From Genetic Variants to Immune Dysregulation: Multi-Omics Mendelian Randomization Decodes Molecular Networks in Rosacea

Abstract

Background

Rosacea is a chronic inflammatory skin disease with incompletely understood pathogenesis.

Objective

This study aimed to identify key genes and regulatory mechanisms using multi-omics Mendelian randomization (MR) and machine learning, and to validate the findings in animal experiments.

Methods

Bidirectional MR analysis using eQTL, pQTL, and mQTL datasets (GWAS ID: GCST90476178) identified rosacea-associated genes. LASSO regression and random forest were used to screen key genes. Immune infiltration, GSEA/GSVA, non-coding RNA network analysis, and molecular docking were used to explore gene functions and drug targets. Real-time quantitative polymerase chain reaction (RT-qPCR) and Western blot validated findings in a rosacea mouse model.

Results

Five key genes (GSTM3, NEO1, UBE2M, CD300A, and EPHB4) were identified. These genes correlated with plasma cells and Tregs and were enriched in Wnt, transforming growth factor (TGF)-β, and interleukin (IL)-17 pathways. Molecular docking predicted the binding of CD300A to bisphenol A and GSTM3 to niacinamide. RT-qPCR confirmed gene expression changes. Western blot supported IL-17 pathway activation.

Conclusion

This study reveals five key genes and immune-related mechanisms in rosacea, offering novel targets for diagnosis and therapy.

INTRODUCTION

Rosacea, a chronic facial inflammatory disease¹,², shows a 5.1% prevalence with notable early onset (25–39 years)³ and significant ethnic variation³,⁴,⁵. It comprises four subtypes (erythematotelangiectatic, papulopustular, phymatous, and ocular)²,⁶, presenting with symptoms like persistent redness, papules, and telangiectasia⁷. While rosacea mechanisms remain partially understood, it likely arises from gene-environment interactions. Key drivers include innate immune activation, neurovascular dysregulation, skin barrier dysfunction, microbial dysbiosis, and systemic inflammation⁸. TLR2 overexpression induces kallikrein 5-mediated cleavage of cathelicidin into LL-37, triggering inflammation and vasodilation⁹. Neuropeptides and TRPV channels contribute to facial flushing and stinging, while disrupted lipid balance and pH shifts weaken the skin barrier¹⁰. Skin barrier dysfunction, with altered lipid composition and pH imbalance, amplifies immune responses¹¹. Increased Demodex folliculorum density and its bacteria further activate TLR pathways¹². Gut dysbiosis may also affect skin immunity through the gut-skin axis¹³, with associations to systemic diseases reinforcing rosacea as a systemic inflammatory condition¹⁴.

The treatment of rosacea should be adjusted according to the phenotype. Common treatments include topical ivermectin, oral doxycycline, and vasoconstrictors⁶. Brimonidine, pulsed-dye laser, and intense-pulsed light reduce redness and telangiectasia, though effects are transient¹⁵. Inflammatory bumps respond well to ivermectin and low-dose doxycycline¹⁶. Supportive skincare and trigger avoidance are key for maintenance¹⁷. Current treatments have limited efficacy for persistent erythema, and lack personalized options¹⁵,¹⁸. Emerging therapies such as minocycline foam and botulinum toxin A show promise but lack large-scale evidence¹⁶,¹⁸.

Bioinformatics and large-scale transcriptomics facilitate gene and pathway mapping for diseases. In rosacea, microarray studies reveal altered expression of IL-1B, MMP9, CXCL8, and TLR8, suggesting activation of innate immunity and inflammation¹⁹. Analyzing multi-omics data with Mendelian randomization (MR) and machine learning (ML) helps identify causal genes and regulatory networks. MR uses inherited genetic variants, such as single nucleotide polymorphisms (SNPs), to test causal links, reducing confounding and reverse causation in observational studies^{20 , 21}. Despite advances from GWAS, the causal variants and genes underlying rosacea remain unclear. This study uses multi-omics, MR, and ML to identify key genes and mechanisms in rosacea pathogenesis, offering new insights and potential therapeutic targets.

