived from the decomposition analysis of RNA-seq data. The r and p indicate the correlation coefficient and the significance obtained through Spearman’s correlation analysis, respectively. The linear regression line is shown with a blue shaded area indicating the 95% confidence interval of the fitted line. Source data are provided as a Source Data file. *p < 0.01.

We then investigated metagenes associated with sample biases. In addition to the seven metagenes derived from skin appendages and reticulocytes (SKITm22–27, SKITm19), we identified four metagenes associated with sample bias, including outlier sample (SKITm14, SKITm20), sample quality such as RIN score and sequence batch effects (SKITm28, SKITm29) (Fig. 2B, Supplementary Figs. 5 and 6A–D). Uniform manifold approximation and projection (UMAP) revealed a distribution that distinctly reflected the disease state when we excluded the expression of sample bias-related metagenes (Supplementary Fig. 7).

Among the 18 metagenes unrelated to sample bias, 12 were differentially expressed in lesional AD or psoriasis samples in the cross-sectional study (Fig. 2C, Supplementary Fig. 8. Patient demographics and sample information are presented in Supplementary Tables 6–8.). Descriptions of highly-weighted genes, marker genes and biological interpretation for the 12 metagenes are presented in Table 2. SKITm1, SKITm17, SKITm4 and SKITm8, with high contributions of genes associated with leukocyte adhesion, type 2 response, and keratinocyte differentiation respectively, were upregulated in AD samples. SKITm10, characterized by inflammatory keratins and keratinocyte-derived genes inducible by type 17 inflammation, and SKITm15, also enriched with genes associated with type 17 response in keratinocytes with marked weight of SPRR2G, were upregulated in AD and psoriasis lesional samples. SKITm9, also characterized by type 17 response, was upregulated in psoriasis only. SKITm11, associated with type 1 response, was also highly expressed in psoriasis. SKITm2, representing an epidermal barrier gene signature, and SKITm16, enriched for genes involved in dermal extracellular matrix (ECM) organization, were both downregulated in AD and psoriasis. SKITm5 and SKITm6, composed of transcription factor genes such as immediate early genes (IEG), were upregulated in AD. Notably, AD and psoriasis exhibited apparent differences in the non-lesional sites. Non-lesional psoriasis samples presented almost the same profiles as normal controls, whereas non-lesional AD samples showed similar but weaker features than lesional AD, with enhanced immunological and barrier disruption. Most of these 12 metagenes correlated with the gene modules from our previous study which included 219 common AD samples (Supplementary Fig. 9)²⁹.

Table 2 Summary of metagenes differentially expressed in AD or psoriasis

The total expression score of 12 metagenes, representing the overall molecular dysregulation, was significantly higher in lesional and non-lesional AD than in normal samples (Fig. 4A). The metagene expression score correlated with the local and total disease severity (Fig. 4B, C). The expression profiles of representative normal and AD participants with mild, moderate, and severe disease are shown in Supplementary Fig. 10A–H.

Fig. 4: Association between metagene expression and clinical symptoms or blood cytokine levels based on the cross-sectional study.

A The y-axis represents the metagene expression score, which was calculated by the sum of expression values of 12 disease-associated metagenes, namely, SKITm17, SKITm10, SKITm9, SKITm15, SKITm11, SKITm8, SKITm6, SKITm5, SKITm4, SKITm1, SKITm2 and SKITm16. All the expression values of each sample were divided by the maximum expression value in each metagene for normalization (relative expression values). The expression values for SKITm2 and SKITm16 were then subtracted from 1 before the calculation of the metagene expression score. The relative expression value was used for analysis throughout the cross-sectional and longitudinal analysis. The metagene expression score between normal samples and lesional AD and non-lesional AD were tested by Wilcoxon test. The p values were corrected using the Bonferroni method for multiple comparisons. In box plots, the center line represents the median; the lower and upper bounds of the box represent the 25th and 75th percentiles, respectively; whiskers extend to the smallest and largest values within 1.5× the interquartile range from the lower and upper quartiles. N (samples) = 49 (normal), 246 (lesional AD), and 170 (non-lesional AD). *p < 0.01. B The y-axis represents the local severity score, calculated by the sum of clinical severity scores of erythema, papulation/induration, excoriation, and lichenification of the sampled sites. n (samples) = 49 (normal), 235 (lesional AD), and 159 (non-lesional AD). *p < 0.01. C The y-axis represents the total disease severity (EASI). n (samples) = 49 (normal), 234 (lesional AD), and 156 (non-lesional AD). *p < 0.01. For (B, C), the severity scores of normal samples were set to 0. The r and p indicate the correlation coefficient and the significance obtained through Spearman’s correlation analysis, respectively. The linear regression line is shown with a gray shaded area indicating the 95% confidence interval of the fitted line. D Heatmap depicting p values obtained from a regression analysis of the severity of skin phenotypes (erythema, papulation/induration, excoriation, lichenification), the overall severity (EASI), and pruritus score. Results from PN and LA lesions were compared with those of non-lesional samples from the same patients using a paired t-test. Red and blue indicate a significant positive and negative association with metagene expression and the severity of symptoms, respectively. Statistical significance was assessed using a two-sided test. Sample sizes: erythema/papulation/excoriation/lichenification (n = 212), LA (n = 12), PN (n = 16), pruritus (lesional, n = 230; non-lesional, n = 151), EASI (lesional, n = 235; non-lesional, n = 159). E Heatmap displaying p values obtained from regression analysis of plasma/serum cytokine levels. Statistical significance of regression coefficients was assessed using a two-sided test. EDN eosinophil-derived neurotoxin, TSLP thymic stromal lymphopoietin. Sample sizes: lesional, n = 224 samples (206 for CCL17); non-lesional, n = 148 (140 for CCL17). For (D, E), no adjustments for multiple comparisons were made, as the analyses were exploratory in nature. Source data are provided as a Source Data file.

Collectively, our unsupervised decomposition analysis revealed a collection of gene clusters, including both established and unrecognized gene clusters within the context of previous inflammatory skin studies.

Different gene co-expression patterns contribute to various skin phenotypes and disease severity

In the cross-sectional study, we explored the correlation between metagenes and local symptoms (erythema, papulation/induration, excoriation, lichenification), as well as disease severity indicators, including Eczema Area and Severity Index (EASI; a measure of disease severity), and pruritus (patient-oriented pruritus severity score), using a mixed-effects regression model (Eq. (1)). We also evaluated metagene expression in prurigo nodules (PN) and lichen amyloidosis (LA), rare chronic subtypes of AD skin symptoms³⁰, and compared them with non-lesional samples from the same patients (PN: eight patients; LA: six patients) (Supplementary Fig. 1B). Each metagene was associated with distinct skin phenotypes (Fig. 4D and Supplementary Fig. 11A). Specifically:

• Erythema: Negatively associated with SKITm2 and SKITm16 and positively associated with SKITm17, SKITm10, and SKITm5.

• Papulation/Induration: Negatively associated with SKITm2 and SKITm16, and positively associated with SKITm17, SKITm10, and SKITm9.

• Excoriation: Displayed marked SKITm10, SKITm8, and SKITm1 expression. Particularly, high SKITm1 expression suggested leukocyte induction through scratching.

• Lichenification: Uniquely associated with SKITm8.

• PN: Associated with SKITm2, SKITm17, SKITm10, and SKITm8.

• LA: Exhibited significant expression of SKITm11 and SKITm8 metagenes. Notably, SKITm11 expression was distinct in LA.

Disease severity (EASI) and pruritus score were negatively associated with SKITm2 and SKITm16, and positively associated with SKITm17 and SKITm10. Even non-lesional sites displayed a weak but significant association with disease severity and pruritus; EASI was associated with SKITm10, SKITm8, SKITm5 and SKITm4, whereas pruritus was associated with SKITm17 and SKITm15 (Fig. 4D and Supplementary Fig. 11B).

