Causal Effects of Diet on Atopic Dermatitis: A Mendelian Randomization Study Implicating Lipid Pathways and Clinical Implications

Introduction

Atopic dermatitis (AD) is a chronic inflammatory skin disorder of increasing global prevalence, affecting up to 20% of children and 10% of adults, and represents a major contributor to the global burden of skin diseases.^1,2 It severely impairs quality of life and imposes substantial socioeconomic costs.^3,4 Diet is a key modifiable factor that influences systemic lipid metabolism, which in turn regulates immune responses and skin barrier function.^5,6 However, causal evidence directly linking holistic dietary patterns to AD via lipid intermediates remains limited.^7,8

Mechanistic studies indicate that obesogenic diets disrupt lipid homeostasis and fuel inflammation.^9,10 Dysregulated lipid metabolism is a hallmark of various inflammatory skin diseases. For instance, elevated sphingolipids in psoriasis¹¹ and aberrant sebaceous lipid synthesis in rosacea¹² underscore the broader relevance of lipid pathways in cutaneous inflammation. Key lipids such as sphingomyelins and free fatty acids—implicated in these conditions—modulate immune activation and barrier integrity,¹³ and clinical observations link dyslipidemia to AD severity.^14,15 Nonetheless, prior research has predominantly focused on isolated nutrients (eg., ω-3 fatty acids), overlooking the complex effects of dietary patterns. Furthermore, establishing a causal diet-AD pathway presents a high-dimensional computational challenge, requiring modeling of complex exposure-mediator networks while controlling for confounding.

Mendelian randomization (MR) has emerged as a powerful tool for causal inference by using genetic variants as instrumental variables to mitigate confounding.^16,17 While MR has linked specific lipids (eg., LDL-C, PUFAs) to AD risk,^18,19 a critical gap persists: no study has systematically investigated the broad lipid metabolome as a mediator between dietary patterns and AD. Applying MR to this high-dimensional mediation context also requires robust methods to handle multiple testing and potential biases like horizontal pleiotropy.

To address this gap, we implement a two-step, high-dimensional MR framework. Using genetic instruments for high-fat/high-sugar dietary patterns. Our study aims to causally identify key lipid mediators, moving beyond associations to extract actionable mechanistic insights for AD prevention.

Methods

Study Design

This study adopted a two-step MR framework to investigate how genetically influenced dietary habits affect AD risk through circulating lipid metabolism, using genome-wide association study (GWAS) summary data (Figure 1). In the first step, we evaluated the cause-and-effect link between dietary behaviors (including high-fat and high-sugar intake, and ω-3 fatty acids) and circulating lipids (cholesterol, triglycerides, and lipoprotein components). During the second stage, we evaluated the causal effect of lipid profiles on AD risk and quantified the mediation proportion using the product method. To confirm the validity and independence of the IVs, we employed rigorous genetic screening criteria, including linkage disequilibrium (LD) clumping and F-statistics, which help minimize confounding. Ethical review for this study was formally waived by the Ethics Committee of Shenzhen Hospital (Fu Tian) of Guangzhou University of Chinese Medicine. This decision was based on the use of de-identified, public GWAS summary data and is in accordance with Article 32, Items 1/2 of the “Measures for Ethical Review of Life Science and Medical Research Involving Human Subjects” (2023, China). An official exemption certificate has been provided as Supplementary Material. All original GWASs obtained ethical approval and participant consent in their primary studies.

Figure 1 Causal pathway from dietary habits through lipid metabolism to AD risk. Bold text indicates main sections (Datasets, Data harmonization and IV selection, MR analyses, Sensitivity tests) and key methods (IVW method, MR-Egger method, Cochran’s Q test). Abbreviations: GWAS, genome-wide association study; GLGC, Global Lipid Genetics Consortium; IV, instrumental variable; MR, Mendelian randomization.

Data Sources and Variable Definitions

Genetic instruments for 83 dietary habits were derived from the UK Biobank study, with summary statistics via the GWAS Catalog (https://www.ebi.ac.uk/gwas/).²⁰ Instrumental variables were selected following a standardized procedure, including single-nucleotide polymorphism (SNP) filtering (P < 5×10⁻⁵), linkage disequilibrium (LD) pruning (r² < 0.001 within 10 Mb), data normalization, and bias correction using the TwoSampleMR R package. Circulating lipid profiles, serving as mediator variables, comprised 241 serum lipid markers sourced from the following:

1. Omicscience database: provided 139 lipid species (eg., triglycerides, sphingomyelins) (N = 9,363).²¹
2. Johannes et al: provided 98 lipoprotein subfractions, including Very Low-Density Lipoprotein/Low-Density Lipoprotein (VLDL/LDL) particle size distributions (N = 24,925).²²
3. Global Lipids Genetics Consortium (GLGC): provided conventional lipid measures, including total cholesterol (TC), triglycerides (TG), low-density lipoprotein cholesterol (LDL-C), and high-density lipoprotein cholesterol (HDL-C) (N = 1,320,016).²³

Outcome data for atopic dermatitis were obtained from the IEU Open GWAS platform (ID: ebi-a-GCST90027161), based on the FinnGen R10 cohort. Cases were defined according to ICD-10 code L20 and further verified through dermatologist-reviewed electronic medical records. The total sample size of 796,661 individuals (22,474 cases).