Botanical molecules, such as flavonoids, with anti-inflammatory and antioxidant properties, are increasingly used in the treatment of inflammatory skin disorders²². Our prior work showed that Vicenin-2 reduces ultraviolet (UV)-induced oxidative stress, inflammation, and apoptosis by inhibiting MAPK/MMP pathways and GSK3β expression²³,²⁴. Given its mechanisms, we hypothesized that Vicenin-2 may benefit rosacea. This study integrates multi-omics, MR, and ML to identify key genes and mechanisms in rosacea, and uses a rosacea-like mouse model to validate predicted gene expression and pathways.

MATERIALS AND METHODS

Data download

1) We retrieved the GSE65914 dataset (GPL570) from NCBI GEO, comprising 20 control and 38 disease samples. This bulk transcriptome data was used for ML, immune infiltration, and GSEA/GSVA.

2) Exposure data (eQTL): eQTL summary statistics were acquired from the eQTLGen Consortium (https://www.eqtlgen.org). This resource provides genome-wide meta-analyses of blood-based gene expression from a large, predominantly European population.

3) Exposure data (pQTL): pQTL data were sourced from the deCODE genetics repository (https://www.decode.com/summarydata/). We used the 2021 dataset (PMID: 34857953), which comprises GWAS results for 4,907 plasma protein levels measured in 35,559 individuals of European ancestry.

4) Exposure data (mQTL): mQTL data were from a previously reported article (PMID: 38548728). Our analysis specifically utilized the European cohort data, which covers 426,636 mQTL traits derived from 3,701 whole-blood samples.

5) Outcome data: Rosacea GWAS summary statistics were obtained from the EBI GWAS Catalog under accession ID GCST90476178. This dataset, primarily from individuals of European descent, includes 14,815 cases and 427,372 controls.

MR of eQTL, pQTL, and mQTL

SNPs associated with each exposure at p<1e-5 were LD-clumped (r²<0.001, 10 Mb window) and filtered for instrument strength (F>10). Causal effects on rosacea (GWAS Catalog outcome) were estimated by IVW, MR-Egger (InSIDE), weighted median, and weighted mode; with a single instrument, the Wald ratio was used. Results across estimators were synthesized to infer overall effects.

Heterogeneity test

For each instrument set, Cochran’s Q was computed as the weighted sum of squared deviations from the pooled effect. Under homogeneity, Q follows χ² with df = number of SNPs minus one. p>0.05 was interpreted as no heterogeneity.

Sensitivity analysis

Leave-one-out analysis removed each SNP in turn and recalculated the causal estimate to assess influence and robustness relative to the all-SNP model.

Feature selection with LASSO and random forest

Predictors were narrowed using LASSO logistic regression (glmnet) and a bootstrap-trained random forest. At each split, a random subset of variables was considered; importance was quantified by permutation %IncMSE. The top five features advanced to downstream work.

Immune infiltration

Bulk RNA profiles were deconvolved by CIBERSORT to estimate 22 leukocyte subsets. Gene-infiltrate relationships were examined for T-cell, B-cell, plasma-cell lineages and others.

GSEA analysis

Samples were stratified by high/low gene expression. Gene Set Enrichment Analysis (MsigDB v7.0) identified pathways with adjusted p<0.05; ranking used the normalized enrichment score (NES).

GSVA analysis

Gene Set Variation Analysis generated per-sample pathway scores from curated MSigDB sets, enabling comparison of functional signatures across specimens.

Non-coding RNA regulation

miRNA-mRNA links for key genes were predicted via miRcode and visualized in Cytoscape to outline putative post-transcriptional control.

Transcription factor (TF) regulatory network

RcisTarget inferred TFs regulating key genes. NES values derived from area under the curve recoveries of target lists per motif; motif identities were assigned by sequence evidence and similarity to reference motifs.