These results suggested that the 1-mm skin sample serves as a reliable indicator of both local skin phenotypes and overall severity in AD patients. Our findings revealed distinct metagenes associated with various local skin symptoms, highlighting their differential significance in lesions and non-lesions.

Blood cytokine levels are associated with multiple skin metagenes

We observed that most blood cytokines were associated with the expression of multiple metagenes in the skin (Fig. 4E) using a mixed-effects regression model (Eq. (2)). Notably, CCL27, IL-22, and TNF-α levels in the blood were positively associated with SKITm17, SKITm10, SKITm2, and SKITm16 in lesional sites, with CCL27 exhibiting the highest level of significance. CCL17 (thymus and activation-regulated chemokine, TARC) showed significant yet weaker associations with SKITm17, SKITm10, and SKITm2. At lesional sites, eosinophil-derived neurotoxin (EDN) was associated with SKITm17, a type 2 response metagene, but not with other inflammation-related metagenes. Conversely, blood levels of IL-17 and CCL20, cytokines associated with type 17 inflammation³¹, positively correlated with SKITm10, validating the metagene as a “type 17 response-related metagene.” Notably, CCL20 exhibited a stronger and more significant association with SKITm10 than with the IL-17 level. Furthermore, the expression levels of non-lesional SKITm15, SKITm5, and SKITm4 were associated with more blood cytokines than those of lesional samples. Among blood clinical biomarkers, LDH demonstrated the highest association with the expression of most skin metagenes and disease severity, whereas eosinophil counts were specifically related to SKITm17, a type 2 inflammation response metagene (Supplementary Fig. 12) (Eq. (3)).

These results suggest that while most blood cytokines are related to multiple metagenes and overall severity, EDN, IL-17, and CCL20 could potentially be utilized to assess immunophenotypes in AD lesions.

Correlation between dupilumab treatment outcomes and baseline metagene expression

In the longitudinal analysis, we examined skin samples from 24 patients sampled monthly during dupilumab treatment, with patient demographics provided in Supplementary Table 9. Because dupilumab can induce facial erythema with a distinct etiology from pre-existing AD, we evaluated the treatment efficacy by modified EASI (mEASI), which is calculated by excluding the scores of the head and neck from the total EASI score^{32,33,34,35,36} (clinical photo of a patient who developed dupilumab-associated facial erythema is presented in Supplementary Fig. 13). Patients were classified into three groups based on mEASI during treatment: “early” responders, who achieved a mild disease level (defined by mEASI ≤ 6.3) within the first month and maintained that level for 6 months; “poor” responders, who did not achieve the mild level during the entire 6-month study period; and “intermediate” responders, who did not meet the criteria for either “early” or “poor” responders (Fig. 5A). These response trends persisted for over a year for most patients (Supplementary Fig. 14).

Fig. 5: Metagene expression and blood cytokine features of patients with different dupilumab responses, based on the longitudinal study.

A Time-course of disease severity (modified Eczema Area and Severity Index: mEASI) in patients treated with dupilumab (n = 24 patients). The blue line represents 6.3, which is the cutoff for mild disease severity in the mEASI. B Correlation between the disease severity score (mEASI) and treatment response groups before treatment. The x-axis represents the responder group: +++ “early responders,” ++ “intermediate responders,” and + “poor responders.” n = 5 (early responders), 11 (intermediate responders), 8 (poor responders). Correlation between treatment outcomes and the expression of skin metagenes (C), skin marker genes (D), and blood cytokines (E) before treatment. n (samples) = 19 (lesional AD skin), 15 (non-lesional AD skin), 19 (blood). In (C, D), the red and blue dots represent lesional and non-lesional samples, respectively. In (B–D), in box plots, the center line represents the median; the lower and upper bounds of the box represent the 25th and 75th percentiles, respectively; whiskers extend to the smallest and largest values within 1.5× the interquartile range from the lower and upper quartiles. The p values indicate the significance obtained using Spearman’s correlation analysis. *p < 0.01, †p < 0.05. Source data are provided as a Source Data file.

We then investigated the clinical and molecular features of the response groups before treatment. Clinically, the pretreatment mEASI significantly differed among the three groups, with lower disease severity being associated with a better response to dupilumab (Fig. 5B). Skin transcriptomic analysis revealed a significant correlation between lesional SKITm16 expression and treatment responses (p = 6.3e-03). Additionally, marginal associations were observed between lesional SKITm10 expression and non-lesional SKITm5 and SKITm4 expression (p = 2.5e-02, 3.6e-02, 1.8e-02, respectively, Fig. 5C).

To identify potentially useful standalone biomarkers, we further assessed the pretreatment expression levels of SKITm16, SKITm10, and SKITm5 marker genes. SKITm4, lacking marker genes, was excluded from this analysis. We found significant associations with treatment outcomes in lesional samples for three SKITm16 genes (OGN, RECK, PDGFRL) and one SKITm10 gene (FBLIM1). In non-lesional samples, significant associations were found for SKITm5 genes (TOB2, EGR1, MCL1, CBX4, IER2, PLK2, CSRNP1, EGR3, PMAIP1) (p < 0.01, Fig. 5D). Supplementary Fig. 15 displays marker genes with marginal correlations (p < 0.05). For these analyses, five patients who were receiving cyclosporine at baseline were excluded, and four patients with no non-lesional sites on their trunks were excluded from the non-lesional skin analysis.

Pretreatment levels of blood cytokines were also analyzed, which showed significant associations with treatment outcomes for IL-22 (p = 2.4e-03) and IL-18 (p = 1.1e-03), while CCL20 exhibited marginal significance (p = 5.0e-02) (Fig. 5E and Supplementary Fig. 16). Although IL-17 did not exhibit a significant correlation, IL-17 blood levels in most early responders were nearly zero. EDN, associated with skin SKITm17 expression (type 2 immune response) in the cross-sectional analysis, was not significantly associated with treatment responses. To further evaluate biomarkers for treatment effect, we validated the correlation between pretreatment levels of skin and blood biomarkers and disease severity after treatment. We found that pretreatment levels of lesional SKITm10, non-lesional SKITm15, plasma IL-22, and IL-18 correlated with the average severity score after 5 and 6 months of treatment (Fig. 6A–C).

Fig. 6: Correlation between pretreatment skin gene expression, blood cytokine levels, and disease severity after dupilumab treatment.

Analysis of pretreatment levels of (A) skin metagene expression, (B) marker gene expression of SKITm10, and (C) blood cytokines. The x-axis represents the averaged values of AD disease severity score (mEASI) after 5 and 6 months of dupilumab treatment, and the r and p indicate the correlation coefficient and the significance obtained through Spearman’s correlation analysis, respectively. n (samples) = 19 (lesional AD skin), 15 (non-lesional AD skin), 19 (blood). *p < 0.01, †p < 0.05. Source data are provided as a Source Data file.

Dupilumab suppresses type 2 inflammation and restores the epidermal barrier and ECM-related genes

We evaluated the chronological effect of dupilumab on skin mRNA profiles (Fig. 7A) using a mixed-effects regression model (Eq. (4)). Average metagene expression levels in the 24 patients during treatment were compared with respective pretreatment levels. The most notable changes were observed in SKITm17 (associated with a type 2 immune response) and SKITm16 (ECM-related gene signature) in both lesional and non-lesional samples, with expression levels almost shifting to the normal range after 3 months. Lesional SKITm2 expression also recovered slowly but did not reach the normal range. SKITm10, SKITm6, and SKITm5 in the lesional samples and SKITm4 in the non-lesional samples showed a partial response. SKITm11 expression was low among the sampled patients, showing no change in the first 6 months of dupilumab treatment. Blood CCL27, CCL17, IL-22, IL-16, and IL-18 levels consistently decreased (Supplementary Fig. 17) (Eq. (5)). These results confirmed the strong effect of dupilumab in suppressing type 2 inflammation and restoring the epidermal barrier. Additionally, the findings suggested that dupilumab also helps restore the pathological gene expression in fibroblasts, while the effect on type 17 response in the skin is limited.