All datasets were confined to individuals of European ancestry to minimize bias from population stratification. Detailed characteristics of the data sources are provided in Table 1 (accessed March 25, 2025).

Table 1 Summary of GWAS Sources, Sample Sizes, and Traits for Exposure, Mediator, and Outcome Variables

Dietary Exposure Assessment

The key exposure variable “Other oil-based spread” was derived from the UK Biobank touchscreen questionnaire item “Spread type” (Field ID 1428; https://biobank.ndph.ox.ac.uk/showcase/field.cgi?id=1428). This variable represents a mutually exclusive category that includes generic, non-branded spreads predominantly composed of vegetable oils (eg., sunflower, rapeseed, or olive oil blends), distinct from butter, branded functional margarines (eg., Flora Pro-activ/Benecol), or low-fat spreads. This classification aligns with nutritional surveillance studies of UK spread consumption and prior methodological work utilizing UK Biobank dietary data.^20,24

Construction and Assumption Validation of Instrumental Variables

Instrumental variables (IVs) were constructed following the core assumptions of Mendelian randomization. The relevance assumption required all SNPs to show significant associations with the exposure variable—dietary traits or lipid phenotypes—at a threshold of P < 1×10⁻⁵, and an F-statistic >10 to minimize weak instrument bias. The independence assumption mandated that the selected SNPs be independent of known confounders, verified using PhenoScanner v2 with the same significance threshold (P < 1×10⁻⁵). The exclusion restriction assumption stipulated that SNP effects on the outcome should occur solely through the exposure, excluding horizontal pleiotropic effects.

SNP selection involved LD clumping with an r² < 0.001 within a 10,000 kb window. SNPs exhibiting an allele frequency (MAF) ≤0.01 were excluded, and palindromic SNPs with MAF values ranging from 0.42 to 0.58 were removed to prevent strand ambiguity. Additionally, SNPs that were directly linked to the outcome (P < 1×10⁻⁵) were discarded. Pleiotropy was controlled using the MR-PRESSO global test (P < 0.05) to identify and exclude outlier SNPs. Detailed methodological steps, including specific tools used for SNP selection and bias correction, are described in the Supplementary Methods section.

Mendelian Randomization Analysis

To investigate the potential mediating role of lipid traits (mediator) in the causal pathway from dietary habits (exposure) to AD (outcome), We employed a two-stage MR design. In the first stage, the effects of dietary habits on 241 lipid traits were estimated, followed by evaluation of the impact of these lipid traits on AD in the second stage. The primary statistical method used was inverse-variance weighted (IVW) regression. Sensitivity analyses included MR-Egger, weighted median, weighted mode, and simple mode approaches, along with leave-one-out analyses. Heterogeneity was assessed via Cochran’s Q test and I² statistic, and horizontal pleiotropy was evaluated using the MR-Egger intercept. Multivariable MR (MVMR) was applied if I² statistic >50% to account for potential confounding.

Mediation Effect Assessment

Mediation effects were assessed within the two-stage MR framework using the product-of-coefficients method. For each lipid trait, the indirect effect (β_mediation) was calculated as the product of β₁ (dietary habit → lipid trait) and β₂ (lipid trait → AD), and the total effect of dietary habits on AD was denoted β₃. The proportion mediated was calculated as: A mediation effect was considered significant if (1) the indirect and total effects were in the same direction (β_mediation > 0) and (2) P < 0.05 with an absolute proportion mediated >5%, minimizing false positives from weak mediation.

Reverse MR Analysis

We conducted a reverse MR analysis to examine the direction of causality. This tested whether lipid levels influenced dietary habits (lipid trait → dietary habits) or whether AD affected lipid levels (AD → lipid trait). The “diet → lipids → AD” causal pathway was considered more plausible if effect estimates for all reverse directions were statistically non-significant (P > 0.05).

Software Implementation

The analyses were conducted with R software (version 4.1.0), utilizing key packages such as “TwoSampleMR” (v0.5.6), “MRPRESSO” (v1.0), and “MendelianRandomization” (v0.5.1). Figures were generated using “ggplot2” (v3.3.5) and “forestplot” (v2.0.1). The tests were two-sided, and a P value under 0.05 was regarded as statistically significant.