Molecular docking

Protein 3D models: AlphaFold (https://alphafold.com/). Small molecules: CTD (https://www.ctdbase.org/) and PubChem (https://pubchem.ncbi.nlm.nih.gov/). AutoDock used a genetic-algorithm search (50 independent runs) per protein–ligand pair; the lowest-energy pose was inspected and rendered in PyMOL.

Animal experiments

Approved by the Animal Ethics Committee of North Sichuan Medical College (NSMC2025-DW-011; ARRIVE compliant). Female BALB/c mice were randomized into six groups: PBS control, LL-37 model, LL-37 + vehicle, and three Vicenin-2 dosage groups (0.025%, 0.05%, 0.1%). Vicenin-2/vehicle was applied topically once daily for 7 days. Except controls, mice received intradermal LL-37 on days 6 and 7 (20 µl of 640 µM per injection, 12-hour interval). On day 8, erythema was quantified, and skin samples were collected for histological and molecular analyses.

RNA extraction and real-time quantitative polymerase chain reaction (RT-qPCR)

Total RNA (Trizol) was reverse-transcribed (HiScript II) and quantified by SYBR Green RT-qPCR (Vazyme). Expression was normalized to GAPDH and calculated by 2^-ΔΔCt.

Western blotting

Proteins from skin tissue were resolved by SDS-PAGE, transferred to PVDF, blocked, and probed with primary then HRP-conjugated secondary antibodies. Signals were detected by ECL and recorded on film.

Statistical analysis

MR assumptions: relevance (strong exposure–instrument link), independence (no confounder association), and exclusion restriction (no pathway to outcome other than via exposure; violations imply horizontal pleiotropy). Computations used R 4.3.0; animal data were analyzed in GraphPad Prism 10.1 (La Jolla, CA, USA). Results are mean ± SD. Group comparisons used unpaired Student’s t-tests or one-way ANOVA; two-sided p<0.05 was considered significant.

Ethics approval and consent to participate

The bioinformatics analyses in this study were based entirely on publicly available, de-identified datasets, and therefore did not require ethical approval or informed consent. The experimental animals were obtained from the Animal Experimentation Center of North Sichuan Medical College. Animal experiments were approved by the Animal Care and Use Committee of North Sichuan Medical College and meet the guidance of the Medical and Ethics Committee of North Sichuan Medical College (Sichuan) (NSMC2025-DW-011).

RESULTS

eQTL MR analysis

Rosacea outcome data were pooled under ID GCST90476178. We assessed causal effects for 493 gene-outcome pairs with significant eQTL signals. Of these, 247 genes (including KCTD3, IFIH1, PIM2, HSPA1A, and LSM1) had odds ratios (ORs)<1, indicating a protective role, while 246 genes (such as UBE2J2, SLC38A6, SURF4, SOX7, CXCR4) had OR>1, suggesting increased risk. Leave-one-out sensitivity analyses showed that removing any single SNP did not meaningfully change the pooled causal estimates or their 95% confidence intervals, indicating that no individual instrumental variable drove the associations. Additionally, no significant horizontal pleiotropy or heterogeneity was found (MR-Egger intercept p>0.05; Cochran’s Q test p>0.05), supporting the robustness and consistency of the eQTL-based MR results (Supplementary Fig. 1A).

pQTL MR analysis

pQTL data from deCODE were analyzed across 206 genes. Ninety-six genes, including SST, CRH, LRRC3, MAPK1, and FBXL4, had OR<1, indicating lower rosacea risk, while 110 genes, such as WWC1, QKI, PTPRD, ARF3, and IL2, showed OR>1, suggesting increased risk. Sensitivity analyses using the leave-one-out approach demonstrated that the causal estimates remained stable after sequential removal of individual SNPs, suggesting that the identified associations were not disproportionately influenced by any single genetic instrument. Consistently, MR-Egger regression and heterogeneity tests revealed no evidence of horizontal pleiotropy or between-instrument heterogeneity (all p>0.05), further supporting the reliability of the pQTL-based MR findings (Supplementary Fig. 1B). All 206 genes met heterogeneity criteria, and leave-one-out sensitivity analysis confirmed stability (Supplementary Fig. 1B).