Fig. 7: Transition in metagene expression during dupilumab treatment based on the longitudinal study.

A Average expression of skin metagenes during dupilumab treatment in 24 patients. The p values indicate the significance of the differences in metagene expression compared with the pretreatment level (month 0) as determined by a linear mixed-effects regression analysis. n = 24 (AD), 26 (normal). *p < 0.01. P values were corrected using the Bonferroni method for multiple comparisons. Source data are provided as a Source Data file. B Averaged transition in skin metagene expression among different responder groups. In (A, B), the mean and SEM (standard error of the mean) are presented as red and blue points and bars, respectively, for lesional and non-lesional AD skin samples. Green highlighted shadows represent the 95% confidence interval (CI) for the mean expression in healthy individuals. n = 5 (early responders), 11 (intermediate responders), 8 (poor responders).

Long-course sampling demonstrates skewed inflammation in patients with poor responses

Our subsequent interest was to investigate skin immunophenotypes under dupilumab treatment. Among the three responder groups, the average transition in skin metagene expression exhibited notable patterns (Fig. 7B). Specifically, SKITm17, associated with a type 2 immune response, was strongly suppressed across all three responder groups. Conversely, SKITm10, associated with a type 17 response, was consistently elevated in the poor-response group. Furthermore, SKITm15 and SKITm5 expression were higher in non-lesional sites of poor responders during treatment (Fig. 7B and Supplementary Fig. 18). The average transition in blood cytokine/chemokine levels was not notably different among the responder groups (Supplementary Fig. 19).

The residual expression of lesional SKITm10 in poor responders persisted over the long term. In the longitudinal analysis, eight patients were sampled sequentially for over 1 year; two patients were classified into the early, two into the intermediate, and four into the poor-response groups in the previous analysis. The transition in metagene expression and its relative expression compared with those at baseline in poor responders are presented in Fig. 8A, B, respectively (results of all 12 metagenes and blood cytokines are presented in Supplementary Figs. 20 and 21). SKITm17 expression was suppressed in almost all patients throughout dupilumab therapy, with the exception of Pt12, where SKITm17 increased when the dosing interval was extended from every other week to once per month. Conversely, poor responders exhibited skewed expressions toward SKITm10. Skewing toward type 17 response-associated cytokines was less evident in the blood (Fig. 8B).

Fig. 8: Long-term observation of skin metagenes and blood cytokines in dupilumab-treated patients.

A Dynamics of skin metagene expression and blood cytokines in individual patients during the long-term course of dupilumab treatment. Green highlighted shadows represent the 95% CI for the mean expression in healthy individuals. B Transition in relative values compared with the baseline skin metagene expression and blood cytokines in four poorly responding patients. C Clinical images of the four poorly responding patients before and after treatment. No apparent changes in clinical phenotypes were observed during the course of treatment.

Further observation of individual patients revealed notable results. Two poorly responsive patients (Pt21 and Pt23) showed persistently high expression of SKITm5 (IEGs) at non-lesional sites (Fig. 8A); both patients had common clinical phenotypes of generalized erythema (Fig. 8B). Pt17 had extensive areas of LA, consistently observed in the longitudinal study. These results suggested that LA eruptions were persistently associated with high SKITm11 expression (indicating type 1 inflammation). Notably, SKITm10 expression, initially low before dupilumab treatment, increased in LA lesions following dupilumab initiation. No patients developed any evident new skin phenotypes throughout treatment (Supplementary Fig. 22). Finally, the marker genes showed trends similar to those of metagene expression in eight patients (Supplementary Fig. 23), highlighting the potential utility of marker genes for disease monitoring and as additional therapeutic targets in poor responders to dupilumab.

Discussion

Human clinical research often struggles to balance invasiveness and information richness. To address this, we profiled gene expression in large-scale whole-skin samples using an unbiased approach to explore the molecular heterogeneity of AD. Consequently, our 1-mm punch method yielded ample samples for comprehensive analyses with minimal invasion. Additionally, unsupervised decomposition successfully captured various keratinocyte and fibroblast gene expression profiles, revealing both recognized and unrecognized gene signatures underlying AD pathogenesis. This is the first comprehensive study to examine the gene expression profiles associated with local AD phenotypes, overall severity, and treatment outcomes.

Conventional analytical approaches often fall short due to disease population diversity and potential sampling bias. Techniques such as gene set variation analysis and CYBERSORT have been used to assess inflammatory pathways in AD^2,8,37; however, detailed reference data for various cell populations in the skin are limited. Single-cell studies can identify novel cell populations¹, but comprehensive understanding requires many patient samples, and single-cell analysis is typically performed on a limited cohort.

We overcame these challenges using unsupervised learning with NMF to profile numerous whole-skin samples in an unbiased manner, revealing the molecular heterogeneity of AD and responses to different inflammatory signals. NMF, a potent decomposition method widely used in bioinformatics^38,39, extracts gene sets that are co-expressed in the original data, providing a low-rank approximation of high-dimensional expression data. The gene sets, or metagenes, obtained by NMF, have the following features: (1) genes with high weights across multiple metagenes uncover the contribution of single genes to multiple pathways, and (2) genes that are highly weighted in a single metagene serve as specific biomarkers for each pathway. For example, KRT17 is upregulated in both normal sebaceous glands and interfollicular keratinocytes in psoriasis⁴⁰, highlighting the need for caution when evaluating the disease state solely based on the gene expression levels per se. In addition, genes such as FLG, FLG2, and LOR, widely recognized as skin-barrier-related genes, exhibited high weights in multiple metagenes, suggesting their involvement in various skin processes. This may explain why the epidermal barrier metagene (SKITm2) did not significantly correlate with the skin modules in our previous study using weighted gene co-expression network analysis (WGCNA), where each gene is clustered to a single module (Supplementary Fig. 9)²⁹. The identified marker genes specific for each metagene serve as potential clinical biomarkers for assessing diverse skin barrier and inflammatory states, as well as for future precision medicine in AD.

Thus, NMF uncovered a significant collection of gene signatures in AD skin tissue reflecting responses to different inflammatory signals, with minimal influence from sample bias related to skin appendages and technical factors.

In previous studies, gene expression signatures in AD have been compared among lesional, non-lesional, and normal skin, but their relevance to different AD eruptions has been less explored. Our study clarifies the contribution of “AD-associated” molecular profiles to various AD skin phenotypes. The following discusses the metagenes identified by NMF, comprising genes involved in linked to type 2, 17, and 1 inflammatory responses, as well as barrier-related genes.

The SKITm17 metagene, associated with type 2 response, correlated strongly with clinical signs of AD, including local symptoms, severity, and pruritus. Signature genes included key chemokines such as CCL18 and CCL13 produced by macrophages and MMP12 expressed in DCs, and genes such as COL6A5, POSTN, CCL2, and CCL19 expressed in a recently reported unique fibroblast population in AD skin^1,41. Th2-responding fibroblasts expressing COL6A5, POSTN, and IL13RA1 have also been identified in eosinophilic fasciitis in humans⁴². Our findings emphasize the crucial role of fibroblasts and macrophages in the skin with type 2 inflammation.

Multiple metagenes, including SKITm9, SKITm10, and SKITm15, shared common high-weight genes known to be inducible by type 17 inflammation in keratinocytes^{17,19,21,22,23,24}. Our findings revealed distinct correlations among these metagenes, marker genes, and clinical parameters. SKITm10 was linked to lesional sites, while SKITm15 was associated with non-lesional sites. The involvement of these metagenes in our samples may be influenced by race, as all participants were Japanese, a population shown to have type 17 polarization in AD¹⁸.

