We identified transcriptomic studies from GEO, ArrayExpress, and additional repositories through a structured multi-step screening workflow, as depicted in Figure 1a . After removing duplicates (n = 9) and screening 1,080 records by title and abstract, 96 records were retained for full-text or protocol assessment. Of these, 88 were excluded for reasons such as non-scarring alopecia (n = 31), absence of mRNA expression data (n = 29), lack of transcriptomic results (n = 22), or other design limitations (see full criteria in Information).

A total of eight datasets were included in the qualitative synthesis, and six were deemed eligible for meta-analysis based on platform compatibility and data availability ( Supplementary Tables S3 - S10 ). These comprised 134 patients with primary lymphocytic scarring alopecias (PLSAs)—including lesional and non-lesional samples—and 49 healthy controls (HCs), analyzed using bulk RNA-seq (Illumina HiSeq) or Affymetrix microarrays (U133 Plus 2.0 or Clariom S) ( Table 1 ). Control representation included 19 individual samples and three pooled samples, each composed of RNA from 10 individuals. Disease subtype distribution included FFA (n = 48), LPP (n = 45), and CCCA (n = 41). PsPB samples were reserved for exploratory analyses and excluded from the meta-analysis due to platform incompatibility (Operon v2 21k).

All samples underwent harmonized preprocessing, including log₂ transformation, gene symbol mapping, and batch correction to ensure consistency across studies ( Figure 1b , Supplementary Figure S1 ). Detailed reasons for exclusion of specific datasets and repository-level breakdowns are reported in Supplementary Tables S3 - S8 .

A total of 2,509 differentially expressed genes (DEGs) were identified across PLSAs following batch correction and meta-analysis, including 359 upregulated and 2,150 downregulated transcripts ( Figure 2f , Supplementary Figures S1 , S2 ). A robust core transcriptomic signature was shared by all three subtypes, comprising 53 consistently upregulated and 871 downregulated genes ( Figures 2a, b , Supplementary Table S11 ). Subtype-specific DEG counts were: FFA (219 up, 1,212 down), LPP (112 up, 1,199 down), and CCCA (158 up, 1,931 down) ( Figures 2c-e ). Notably, FFA exhibited the highest number of uniquely upregulated genes (n=27), whereas CCCA displayed the greatest number of uniquely downregulated genes (n=309), underscoring both shared and distinct transcriptional programs.

All scarring alopecia subtypes exhibited both shared and distinct patterns of functional dysregulation.

Over-representation analysis (ORA) of GO and Reactome terms consistently highlighted activation of TGF-β signaling, fibroblast–stroma interactions, and epithelial–mesenchymal transition (EMT) pathways ( Figure 4a ). Gene set variation analysis (GSVA) confirmed significant downregulation of sebocyte-specific transcriptional programs, particularly in FFA and CCCA, underscoring sebaceous gland loss as a shared pathological feature ( Figure 4b ).

Subtype-specific signatures were also apparent. CCCA showed upregulation of fatty acid metabolism and leukocyte migration pathways, along with repression of negative immune regulators, suggestive of active stromal remodeling with limited immune diversity. In FFA, we observed increased activation of keratinocyte and stromal pathways, elevated macrophage and stromal GSVA scores, but minimal lymphocytic enrichment, pointing to a predominantly epithelial–stromal crosstalk. Conversely, LPP exhibited marked enrichment of keratinocyte, fibroblast, and T cell (CD4 ⁺/CD8 ⁺) signatures, consistent with a T cell–mediated inflammatory phenotype.

All three PLSA subtypes exhibited a conserved stromal remodeling program, marked by increased fibroblast and adipocyte signatures, alongside elevated StromaScore and MicroenvironmentScore, consistent with a shared fibrotic axis. Sebocyte depletion was most pronounced in FFA, aligning with its characteristic glandular atrophy ( Figure 5 ). Despite this shared stromal expansion, immune profiles diverged markedly across subtypes. CCCA displayed robust fibroblast expansion and enrichment of plasmacytoid dendritic cells (pDCs), monocytes, and basophils, but minimal involvement of adaptive immune cells. This innate-skewed profile suggests limited antigen presentation and may explain the lack of overt lymphocytic infiltration. In contrast, LPP exhibited a chronic adaptive immune signature, including CD4 ⁺ central and effector memory T cells, CD8 ⁺ T cells, and loss of naive B and plasma cells, indicating persistent epithelial–immune crosstalk. FFA featured M2 macrophage and keratinocyte enrichment, Treg accumulation, and reduced sebocyte and memory B-cell signatures, suggesting fibrosis driven by innate responses within a disrupted epithelial barrier. Additional distinctions included opposite trends in granulocyte lineages, such as increased basophils in CCCA vs depletion in LPP, and progenitor cell expansion (e.g., pro-B, CLP, CMP), which differed by subtype. Together, these data define a core stromal program shared across scarring alopecias, coupled with disease-specific immune microenvironments that may drive divergent clinical phenotypes and therapeutic responses.