mQTL MR analysis

We evaluated 1,422 mQTL-anchored genes using the IVW method (p<0.05). Of these, 672 genes, such as ZNF614, PTPN20A, PTPN20B, MIR1279, and MPST, were associated with reduced rosacea risk (OR<1), while 750 genes, including SNORD114-13, GLYCAM1, PRO0611, SNORD114-12, and OR10S1, were linked to higher risk (OR>1). Leave-one-out analyses indicated that removal of any individual SNP did not result in substantial changes in the overall effect estimates, confirming the robustness of the mQTL-derived causal inferences. Moreover, the absence of significant MR-Egger intercepts and heterogeneity (all p>0.05) suggested minimal influence of horizontal pleiotropy or confounding across instrumental variables (Supplementary Fig. 1C).

Reverse MR and ML algorithm to identify key genes

MR results from eQTL and pQTL were compared, and protective and risk gene sets were intersected using a Venn diagram. This identified one low-risk gene (RPS5) and three high-risk genes (EPHB4, SHMT1, TWSG1). Additionally, six low-risk genes, including PF4V1, MEMO1, CYB5R1, CXCL1, GSTM3, and CD300A, and two high-risk genes (UBE2M, NEO1) were identified from the intersection of mQTL and pQTL data (Supplementary Fig. 2A-D). Reverse MR analysis showed no causal effect on pQTL levels for these 12 overlapping genes (Supplementary Fig. 2E). LASSO regression and random forest models identified key genes, with LASSO selecting nine candidates (Supplementary Fig. 3A and B) and the random forest ranking five top features (Supplementary Fig. 3C). The intersection of both lists resulted in five shared genes: GSTM3, NEO1, UBE2M, CD300A, and EPHB4, identified as key targets for further study (Supplementary Fig. 3D).

Immune infiltration

The rosacea microenvironment showed elevated plasma cells and activated CD4 memory T cells, but fewer activated dendritic and NK cells (Fig. 1A-C). Key genes correlated distinctly with immune cell populations: EPHB4, GSTM3, and NEO1 were positively associated with resting dendritic and mast cells, but negatively with plasma cells and macrophages. UBE2M and CD300A showed opposing correlations with various immune cells (Fig. 1D).

Fig. 1. Immune infiltration analysis. (A) Relative percentages of immune cell subpopulations. (B) Correlation between immune cells, where blue represents negative correlation and red represents positive correlation. (C) Differences in immune cell content between control and disease samples. (D) Correlation between key genes and immune cells.

Immunoregulatory analysis

Key genes were associated with various immune modulators. EPHB4, GSTM3, and NEO1 correlated positively with CCL28 and CXCL14, but negatively with CCL4, CCL5, and several MHC and immunoinhibitor genes. Conversely, CD300A and UBE2M showed positive links to CCL18, CCL19, and specific immunostimulators (Fig. 2A-E).

Fig. 2. Relationship between key genes and immune factors. (A-E) Correlation of key genes with chemokines, immunoinhibitors, immunostimulators, MHC, and receptors.

GSEA

To delineate pathway contexts for the candidate genes, GSEA was performed. Enrichment for GSTM3 was observed in the Wnt, Hippo and transforming growth factor (TGF)-β pathways (Fig. 3A). NEO1 showed enrichment for estrogen signaling, TGF-β signaling and apoptosis (Fig. 3B). UBE2M mapped to chemokine signaling, tumor necrosis factor (TNF) signaling and the pathway annotated for Chagas disease (Fig. 3C). CD300A was linked to interleukin (IL)-17 signaling, nuclear factor (NF)-κB signaling and differentiation of Th1 and Th2 cells (Fig. 3D). EPHB4 was associated with TGF-β signaling, cAMP signaling and adherens junctions (Fig. 3E). These patterns indicate plausible routes through which the genes may influence disease progression.