The SKITm11 metagene, linked to LA, a chronic AD type, includes genes implicated in type 1 inflammation. In AD, genes such as MX1, STAT1, CXCL9, and CXCL10 have been implicated in type 1 inflammation with “chronic” lesions and African-American AD; however, other studies have shown conflicting results^37,43,44. The inconclusive results in these studies may stem from inadequate consideration of local skin phenotypes and inconsistent definitions of “chronic disease.” Discrepancies arise, as some define “chronic AD” based on the disease duration (e.g., childhood-onset adult AD)⁴⁴, whereas others focus on the local skin lesion duration (e.g., >72 h)⁴³. By examining eruption types, we confirmed the clinical significance of type 1 inflammation in AD skin rather than relying on divergent definitions across studies.

Within the barrier-related metagene SKITm2, we identified unrecognized genes as potential biomarkers for barrier disruption. The high-weighted genes in SKITm2, such as FLG, FLG2, and LOR, are well-known as genetic risk factors in AD^7,45. Nevertheless, their utility as universal biomarkers for barrier function has not been shown^46,47. Here, within the barrier-related metagene SKITm2, we identified unrecognized genes, namely PHYHIP, UGT3A2, CACNA1H, and BTC, as potential biomarkers for barrier disruption. Their roles in skin barrier functions, however, remain to be explored⁴⁸. Our emphasis on potential markers resulting from unbiased analysis warrants a comprehensive evaluation of skin barriers in AD, transcending conventional gene markers rooted in one-sided hypotheses.

SKITm8, linked to epidermal terminal differentiation in the granular layer^49,50, was upregulated in chronic AD phenotypes, such as lichenification, PN, and LA resulting from repetitive scratching⁵¹. These eruptions also share common histological features, such as hyperkeratosis and hypergranulosis⁵². HRNR, a paralog of FLG2 and a marker gene of SKITm8, has shown conflicting expression patterns in AD lesions in previous studies^53,54, highlighting the importance of considering eruption types. Thus, genes in SKITm8 may reflect “abnormal” epidermal differentiation following scratching in AD skin.

Our investigation on non-lesional sites indicated distinctive pathology in AD compared with that of psoriasis. In AD, even clinically asymptomatic skin exhibits barrier dysfunctions and subclinical inflammation^55,56, whereas, in psoriasis, non-lesional skin resembles healthy control skin⁵⁷. This distinction emphasizes the importance of evaluating non-lesional sites in AD.

This study revealed unique associations between detailed clinical symptoms and conventional AD-associated genes. Unlike previous transcriptomic studies that often lacked comprehensive clinical data, our study included this crucial information, allowing us to uncover the intricate relationships between molecular profiles and clinical manifestations. This integrated approach underscores the importance of detailed clinical data in understanding the heterogeneous nature of diseases, such as AD, and provides consistent, accurate insights into complex disease pathogenesis.

Beyond the reidentification of known AD-associated genes, our unbiased approach revealed two significant gene clusters not well-recognized in previous AD studies: IEG-related genes (SKITm5) and fibroblast-derived genes associated with ECM organization (SKITm16)⁵⁸.

The role of IEGs in AD pathogenesis is not well understood; however, in vitro studies suggest their association with “tissue-based memory.” IEGs are transcription factors rapidly and transiently transcribed in response to acute stress or proliferation signals, with expression levels that are sensitive to timing^59,60. Repetitive IL-13 stimulation in alveolar epithelial cells leads to sustained IEG activation and hyper-reactivity, resulting in increased cytokine and chemokine production⁶¹, implying a role in chronic type 2 inflammation in asthma, known as “tissue-based memory”⁶². Our results suggest that IEG-related genes (SKITm5) in non-lesional sites could serve as biomarkers for disease severity and dupilumab treatment outcomes (Figs. 4D and 5C). This supports the idea that IEG expression may reflect “latent inflammation” in non-lesional AD skin. Additionally, the inclusion of epidermal growth factor receptor ligands in the same metagene suggests underlying activation of keratinocytes.

SKITm16, comprising fibroblast-derived genes, showed a strong negative association with disease severity and dupilumab treatment outcomes. Skin fibroblasts are crucial in ECM organization and immune regulation⁶³, but the molecular features of hypofunctional fibroblasts and their clinical impact on skin diseases are largely unknown. Single-cell studies have highlighted fibroblast heterogeneity in normal skin⁵⁸, yet their utility as biomarkers in diseased skin remains uncertain. Genes in SKITm16, specifically fibroblast-derived genes, have been linked to other skin diseases. For example, CCDC80 was identified in a genome-wide association study of AD⁶⁴, and LOX, overexpressed in the lungs and skin of patients with systemic sclerosis⁶⁵, was associated with dupilumab treatment outcomes in our study. Downregulation of LOX expression impairs ECM cross-linking, affecting its structural integrity and stability⁶³. RECK and PI16, expressed in endothelial cells, inhibit MMP2 involved in ECM breakdown and are suppressed by inflammatory cytokines such as IL-6 in human endothelial cells⁶⁶, suggesting their responsiveness under inflammatory conditions. The downregulation of RECK is also linked to poor prognosis in multiple cancers^66,67. These findings highlight the significant role of impaired ECM organization in the inflammatory pathogenesis of AD and suggest that these genes could be important biomarkers for assessing disease severity and treatment responses.

To achieve personalized AD therapy, understanding both the immunophenotypes and personal molecular signatures under targeted treatments is essential. Identifying pretreatment molecular markers can help predict dupilumab responses, while understanding treatment dynamics can inform alternative strategies.

In the longitudinal study, we found that four metagenes (SKITm10, SKITm16, SKITm5, and SKITm4) and clinical severity were associated with treatment response to dupilumab. Notably, the expression of SKITm10 and its marker genes, linked to type 17 response, were associated with limited treatment outcomes, unlike those linked to barrier impairment and type 2 inflammation (Fig. 5C). Of note, the pretreatment severity of patients was strongly correlated with an increased type 17/type 2 ratio, characterized by skewed expression of SKITm10 relative to SKITm17, and severe patients who exhibited a pure type 2 profile were rare (Supplementary Fig. 24). Given that many biomarkers related to dupilumab response were confounded by disease severity, further investigation is necessary to assess their independence from disease severity.

Temporal analysis of the cohort revealed that dupilumab had the most robust effect on SKITm17, a type 2 response-related metagene, consistent with previous reports^2,41. This effect was continuous for over a year, with recovery observed in barrier-related and ECM organization-related metagenes, SKITm2 and SKITm16.

In contrast, poor responders exhibited skewed expression of SKITm10, characterized by high weights of keratinocyte-derived genes responsive to type 17 inflammation (Figs. 7B, 8A, B). The potential skewing toward a Th17 response under Th2 pathway blockade has been reported^68,69, while other reports have been inconsistent^2,70,71,72. Our study’s unique features, including shorter sampling intervals and stratified analysis, provided new insights. This suggests that patients with inadequate dupilumab response might benefit from JAK inhibitors, known to downregulate type 1, 2, and 17 responses⁷³.

Furthermore, sustained upregulation of IEGs (SKITm5 metagene) in some poor responders suggests an additional therapeutic target. Recently developed topical drugs that inhibit the nuclear transport of transcription factors, such as AP-1, a major component of IEGs⁷⁴, might aid in further disease control for those patients with sustained IEG expression during systemic treatment. Our findings help identify potential genetic biomarkers to stratify populations that would benefit from IEG-targeted therapy in AD.