Supplementary Figure S6 explores the alignment between GSVA enrichment scores from the meta-analysis and transcriptomic responses to brepocitinib at week 24 in a phase 2a trial. Inflammatory pathways enriched in LPP and FFA—such as Th1/IFNγ, JAK-STAT, Th17, Th22, and NK cell activation—were downregulated following treatment, indicating potential pharmacologic reversibility. This modulation was most pronounced in FFA, consistent with its strong baseline inflammatory signature. In contrast, CCCA exhibited minimal modulation across these pathways. Notably, follicular keratin programs—suppressed across all subtypes, particularly in CCCA—showed upregulation post-treatment, suggesting partial restoration of follicular gene expression. Fibrotic and extracellular matrix pathways remained unchanged or were further elevated, implying limited efficacy of JAK inhibition on fibrotic remodeling.

Downregulated genes across PLSA subtypes—particularly in FFA and CCCA—were significantly enriched for cardiometabolic pathways, including adipogenesis, lipid storage, insulin signaling, and brown adipose tissue regulation ( Figure 6 ). These transcriptional alterations encompassed reduced expression of key regulators of cholesterol metabolism (e.g., decreased HDL/LDL ratio, triglyceride biosynthesis), adipokine signaling, and glucose homeostasis. In FFA, strong associations were observed with markers of insulin resistance, type 2 diabetes, and impaired adipose morphology, aligning with reported clinical comorbidities. CCCA exhibited similar trends with additional links to arterial stiffness, vascular tone regulation, and myocardial remodeling.

Conversely, upregulated genes were enriched for immune-mediated cardiovascular traits such as monocyte/macrophage activation, vascular inflammation, and coronary artery disease. Traits including coronary stenosis, myocardial infarction, and atherosclerosis were transcriptionally aligned with the inflammatory signatures of PLSAs. Collectively, these findings suggest that chronic scalp inflammation in scarring alopecias may engage systemic cardiovascular risk pathways at both metabolic and immunologic levels, supporting epidemiologic observations and underscoring the relevance of comorbidity screening in affected patients.

This systematic review and meta-analysis adheres to PRISMA 2020 guidelines and a pre-registered PROSPERO protocol, enhancing transparency and minimizing selection bias. We integrated six harmonized transcriptomic datasets—including two previously unpublished—across RNA-seq and microarray platforms. To reduce technical variability, we applied batch correction and linear mixed-effects modelling. Risk of bias was formally assessed, and robustness was further evaluated through leave-one-study-out (LOSO) sensitivity analyses. Our multilayered framework—spanning gene-level, modular, pathway-level, and cell-type deconvolution analyses—enabled the detection of conserved and subtype-specific disease signatures not apparent in individual studies.

In contrast to the recent narrative review by Bao et al. (61)—which emphasized dysregulated lipid metabolism and sebaceous gland atrophy in CCCA without applying quantitative synthesis—our meta-analysis offers a statistically rigorous framework that integrates data across multiple transcriptomic studies. While we confirmed key features noted by Bao et al., including sebocyte depletion and metabolic dysfunction, our analysis also revealed novel, subtype-specific insights: distinct immune cell infiltration profiles, epithelial–mitochondrial collapse, and transcriptomic links to cardiovascular traits, particularly pronounced in FFA and LPP. Furthermore, by incorporating unpublished datasets and conducting drug repurposing analyses, we identified tractable molecular targets with potential therapeutic relevance.

Nonetheless, several limitations should be acknowledged. Bulk transcriptomics inherently lacks single-cell resolution, limiting precise cellular attribution. Additionally, incomplete clinical metadata (e.g., disease stage, treatment exposure) in some studies restricted the possibility of subgroup stratification. To enhance statistical power and subtype coverage in this rare disease group, we included a subset of studies with moderate to high risk of bias (e.g., pooled controls, unavailability of raw data, or unpublished status). Their inclusion was justified by the scarcity of available data and supported by LOSO analyses, which confirmed that core disease signatures—particularly those involving immune–fibrotic programs—remained stable. In fact, the exclusion of lower-quality datasets occasionally sharpened signal detection, though we recognize that these studies may introduce analytical noise and should be interpreted cautiously.