Fig. 3. GSEA analysis of key genes. (A-E) KEGG signaling pathways involved with key genes, including pathway regulation and associated genes.

GSVA

GSVA yielded consistent signals. GSTM3 scored highly for ESTROGEN_RESPONSE_LATE and PANCREAS_BETA_CELLS (Fig. 4A). NEO1 aligned with ESTROGEN_RESPONSE_LATE and UV_RESPONSE_DN (Fig. 4B). UBE2M was positively associated with XENOBIOTIC_METABOLISM and OXIDATIVE_PHOSPHORYLATION (Fig. 4C). CD300A correlated with INTERFERON_GAMMA_RESPONSE and ALLOGRAFT_REJECTION (Fig. 4D). EPHB4 was linked to UNFOLDED_PROTEIN_RESPONSE and MITOTIC_SPINDLE (Fig. 4E). Taken together, these gene set activities support the view that the key genes may modulate disease progression through the above biological.

Fig. 4. GSVA analysis of key genes. (A-E) GSVA analysis of key genes, with blue representing high-expression genes and the corresponding signaling pathways, and green representing low-expression genes and related pathways. The background gene set is “hallmark.”.

Non-coding RNA network and TF regulation

Upstream regulators of the five key genes were inferred via miRcode, identifying 73 miRNAs and 151 interactions (Fig. 5A). TF enrichment analysis using cumulative recovery curves highlighted cisbp_M0684 as the top motif (NES=5.21), with the complete TF-motif network detailed in Fig. 5B and C.

Fig. 5. miRNA and transcriptional regulatory network of key genes. (A) miRNA network of key genes, with yellow representing mRNA and purple representing miRNA. (B) Transcriptional regulatory network of key genes, where red represents key genes and green represents transcription factors. (C) Display of all motifs enriched by key genes and their corresponding transcription factors.

NES: normalized enrichment score, AUC: area under the curve, TF: transcription factor.

Molecular docking

We evaluated three protein-ligand pairs: CD300A with Q9UGN4-bisphenol A, EPHB4 with P54760-bisphenol A, and GSTM3 with P21266-Niacinamide. The predicted binding energies were as follows: CD300A with Q9UGN4-bisphenol A, -6.81 kcal/mol; EPHB4 with P54760-bisphenol A, -4.73 kcal/mol; GSTM3 with P21266-Niacinamide, −4.00 kcal/mol (Supplementary Fig. 4A-C).

Animal experiments

We tested the hypothesis in an LL-37-induced rosacea-like mouse model (Methods). Model mice developed marked erythema, which topical Vicenin-2 reduced dose-dependently (Fig. 6A). ImageJ-based erythema area and scores corroborated improvement (Fig. 6B); procedures are in Supplementary Tables 1 and 2. RT-qPCR showed downregulated low-risk and upregulated high-risk genes in the model, consistent with predictions (Fig. 6C); raw data and curves are in Supplementary Tables 3 and Supplementary Fig. 5. Western blots showed elevated IL-17/MAPK pathway proteins in the model, dose-dependently suppressed by Vicenin-2, most in the high-dose group (Fig. 6D); original bands are in Supplementary Fig. 6. Statistics confirmed reductions in IL-17RA, ACT1, TRAF6, p-TAK1/TAK1, and p-P38/P38 with Vicenin-2 (Fig. 6E-I). These findings suggest Vicenin-2 mitigates rosacea-associated inflammation by modulating IL-17/MAPK signaling.