Our study had some limitations. For example, all participants were Japanese, warranting further research on other races. Additionally, we could not directly evaluate immune cells, presumably due to their small numbers in the samples. Nevertheless, our unbiased approach, fortified by minimally invasive skin sampling and integrative analysis with comprehensive unsupervised decomposition method and detailed clinical information, helped elucidate the complex links between heterogeneous molecular signatures and clinical AD features in skin tissue. The cross-sectional analysis highlighted the distinct clinical significance of both recognized and previously unrecognized gene sets in AD skin. The longitudinal analysis identified potential biomarkers and additional targets for dupilumab treatment. The marker genes identified offer a robust framework for assessing each pathway, with potential applications across various study settings, including other datasets and small-sample-size research. Altogether, our findings provide comprehensive insights into the complex molecular heterogeneity of AD, paving the way for future rational use of molecular-targeted treatments in personalized medicine.

Methods

Patients and sample collection

The study protocol was approved by the Keio University School of Medicine Ethics Committee (Approval Numbers: 20160377, 20150325, and 20130384), Nippon Medical School Ethics Committee (29-03-916), Tokyo Teishin Hospital (1054), and RIKEN Yokohama Ethics Review Committee (Approval Number H28-24), aligning with the Declaration of Helsinki. Patients with AD, plaque-type psoriasis visiting Keio University Hospital, Nippon Medical School Hospital, and Tokyo Teishin Hospital between 2018 and 2022 were invited to participate. AD was diagnosed according to the criteria of Hanifin and Rajka⁷⁵, and psoriasis diagnosis was based on clinical symptoms or previously obtained skin histopathology, reviewed by at least two board-certified dermatologists. Normal control skin was obtained from healthy volunteers who were recruited at Keio University Hospital, defined as individuals with no inflammatory skin diseases. Exclusion criteria included age <15 years, pregnancy, and a history of lidocaine allergy, bleeding disorders, and complications of other inflammatory skin diseases. Written informed consent was obtained from all participants: 160 patients with AD, 18 patients with psoriasis, and 26 healthy volunteers (Supplementary Fig. 1). Participants who underwent skin sampling were compensated with 10,000 JPY. The sex (biological attribute) of the participants was determined based on self-reporting.

A 1-mm punch biopsy (Kai Medical, Japan, #BPP-10F) was used for skin sampling, using local anesthesia with topical lidocaine/prilocaine cream (EMLA cream; Sato Seiyaku, Tokyo, Japan, #1219800N1023) or 1% lidocaine injection (Xylocaine Injection Polyamp 1%; Sandoz Pharma K.K., Tokyo, Japan, #1214400A7055). Lesional and non-lesional samples were obtained based on consent, and “lesional” samples were defined by the presence of one or more skin symptoms: erythema, papulation/induration, excoriation, and lichenification. Sampling sites were primarily selected based on the most severe or characteristic lesions of the posterior trunk in patients with AD or psoriasis. Sampling was performed at multiple time points after obtaining patient consent.

During sampling, we avoided hair follicles by selecting sampling sites through dermoscopy, a microscopy tool commonly used in dermatological clinical practice. This precaution was adopted to mitigate potential confounding effects on the transcriptomic data. However, in some patients, hair follicles were undetectable, likely owing to edema or thickening of the epidermis. To acquire control data, we collected two skin samples from healthy volunteers: one from a non-hair follicle site and the other from a hair follicle site.

Each sampling site was rated from 0 to 3 based on the local severity of erythema (redness), papulation/induration (swelling of the skin), excoriation (erosion due to scratching), and lichenification (leathery thickening of the skin with exaggerated skin markings), following the EASI scoring system for AD patients⁷⁶. Excoriated samples were obtained from epithelialized skin adjacent to excoriated erosions to avoid bias arising from a lack of epidermis. Prurigo nodules (PN, hard bumps) and lichen amyloidosis (LA, discrete, hard, small reddish-brown papules), which are rare chronic subtypes of AD skin symptoms³⁰, were also evaluated. Examples of skin symptom annotations are presented in Supplementary Fig. 1B.

For AD patients, the severity of the whole body was assessed using EASI, and the patient-oriented pruritus score was recorded at the time of sampling. The Modified EASI (mEASI) was calculated by excluding the scores of the head and neck from the total EASI score³². Upon EASI scoring, PN was assessed as lichenification scores in accordance with the Harmonizing Outcome Measures for Eczema (HOME) guidance⁷⁶. LA, another subtype of chronic AD skin symptoms, was also scored as lichenification. The suggested EASI range for a “mild” condition is 0–7, 7.1–21.0 for “moderate,” and 21.1–50 for “severe,” based on previous literature³³. Therefore, the “mild” mEASI was set at 6.3, calculated as 7×64.8/72 (based on the maximum score of EASI being 72 and that of mEASI being 64.8). The diagnosis of skin diseases, evaluation of severity scores, and selection of sampling sites involved at least two board-certified dermatologists.

Medications for AD, psoriasis vulgaris, and other complications within 1 week of skin sampling were recorded. AD patient treatments included topical moisturizers, steroids, tacrolimus, JAK inhibitors, oral cyclosporine, nemolizumab, and dupilumab. Psoriasis patient treatments included topical steroids, topical vitamin D analogs, TNF-alpha inhibitors, and IL-17 inhibitors. Topical treatments were marked as “used” if the ointment had been applied to the sampled area within 1 week of sampling. Dupilumab was administered to adults with severe AD meeting the indication criteria under the Japanese national insurance; these included a body surface area percentage >10%, a Validated Investigator Global Assessment score higher than 3, and EASI > 16, despite a history of treatment with topical steroids (strong class or higher) or topical calcineurin inhibitors for at least the last 6 months, or patients with hypersensitivity to topical medications, rendering it difficult to continue topical treatments.

RNA extraction and sequencing

Samples for RNA transcriptome analysis were immediately submerged and stored in RNAlater solution (Thermo Fisher Scientific, Waltham, MA, USA, #AM7020) at 4 °C and subsequently frozen at −80 °C for storage. RNA was extracted from the samples using TRIzol (Thermo Fisher Scientific, Waltham, MA, USA, #15596026) and the Direct-Zol RNA MiniPrep Kit (ZYMO RESEARCH, Irvine, CA, USA, #R2052) following the manufacturer’s protocol. For quality control, the RNA integrity number (RIN) was calculated using an RNA Pico 6000 kit (Agilent Technologies, Santa Clara, CA, USA, #5067-1513) and Agilent 2100 Bioanalyzer (Agilent Technologies, Santa Clara, CA, USA, #G2939A) with 2100 Expert software (version B.02.07; Agilent Technologies, Santa Clara, CA, USA), following the manufacturer’s instructions. The RNA concentration was measured using a Qubit fluorometer (Thermo Fisher Scientific, Waltham, MA, USA). Samples with RNA concentrations below the measurement range of the Qubit fluorometer (below 1 ng/μl) were measured using a Quantus Fluorometer (Promega Benelux, Leiden, the Netherlands) or estimated from the electrophoresis results using the bioanalyzer. The libraries were prepared using a NEBNext Ultra RNA Library Prep Kit for Illumina (New England Biolabs, Ipswich, MA, USA, #E7530). A total of 10 ng of RNA (for smaller samples, 1 ng) was used for library preparation (18 or 21 PCR cycles, depending on the amount of extracted RNA). Sequencing was performed using an Illumina HiSeq 1500 or 2500 to obtain 15–20 million reads using a 50 bp single-end read configuration. Reads were aligned to the Ensembl GRCh38 human genome assembly using STAR⁷⁷, and gene counts and Transcripts Per Kilobase Million (TPM) were calculated using the R package Rsubread⁷⁸. The quality data for RNA-sequence has been shown in Supplementary Data 5.