Finally, the PsPB subtype was excluded from integrative meta-analysis due to platform incompatibility (Operon v2, 21k array). However, exploratory reanalysis of its data (GSE11905) revealed substantial transcriptomic overlap with LPP, including upregulation of interferon-stimulated genes (e.g., MX1, OASL, STAT1), CD8 ⁺ T cell cytotoxic markers (e.g., GZMB, PRF1), and downregulation of follicular keratins and desmosomal genes. These immune-epithelial signatures suggest that PsPB may lie within the LPP spectrum. Although these findings derive from a single, legacy microarray dataset and cannot support definitive conclusions, they highlight the need for renewed transcriptomic profiling of PsPB using modern RNA sequencing platforms to clarify its nosological status almost two decades after its initial description.

This systematic review and meta-analysis followed PRISMA 2020 guidelines and was prospectively registered in PROSPERO (ID: CRD42024559969) (31).

Transcriptomic studies were systematically searched in GEO (32), ArrayExpress (33), MEDLINE, ClinicalTrials.gov, and grey literature sources up to March 23, 2024. Eligible studies included RNA-seq or microarray data from human scalp biopsies of patients with primary lymphocytic cicatricial alopecias (FFA, LPP, CCCA, PsPB) and matched healthy controls or baseline lesional samples. Exclusion criteria included non-human models, non-scarring alopecias, incompatible data formats, or non-mRNA profiling.

Two reviewers independently screened studies and extracted metadata including alopecia subtype, platform, sample type and size, control characteristics, and completeness of annotations.

Dataset-level risk of bias was assessed using a modified ROBINS-I tool adapted for transcriptomic studies, covering six domains: sample selection, platform consistency, metadata completeness, case–control comparability, peer-review status, and conflict of interest disclosure (34, 35).

Microarrays were normalized using RMA; RNA-seq data using variance-stabilizing transformation (VST). DEGs were identified using limma (with voom for RNA-seq), applying FDR-adjusted p < 0.05 and |log₂FC| > 0.5. Intra-dataset correlation was addressed using duplicateCorrelation.

Batch correction across datasets was performed with ComBat. Meta-analysis was conducted using limma with random-effects modeling. Cell-type proportions were estimated with xCell, and LOSO sensitivity analysis was used to evaluate robustness and identify influential datasets.

We applied Overrepresentation Analysis (GO, Reactome) and GSVA on R using curated and custom gene sets representing immune activation, epithelial remodeling, fibrosis, lipid metabolism, cellular stress, and cardiometabolic traits.

PPI networks were built using STRING (score ≥ 0.7), limited to DEGs, and functionally clustered into disease-specific modules (36).

Candidate compounds were identified through enrichment of DEG modules (from PPI networks) against LINCS L1000 and PharmGKB using GeneCodis 4 (37), prioritizing drugs predicted to reverse pathogenic expression patterns.

Transcriptomic profiles from our meta-analysis were directionally compared with TLDA-based brepocitinib trial results at week 24, using visual plots of GSVA and log₂FC for exploratory interpretation.

Differential gene expression was assessed using the limma package, applying empirical Bayes moderation and a random-effects model with the duplicateCorrelation function to account for within-study dependencies. Meta-analyses were conducted separately for each contrast (FFA vs. control, LPP vs. control, CCCA vs. control), and results were considered significant at false discovery rate (FDR) adjusted p < 0.05 and |log₂FC| > 0.5.

For pathway-level analysis, Gene Set Variation Analysis (GSVA) was performed on batch-corrected expression data using both curated (MSigDB) and custom gene sets related to immunity, epithelial biology, fibrosis, metabolism, and cardiovascular traits. Differences between groups were assessed using Wilcoxon rank-sum tests, with FDR correction.

Overrepresentation analysis (ORA) was performed using clusterProfiler and GeneCodis, employing hypergeometric testing with Benjamini-Hochberg correction. Only pathways with adjusted p < 0.05 and minimum gene count ≥ 3 were retained.

In drug repurposing analyses, candidate compounds from LINCS L1000 and PharmGKB were ranked by enrichment scores and p-values derived from permutation-based gene set tests. Reversal of disease expression signatures was prioritized using negative connectivity scores and pathway concordance.

All analyses and plots were implemented in R version 4.3.2 and Bioconductor 3.17.