Fig. 6. Animal experiment results. (A) Representative images of erythema on mouse backs from different treatment groups. (B) Quantification of erythema scores (n=5 per group). (C) Quantification of erythema area (n=5 per group). (D) Real-time quantitative polymerase chain reaction analysis of relative mRNA expression levels of key genes (GSTM3, CD300A, NEO1, UBE2M, and EPHB4) in normal and model groups. (E) Western blot analysis of key proteins in the IL-17/p38 MAPK signaling pathway in skin samples from different treatment groups. (F-J) Statistical analysis of relative expression levels of IL-17RA, ACT1, TRAF6, p-TAK1/TAK1, and p-P38/P38, with GAPDH as the internal control. Statistical significance was assessed using one-way ANOVA followed by appropriate post hoc tests.

ns, no significant difference.

^*p<0.05; ^**p<0.01; ^***p<0.001; ^****p<0.0001.

DISCUSSION

Rosacea’s relapsing nature underscores the need for precision medicine²⁵,²⁶. MR analysis uses genetic variants to estimate causal effects, reducing confounding and reverse causation bias²⁷. Unlike previous MR studies focusing on lifestyle or systemic commodities²⁸,²⁹. Our multi-layer MR framework integrated eQTL, pQTL, and mQTL data to identify hundreds of variants. Through rigorous sensitivity testing and ML (LASSO/random forest), we identified five key genes (GSTM3, CD300A, NEO1, UBE2M, and EPHB4). This causal-to-mechanistic approach illustrates how expanding GWAS resources can refine therapeutic targets and inform functional research in rosacea.

Pinpointing pivotal genes can inform improved strategies for rosacea care. In GWAS, Chang and colleagues detected risk variants in HLA-DRA and BTNL2, suggesting these loci as therapeutic targets to lessen inflammatory features³⁰. Additionally, Buhl et al reported upregulated expression of CXCL8 and IL-18 in rosacea, closely linked to inflammation and angiogenesis, providing potential molecular targets for therapy³¹. This study, through an integrated analysis of MR, ML, and functional annotation, identified five key genes (GSTM3, CD300A, NEO1, UBE2M, EPHB4) and their associated signaling pathways closely linked to rosacea pathogenesis, offering novel avenues for optimizing disease management.

GSTM3, an antioxidant gene, shows increased expression in rosacea lesions, likely as a compensatory mechanism against oxidative stress. This aligns with GWAS studies linking antioxidant pathways to disease susceptibility³⁰,³². NEO1, a regulator of blood vessel stability, may be crucial in countering the persistent erythema and telangiectasia characteristic of rosacea³³. Conversely, EPHB4 is vital for proper vascular formation, and its dysregulation could directly contribute to the abnormal vessel morphology seen in patients³⁴. The gene UBE2M amplifies inflammation through the NF-κB pathway, promoting the release of key cytokines like IL-6 and CXCL8, which are known drivers of rosacea³⁵. While the role of the immune receptor CD300A in rosacea is not yet defined, its known function in modulating neutrophil and dendritic cell activity makes it a compelling candidate for future study³⁶. Although its role in rosacea remains uncharacterized, its immunoregulatory function warrants further study. Expression of these genes was confirmed in rosacea-like murine skin by RT-qPCR, consistent with the bioinformatic findings. GSEA and GSVA analyses revealed enrichment in IL-17, NF-κB, TGF-β, Wnt, Hippo, chemokine, TNF, estrogen, and cAMP pathways, highlighting their collaborative roles in inflammation and angiogenesis. Wnt and Hippo pathways regulate keratinocyte proliferation and tissue repair³⁷. TGF-β/Smad2/3 signaling upregulates VEGF and promotes both fibrosis and neovascularization³⁸. Other signaling cascades involving TNF-induced NF-κB activation, chemokine-driven leukocyte adhesion, oxidative phosphorylation, and the unfolded protein response reflect metabolic and proteostatic stress in rosacea skin³⁹,⁴⁰. Immune cell infiltration, particularly involving macrophages, is a critical component of rosacea’s pathological landscape. Single-cell datasets and bioinformatic evaluation indicated greater proportions of M0 and M1 macrophages, gamma-delta T cells, and activated CD4+ memory T cells, accompanied by fewer regulatory T cells and fewer dendritic and NK cells in lesions⁴¹,⁴². Functional studies indicate that M1 macrophage polarization exacerbates inflammation, while macrophage depletion reduces rosacea-like skin inflammation in mouse models⁴³. Moreover, keratinocyte TLR2 activation drives CXCL8-mediated recruitment of neutrophils and macrophages⁴⁴. Our findings align with prior reports: increased plasma cells, activated CD4+ memory T cells, and gamma-delta T cells, with reductions in regulatory T cells, dendritic cells, and NK cells. Importantly, CD300A and UBE2M correlated positively with inflammatory immune subsets, whereas GSTM3, NEO1, and EPHB4 were inversely associated with regulatory checkpoints such as CXCR4 and IDO1. This suggests that key genes may actively modulate immune cell recruitment and function. Moreover, single-cell findings of heterogeneous immune populations, including interferon-stimulated macrophage subsets, support our results. In sum, immune infiltration in rosacea reflects a complex, gene-regulated network driving inflammation. Understanding these gene-immune interactions could inform precision immunomodulatory strategies targeting specific immune cell subsets or checkpoints like PD-L1 and chemokine receptors.