Sample filtering and patient demographics

We excluded low-quality samples from the analysis, specifically samples with low expression of representative structural genes from the stratum corneum, epidermis, and dermis were excluded. Specifically, we excluded samples with counts fewer than 10^1.5 TPM for CDSN (used as a marker of the stratum corneum of the epidermis) or 10³ TPM for KRT10 (used as a marker of the spinous layer of the epidermis). No outliers were observed in the distribution of COL1A1 (used as a dermal fibroblast marker). The criteria were set based on gene distribution (Supplementary Fig. 25). In addition, we removed three samples with a read count lower than 10^6.5 and seven samples with a RIN (an index of RNA degradation) lower than 2 or unmeasurable. To evaluate the sample quality of samples with lower RIN scores (<6), we conducted a correlational analysis of disease severity (EASI) and metagene expression in AD samples, stratified by RIN scores of the samples (<6 and ≥6). The results demonstrated a consistent trend across both RIN groups, with similar correlational relationships to disease severity (Supplementary Fig. 26). Notably, the correlations involving the four key metagenes (SKITm2, SKITm16, SKITm17, SKITm10) in the samples with low RIN score (<6) exhibited the same trend as observed in the regression analysis which used all samples with RIN scores between 2 and 10 (Fig. 4D).

Based on these criteria, 25 of the 1061 samples were excluded. Thus, for further investigation, we used 1036 samples: 951 AD samples (529 lesional and 422 non-lesional samples of 156 patients), 36 psoriasis samples (18 lesional and 18 non-lesional samples of 18 patients), and 49 normal control samples (26 healthy individuals). The patient and sample demographics are presented in Table 1. In addition, 43 patients with AD and two healthy volunteers were sampled multiple times on different dates (Supplementary Table 2). Table 1 presents the age and severity at the time of initial sampling for patients from whom samples were collected at multiple time points. Detailed information on the sampled regions and local symptoms is summarized in Supplementary Tables 1 and 3.

mRNA expression analysis

All bioinformatic and statistical analyses were conducted using R (version 3.6.3)⁷⁹. Figures were created using the ggplot2 package (version 3.3.2)⁸⁰, pheatmap package (version 1.0.12)⁸¹, ComplexHeatmap (version 2.2.0)⁸², VennDiagram (version 1.7.3)⁸³, or the pals package (version 1.6)⁸⁴. Gene counts were normalized based on a variance-stabilizing transformation (vst) using the DESeq2 R package (version 1.26.0)⁸⁵. The normalization of gene counts and differentially expressed gene analysis was also performed using DESeq2 R package. Each sample was assigned to one of the five categorical variables according to the disease and presence of symptoms (“lesional AD,” “non-lesional AD,” “lesional psoriasis,” “non-lesional psoriasis,” and “normal”). This variable was included in the design option of the DEseq function, followed by the vst function. The differentially expressed genes were defined as those with an adjusted p < 0.01 and a log fold change greater than 1 or less than -1. For hierarchical clustering using all samples, we selected the 100 genes with the highest variance in the dataset (Supplementary Fig. 2B). Clustering was performed using the ward. D2 method, applying the Euclidean distance.

Quantitative real-time PCR (qRT-PCR)

qRT-PCR was performed to quantify the expression levels of IL17A, IL17C, IL17F, IL23A, CCL20, IL13, CCL13, CCL18, CCL26. Of the samples used in the cross-sectional analysis, 33 samples of lesional AD skin were used for analysis due to considerations of sample quality and the availability of remaining material for PCR analysis. Total RNA (50–100 ng) was reverse-transcribed into complementary DNA (cDNA) using the SuperScript VILO cDNA Synthesis Kit (Thermo Fisher Scientific, #11754050) in a reaction volume of 20 µL, according to the manufacturer’s instructions. TaqMan assays were performed using the TaqMan Fast Advanced Master Mix for qPCR (Thermo Fisher Scientific, #4444556) in a reaction volume of 20 µL, containing 0.2 µL of the reverse transcription product. The details of the TaqMan Gene Expression Assays used in this study are provided in Supplementary Table 10. Reactions were run on a QuantStudio 12 K Flex Real-Time PCR System (Thermo Fisher Scientific) under the recommended cycling conditions. Gene expression data were analyzed using ExpressionSuite Software v1.3 (Thermo Fisher Scientific) following the ΔΔCt method. Ct values of the target genes were normalized to that of the reference gene (HPRT1) and expressed relative to the reference sample. Correlational analysis with metagene expression derived from RNA-seq analysis was performed using Spearman’s correlation method (Fig. 3).

Blood cytokine levels

CCL17 levels were measured in serum samples, and other cytokines were measured in plasma samples. IL-17 levels were measured using a fully automatic, highly sensitive immune analyzer (HI-1000; Sysmex Corp., Hyogo, Japan). Levels of other plasma cytokines (CCL27, IL-4, IL-5, EDN, CCL20, IL-17, IL-22, CXCL9, IL-16, IL-18, CCL5, MMP9, IL-6, TNF-α, TSLP, IL-10) were measured using a fully automatic immune analyzer (HISCL-5000 or HISCL-800; Sysmex Corp., Hyogo, Japan)^86,87. The normal serum CCL17 level range was obtained from a previous study⁸⁸. Since the distribution of cytokine levels was right-skewed, we used log- logarithmically transformed values for analysis to correct skewness and enhance data visualization and interpretability (Supplementary Fig. 27). A constant of 0.1 was added before the log transformation to prevent undefined values at 0 pg/mL.

NMF and rank estimate

NMF is mathematically formulated as the problem of decomposing a two-dimensional matrix A, as the product of two matrices, W and H. Unlike PCA methods, NMF does not require an orthogonality assumption based on the matrix W. In this study, A was the original RNA expression N×M matrix (N genes and M samples). Matrix W had a size of N×k, where each of the k columns represents a metagene, and element w_ij represents the weight of gene i for metagene j. Matrix H had a size of k×M, where each of the M columns represents the metagene expression pattern of a sample, and element h_ij represents the expression values of metagene i in sample j.

We performed NMF with the NNLM R package⁸⁹ using the vst-normalized data. Before the analysis, we filtered genes with a variance less than 0.1 or a mean less than 5.5 to exclude insignificant genes and genes potentially related to measurement noise. Additionally, Y-chromosome genes were excluded from the dataset to avoid sex-biased results. Consequently, we used the expression data of 11,270 genes from 1036 samples for NMF analysis. The minimum expression value was set to zero by subtracting the minimum value of each gene from the expression value of each gene. We estimated the optimal rank k according to the masking approach^90,91. Specifically, 1% of elements were randomly deleted from the original matrix and then imputed with NMF through ranks 1 to 50. We then compared the imputed values with the original values by calculating the mean squared error (MSE) between the values to evaluate the reconstruction errors. The evaluation procedure was performed as follows:

• 1. Randomly select 1% of elements to be deleted.

• 2. Substitute NA for the selected element of the original matrix A to obtain A_del.

• 3. Decompose A_del into W_del and H_del using the nnmf function with default parameters in the NNLM package.

• 4. Compute the product of W_del and H_del to obtain the reconstructed matrix A_recon.

• 5. Compute the MSE between the original matrix A and the reconstructed matrix A_recon for the deleted elements only.

The aforementioned steps were repeated 20 times for each rank, and the median reconstruction error for ranks 1–50 was used to determine the optimal rank k (Supplementary Fig. 2C). Next, among the optimal ranks of 25–34 that presented the smallest median errors, k = 29 was chosen as the factorization rank for our dataset. Finally, we decomposed the dataset based on the optimal rank k using the nnmf function with max.iter = 10,000 to calculate the metagenes and the loading matrix.

The “metagene” matrix resulted in 11,270 rows and 29 columns, with each value w_ij, which was the weight of gene i in metagene j. The “loading” matrix had 29 rows and 1036 columns; value h_ij represented the expression level of metagene i in sample j. The distribution of gene weights and expression levels in each metagene are shown in Supplementary Figs. 3 and 5.