Molecular docking, a computational method, supports drug screening and target prediction. In the early stage of this study, we used the CTD database to identify multiple potential compounds targeting five key genes (GSTM3, NEO1, UBE2M, CD300A, and EPHB4). This screening highlighted two representative compounds: bisphenol A (BPA) and nicotinamide. The docking results showed that BPA and nicotinamide exhibited strong binding affinity for several key targets (binding energies <−4.0 kcal/mol). BPA, an endocrine disruptor, is linked to cutaneous inflammation via immune targets and PI3K-AKT signaling⁴⁵ and can penetrate skin to sustain local immune activation⁴⁶. Niacinamide (the vitamin B3 amide) is used in dermatology for anti-inflammatory, antioxidant, and barrier support; clinical studies report improvements in erythema and papules in rosacea⁴⁷. Our docking suggests BPA may interact with high-risk genes to disturb immune balance and promote inflammation, whereas niacinamide may act on low-risk genes to modulate anti-inflammatory pathways. However, previous studies have suggested that BPA may be involved in the pathogenesis of rosacea, although a small prospective study found no difference in serum BPA levels between rosacea patients and healthy controls⁴⁸. In contrast, the beneficial effects of nicotinamide in alleviating rosacea have already been demonstrated in clinical studies⁴⁹.

Vicenin-2, a naturally derived flavonoid, exhibits anti-inflammatory, antioxidant, antitumor, and photoprotective properties²³. Recent studies highlight its therapeutic potential in skin diseases, particularly photoaging and inflammatory dermatoses. By acting on the GSK3β-PI3K/Akt axis and AMPK pathway, Vicenin-2 lessens UV-driven skin damage, dampens inflammation and apoptosis linked to photoaging, and curbs influx of M1 macrophages and CD4+ memory T cells²⁴,⁵⁰. Given the partial overlap in the pathogenic mechanisms of rosacea and photoaging⁵¹,⁵², we hypothesized that Vicenin-2 may have therapeutic potential for rosacea. Based on this rationale, we selected Vicenin-2, a novel natural compound, as the focus of subsequent experimental validation in this study. Extensive evidence links the IL-17 pathway to rosacea. Interleukin 17A, primarily from Th17 cells, binds IL-17RA/RC to recruit ACT1 and TRAF6, triggering TAK1 and downstream NF κB and MAPK pathways. This induces IL-6, IL-8, and CXCL1 expression and neutrophil migration. Rosacea lesions show elevated IL-17 together with IL-17RA, and IL-17A blockade mitigates inflammatory and fibrotic changes in a corresponding murine model⁵³,⁵⁴. As one of the major signaling pathways enriched for key genes, our study focused on IL-17 signaling transduction and further validated downstream MAPK signaling pathway activation. Our experimental results showed that topical Vicenin-2 solution significantly alleviated LL-37-induced rosacea-like erythema in mice, possibly by dose-dependently inhibiting IL-17/MAPK signaling. In addition, IL-17RA may activate the p38 MAPK pathway through the ACT1-TRAF6-TAK1 signaling axis, thereby contributing to the local inflammatory cascade in rosacea. These findings suggest Vicenin-2 as a promising candidate for rosacea treatment. Despite supportive preclinical evidence, proof of benefit and risk will depend on adequately powered clinical studies. Future research should further elucidate its molecular targets and therapeutic mechanisms to support its development as a targeted natural therapy for rosacea and related skin disorders.

Nonetheless, this study has certain limitations. First, the public datasets used did not distinguish among clinical subtypes of rosacea, limiting our ability to explore subtype-specific mechanisms. Second, our analyses were restricted to populations of European ancestry, without accounting for potential ethnic differences. Future studies will incorporate publicly available datasets from other ethnic groups, such as East Asian populations, and collect multi-ethnic clinical samples to validate differential expression of key genes and improve the generalizability of our findings. In addition, immune cell infiltration was inferred from deconvolution of public bulk transcriptomic data and therefore represents associative evidence. Experimental validation using flow cytometry and immunohistochemistry will be conducted to directly assess immune cell infiltration and cell-specific expression of key genes in clinical samples. Finally, although Vicenin-2 showed preliminary efficacy in rosacea, the lack of a positive control limited comparison with established therapies; future in vivo and in vitro studies with improved experimental designs are warranted.

In conclusion, this study provides the first systematic identification of five core rosacea-related genes (GSTM3, NEO1, UBE2M, CD300A, EPHB4). By integrating MR with ML (LASSO/Random Forest) and immune infiltration analysis, we linked these targets to plasma cell/Treg dysregulation and metabolic pathways. Furthermore, we identified Vicenin-2 as a novel, promising immune modulator for rosacea treatment.

SUPPLEMENTARY MATERIALS

Supplementary Table 1

Raw erythema area measurements for animal model evaluation (cm²)

Supplementary Table 2

Raw erythema score for animal model evaluation

Supplementary Table 3

Raw RT-qPCR data showing relative expression of GSTM3, CD300A, NEO1, UBE2M, and EPHB4 between control and model group

Supplementary Fig. 1

Mendelian Randomization of eQTL, pQTL and mQTL. (A) The distribution of causal relationship risk ratios and p-values obtained from eQTL Mendelian randomization. (B) The distribution of causal relationship risk ratios and p-values obtained from pQTL Mendelian randomization. (C) The distribution of causal relationship risk ratios and p-values obtained from mQTL Mendelian randomization.

Supplementary Fig. 2

Reverse MR. (A-D) Venn diagram of the intersection of mQTL, eQTL, and pQTL associated genes. (E) Reverse MR analysis reveals 12 causal relationships.

Supplementary Fig. 3

Construction of Lasso and random forest models. (A) Distribution of Lasso coefficients and the gene combination at the minimum lambda value. (B) Ten-fold cross-validation for selecting the tuning parameter to determine the minimum lambda value in the Lasso model. (C) Display of feature genes from the random forest model. (D) Intersection of feature genes selected by Lasso regression and random forest algorithms.

Supplementary Fig. 4

Molecular docking network. (A-C) Interaction between key genes and molecules. Each node represents a molecule, and the connecting lines between nodes indicate docking or interactions between molecules.

Supplementary Fig. 5

Real-time quantitative polymerase chain reaction amplification and melting curves of representative genes.

Supplementary Fig. 6

Original uncropped WB images for Fig. 6E.