We declared results with a p < 0.01 as significant throughout the analysis after NMF. As R does not store values smaller than 2.225074e-308 (obtained based on.Machine$double.xmin), an output p value of 0.e + 00 by R was denoted as a p < 2.2e-308 in the figures and text.

Definition of “highly-weighted” and “marker” genes

To interpret the cellular components and pathways associated with the metagenes, we defined the “highly-weighted” genes in the metagene matrix, which were evaluated from the right-skewed distribution (Supplementary Fig. 3), as genes with z-scores > 3. These were obtained by z-normalizing the weights of a metagene (z-score_g). Of the 11,270 genes, 2368 were selected as “highly weighted” genes. The raw weights of the highly weighted genes within the 29 metagenes are as a heatmap in Fig. 2B. The z-score_g of the highly-weighted genes within the 29 metagenes are presented in Supplementary Data 2 and 3. The maximum number of “highly-weighted” genes in a metagene was 263, with a minimum of 57, averaging 137.

Furthermore, to search for candidate biomarkers of each metagene, we searched for “marker” genes specifically included in each metagene. The z-score (z-score_m), a measure of the specificity of a particular metagene, was calculated by z-normalizing the z-score_g again across the metagene. We extracted genes with a z-score_m > 3 that were highly correlated (Pearson’s correlation coefficient > 0.7, p < 0.01) with metagene expression in all skin samples (Supplementary Fig. 28). These genes were defined as “marker” genes. SKITm3, SKITm4, SKITm6, SKITm13, SKITm14, and SKITm20 had no marker genes that met these criteria. The top 10 “highly-weighted” and “marker” genes are presented in Supplementary Data 4.

As for the biological roles of the metagenes, we verified them using knowledge-based approaches in reference to previous in vivo and in vitro studies based on their documented roles, involvement, or activation in skin disorders. For example, SKITm9, SKITm10, and SKITm15, had high weights of genes activated in the skin through IL-17 or IL-22 and related cytokine stimulation. The representative enrichment analysis of the highly weighted genes in SKITm10 with KEGG pathway analysis performed with enrichR⁹² and Fisher’s exact test are presented in Supplementary Fig. 4 and Supplementary Table 4. The KEGG pathway enrichment analysis included “IL-17 signaling pathway” in the top 10 enriched KEGG pathways (Supplementary Fig. 4). Additionally, we used Fisher’s exact test to evaluate the significance of the overlap between highly weighted genes (Supplementary Table 4) and marker genes (Supplementary Table 5) with genes reportedly induced by IL-17 and its related cytokines in keratinocytes^{21,25,26,27,28}. We constructed a 2 × 2 contingency table to calculate p values and odds ratios.

Statistical analysis of the sample quality and sequence batch

We further studied the association between metagene expression and the symptoms or treatment effects using relative expression values, which were calculated by dividing the expression value of each sample by the maximum expression value in each metagene. The relative expression value was used for analysis throughout the cross-sectional and longitudinal analysis.

Pearson’s correlation coefficients were used to evaluate the correlation between metagene expression and the RIN, total read counts, and the amount of extracted RNA (Supplementary Fig. 6A–C). The amount of RNA was too low to be measured in seven samples, and these samples were thus excluded from this analysis. The RNA amount and total read counts were log₁₀-transformed. Metagenes with correlation coefficients >0.3 or <−0.3 and a p < 0.01 were considered relevant.

The sequence batch effect was also assessed because RNA sequencing was carried out in nine separate sessions owing to the large number of samples. For this analysis, we compared two linear regression models fitted to the data, specifically model 1 with sequence sessions (SQ1–SQ9) as an explanatory variable and model 2 without these sessions. The models were built using a linear model (lm function) and compared by performing an ANOVA test (anova function). To evaluate the effect of sequence sessions on metagene expression depending on the patient background, both models were controlled for sex, the use of cyclosporine, months used after the initiation of dupilumab (0 if not on dupilumab, 1–5 for the first 5 months, and ≥6 months as month 6), disease activity, and the calendar month of the sampled date. The p values in these analyses were corrected using the Bonferroni method for multiple comparisons (Supplementary Fig. 6D).

In summary, SKITm28 was weakly correlated with the RIN (Supplementary Fig. 6A, r = −0.32, p < 2.2e-308). SKITm29 was negatively correlated with the read counts (Supplementary Fig. 6B, r = −0.49, p < 2.2e-308). In addition, SKITm28 and SKITm29 were most significantly associated with sequence sessions (Supplementary Fig. 6D). Thus, we considered SKITm28 and SKITm29 to be related to the sample quality and sequence batch. SKITm14 and SKITm20 were highly expressed in only one sample each (Supplementary Figs. 5 and 6); therefore, we assumed that these two samples were outliers owing to unknown sample bias. The expression of three metagenes (SKITm2, SKITm7, and SKITm10) was correlated substantially with the amount of RNA extracted (Supplementary Fig. 6C). However, the expression of these metagenes did not correlate with the RIN (Supplementary Fig. 6A), and thus, they were considered independent of sample quality.

To evaluate the effect of sample biases on the expression data, we conducted a UMAP analysis using the expression of all metagenes (Supplementary Fig. 7A), as well as a subset of metagenes that was considered unrelated to sample bias (Supplementary Fig. 7B). Specifically, for Supplementary Fig. 7B, we excluded data corresponding to metagenes associated with skin appendages and reticulocyte inclusion (SKITm22–27, SKITm19), outlier samples (SKITm14, SKITm20), and samples affected by quality issues and sequence batch effects (SKITm28, SKITm29). The UMAP analysis was performed using the umap function from the uwot package (version 0.1.8)⁹³.

The association between skin metagenes identified in this study and transcriptomic modules in our prior study was verified as follows²⁹. The previous study was conducted on 315 skin samples, including lesional, non-lesional AD, and normal samples. Of these, 219 samples were common to the present study. First, we assessed the similarity between the highly weighted genes in the 29 metagenes and the highly contributing genes in the 21 skin gene modules using the Jaccard index, which was calculated by dividing the intersection by the union of the genes (Supplementary Fig. 9A). Next, we calculated Spearman’s correlation between the weights of common genes included in the skin module with the highest Jaccard index and the corresponding skin metagene (Supplementary Fig. 9B). For better interpretability, negative weights of skin modules were converted to positive values, and the minimum value was shifted to zero before calculation.

Analysis of association among metagene expression in disease, clinical symptoms, and disease severity in the cross-sectional study

In the cross-sectional study, we analyzed the association between metagene expression levels and the disease severity and local severity of various types of eruptions. Samples from patients not receiving dupilumab treatment were used in this analysis (lesional AD, n = 246; non-lesional AD, n = 170; lesional psoriasis, n = 18; non-lesional psoriasis, n = 18; normal, n = 49; patient demographics are presented in Supplementary Tables 6–8).

Metagene expression between diseases (normal vs. lesional AD, normal vs. non-lesional AD, normal vs. lesional psoriasis, and normal vs. non-lesional psoriasis) was tested using the Wilcoxon test. P values were corrected using the Bonferroni method for multiple comparisons (Fig. 2C and Supplementary Fig. 8). We determined a negative or positive association based on the subtraction of the expected value from the statistical value, with a negative or positive sign indicating the direction of the association. To assess the effect of skin symptom severity (EASI, pruritus score, local severity of erythema, papulation/induration, excoriation, and lichenification) on metagene expression in the sampled skin, we used a linear mixed-effects regression model with random intercepts for each patient. The following model was used for this purpose:

$${{{\rm{Metagene}}}}\; {{{\rm{expression}}}} \sim \, \alpha+\beta 1\,({{{\rm{symptom}}}}\; {{{\rm{severity}}}})+\beta 2+\ldots .\, \\ +\beta k\,({{{\rm{covariates}}}})+\mu \,({patient\; id})+\varepsilon$$

(1)

where α is the overall intercept, β1 is the fixed effect coefficient for the skin symptom levels, β2 to βk are the fixed coefficients for the covariates (age, sex, and treatments; oral cyclosporine, topical steroids, tacrolimus, and moisturizers within 1 week of sampling, as categorical values of 0 or 1), μ represents the random intercept for the patients, and ε is the residual term. We used patient ID as a random effect variable because 23 patients were sampled at least twice, and samples from 59 and three patients were obtained simultaneously from more than two regions at the lesional and non-lesional sites, respectively. The use of topical JAK inhibitors and nemolizumab was not included in the regression model because of the small sample size (four samples from two patients on a topical JAK inhibitor and two samples from one patient on nemolizumab). In the analysis of PN and LA, which are rare skin symptoms of AD, we compared metagene expression in lesional and non-lesional sites obtained from the same patients (PN, eight patients; LA, six patients) using a paired t-test. The results are presented in Fig. 4D and Supplementary Fig. 11. The effects of blood cytokine levels and clinical biomarkers on metagene expression were analyzed using the following model:

$${{{\rm{Metagene}}}}\; {{{\rm{expression}}}} \sim \ {{{\rm{\alpha }}}}+\beta 1\,({{{\rm{blood}}}}\; {{{\rm{cytokine}}}}\; {{{\rm{levels}}}})+\beta 2+ \ldots . \, \\ +\beta k\,({{{\rm{covariates}}}})+{{{\rm{\mu }}}}\,({{{\rm{patient\; id}}}})+{{{\rm{\varepsilon }}}}$$

(2)

$${{{\rm{Metagene}}}}\; {{{\rm{expression}}}} \sim \, {{{\rm{\alpha }}}}+\beta 1\,({{{\rm{blood}}}}\; {{{\rm{clinical}}}}\; {{{\rm{biomarkers}}}})+\beta 2+\ldots .\, \\ +\beta k\,({{{\rm{covariates}}}})+\mu \,({{{\rm{patient}}}}\; {{{\rm{id}}}})+\varepsilon$$

(3)

where β1 represents the fixed effect coefficient for the blood cytokine levels (adding 0.1 and log₁₀-transformed). The same fixed effect variables in β2 to βk were used in the analysis of the skin symptoms. The results are summarized in Fig. 4E and Supplementary Fig. 12. The 95% confidence interval (CI) of predicted values in Supplementary Fig. 11 was visualized using the ggpredict function in the ggeffects package (version 1.1.2)⁹⁴.

In the analysis of expression profiles of samples with the same local severity and phenotype obtained from the same patients (n = 7) at different time points (Supplementary Fig. 29), the local severity was assessed by the total score of erythema, lichenification, and induration/papulation. The clinical phenotypes that exhibited distinctive expression profiles (excoriation, PN, and LA) were not included to exclude the effect of the clinical phenotypes on the expression profiles. In addition, patients on systemic treatment were not included to exclude treatment effects. A paired Wilcoxon test was used to assess the difference in the metagene expression values.

Transition in skin metagenes and blood cytokines in the longitudinal study of patients treated with dupilumab

In the study involving 24 patients treated with dupilumab, we employed a linear mixed model with random intercepts for each patient. Twenty-four patients from whom a sample could be obtained on the day of dupilumab initiation (month 0) and more than four times in the first sequential 6 months based on calendar months (months 1–6) were included in the analysis. A window of ±7 days was allowed for all time points. The patient demographics are presented in Supplementary Table 9. We tested the changes in metagene expression and blood cytokine levels during treatment using the following model:

$${{{\rm{Metagene}}}}\; {{{\rm{expression}}}} \sim \, {{{\rm{\alpha }}}}+\beta 1\,({{{\rm{months}}}}\; {{{\rm{since}}}}\; {{{\rm{the}}}}\; {{{\rm{start}}}}\; {{{\rm{of}}}}\; {{{\rm{treatment}}}}\; {{{\rm{by}}}}\; {{{\rm{dupilumab}}}}) \\ + \beta 2+\ldots .\, +\beta j\,({{{\rm{covariates}}}})+\mu \,({{{\rm{patient}}}}\; {{{\rm{id}}}})+\varepsilon$$

(4)

$${{{\rm{Blood}}}}\; {{{\rm{cytokine}}}}\; {{{\rm{level}}}} \sim \, \alpha+\beta 1 \,({{{\rm{months}}}}\; {{{\rm{since}}}}\; {{{\rm{the}}}}\; {{{\rm{start}}}}\; {{{\rm{of}}}}\; {{{\rm{treatment}}}}\; {{{\rm{by}}}}\; {{{\rm{dupilumab}}}}) \\ + \beta 2+\ldots .\,+\beta j\,({{{\rm{covariates}}}})+\mu \,({{{\rm{patient}}}}\; {{{\rm{id}}}})+\varepsilon$$

(5)

where β1 is the fixed effect coefficient for the number of months since the start of dupilumab, which was transformed into seven categorical variables (months 0–6), β2 to βk are the fixed coefficients for the treatment covariates (oral cyclosporine, topical steroids, tacrolimus, and moisturizers within 1 week of sampling, as categorical values of 0 or 1), μ represents the random intercept for the patients, and ε is the residual term. All skin samples used in this analysis were obtained from the posterior trunk and, whenever feasible, were sequentially sampled from the same skin phenotype within each patient. In addition, the analysis of skin did not include samples with excoriation scores ≥1, since skin excoriation was observed only sporadically in the patients. P-values were corrected using the Bonferroni method for multiple comparisons. The results of the regression analysis on skin metagene expression and blood cytokine levels are presented in Fig. 7A and Supplementary Fig. 17. The upper and lower limits of the 95% CI for the mean of normal controls were calculated by subtracting or adding 1.96× the standard error to the mean. The linear mixed models employed in all analyses were fitted using the lmer function in the lmerTest package (version 1.1-23)⁹⁵.

Response to dupilumab treatment

The mEASI reduction rates in dupilumab-treated patients were assessed in 22 patients from whom the EASI was available at baseline and month 6 (data for two patients were unavailable) (Supplementary Table 9). After 6 months of treatment, 14 of 22 patients achieved a mEASI75 (75% improvement in the EASI from baseline), and 20 of 22 patients achieved a mEASI50 (50% improvement in the EASI); no data were available for two patients.

However, the absolute value of severity scores, rather than the reduction rate, has recently been discussed as a better measure for assessing treatment outcomes and long-term disease control in real-world clinical settings^96,97,98. Accordingly, we classified the patients into three groups based on their absolute severity scores during treatment: the “early” responders achieved and maintained a mild disease level (defined by mEASI ≤ 6.3) within the first month and throughout 6 months of dupilumab treatment, “poor” responders never reached the mild level during the entire 6-month study, while “intermediate” responders did not meet the criteria for “early” or “poor” (Fig. 5A).

We then assessed differences in pretreatment metagene expression and blood biomarkers among the three patient clusters using Spearman’s correlation analysis. For this analysis, response categories were converted to numeric variables (“early” = 1, “intermediate” = 2, “poor” = 3), and correlations were calculated with respect to the expression values of metagenes, marker genes, and blood cytokine levels (Fig. 5C–E, Supplementary Figs. 15 and 16). We also evaluated the correlation between disease severity scores post-dupilumab treatment and pretreatment levels of skin and blood biomarkers using Spearman’s correlation (Fig. 6). Marker genes from SKITm15, which did not exhibit significant results, are not included.

Five patients receiving cyclosporine at the start of dupilumab treatment were excluded from the correlation analysis. Additionally, four patients with no non-lesional regions in the posterior trunk area were excluded from the correlation analysis of non-lesional samples.

Reporting summary

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