Reporting Guidelines

This study was reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology using Mendelian Randomization (STROBE-MR) guidelines. The completed checklist is provided as Supplementary Material.

Results

SNP Selection

Following cluster analysis and application of exclusion criteria, we selected three SNP sets: 6,290 SNPs associated with dietary habits (Supplementary Table 1-1), 7,703 SNPs associated with both lipid traits and AD (Supplementary Table 1-2), and 11,575 SNPs associated with both dietary habits and lipid traits (Supplementary Table 1-3). All GWAS data for exposures, mediators, and outcomes were obtained from independent cohorts with no sample overlap.

Direct Causal Effects of Dietary Habits on AD

Based on IVs for 83 dietary habits, IVW analysis identified five dietary habits significantly associated with AD risk (P < 0.05; Table 2 and Figure 2). High intake of brown bread (OR = 1.78, 95% CI 1.10–2.89) and wholemeal bread (OR = 1.23, 95% CI 1.06–1.43) increased AD risk, whereas muesli (OR = 0.814, 95% CI 0.67–0.99), lamb or mutton (OR = 0.80, 95% CI 0.67–0.95), and other oil-based spread (OR = 0.56, 95% CI 0.34–0.93) was protective. Heterogeneity was low (Cochran’s Q P > 0.05; I² < 25%) and no horizontal pleiotropy was detected. Sensitivity analyses (leave-one-out, funnel, scatter plots) revealed no influential outliers, supported by Supplementary Dataset 1.

Table 2 IVW Estimates of Dietary Exposures on AD Risk

Figure 2 MR forest plot: Causal effects of dietary exposures on AD risk.

Two-Stage MR Analysis of Lipid Traits Mediating the Diet–AD Relationship

To explore whether lipid traits mediate the relationship between dietary patterns and AD, we performed a two-stage MR analysis.

Causal Relationship Between Lipid Traits and AD

In the first stage, we assessed 241 lipid traits as candidate mediators and examined their causal associations with AD.

The IVW method identified 37 lipid traits significantly associated with AD risk (P < 0.05), 28 of which remained consistent after sensitivity analyses (Table 3 and Figure 3). Fourteen traits were positively associated (eg., SM C20:2, OR = 1.13), while 14 were associated with a reduced risk (eg., phosphatidylcholine aa C36:4, OR = 0.96). Omega-6 fatty acids emerged as a significant risk factor (OR = 1.06, 95% CI 1.01–1.11, P = 0.03). Heterogeneity was low (I² = 12%, Cochran’s Q = 32.97, P = 0.28) and no horizontal pleiotropy was observed (MR-PRESSO P = 0.26). Sensitivity analyses (leave-one-out, funnel, scatter plots) revealed no influential outliers, supported by Supplementary Dataset 2.

Table 3 IVW Estimates of Lipid Traits on AD Risk

Figure 3 MR forest plot: Causal effects of Lipid Traits on AD risk.

Causal Relationships Between Dietary Habits and Lipid Traits

In the initial stage of the two-sample Mendelian randomization analysis, 28 lipid traits demonstrated significant associations with AD. Using the IVW method, we identified four significant causal relationships between dietary exposures and lipid traits.

Specifically, consumption of other oil-based spreads was negatively associated with sphingomyelin SM C20:2 levels (OR = 0.47, 95% CI 0.25–0.87, P = 0.02). Brown bread intake exhibited negative associations with both the concentration of very large VLDL particles (OR = 0.56, 95% CI 0.37–0.85, P = 0.01) and total lipids in large VLDL (OR = 0.63, 95% CI 0.42–0.94, P = 0.03). Muesli intake was also negatively associated with phosphatidylcholine PC aa C36:0 levels (OR = 0.73, 95% CI 0.56–0.95, P = 0.02) (Figure 4 and Table 4).

Table 4 IVW Estimates of Dietary Exposures on Lipid Traits

Figure 4 MR forest plot: Dietary exposures → Lipid traits.

No significant heterogeneity (I² = 0–7%, Cochran’s Q test P > 0.35) or horizontal pleiotropy (MR-Egger intercept test P > 0.18; MR-PRESSO global test P > 0.43) was detected. Sensitivity analyses (leave-one-out, funnel, scatter plots) revealed no influential outliers, supported by Supplementary Dataset 3.

Mediating Effect of Dietary Habits on AD and Validation of Causal Direction

Following the two-stage MR mediation framework outlined in Section 2.5, we applied the product method to identify significant diet–lipid–AD pathways. Mediation was deemed significant if (1) the indirect and total effects were in the same direction (β_mediation > 0) and (2) P < 0.05 with an absolute proportion mediated >5%.

The total effect of “other oil-based spread” on AD was β₃ = −0.54 (P < 0.01), with 15.9% mediated by SM C20:2 (β_mediation = −0.091, P < 0.01). For “brown bread”, a total effect of β₃ = 0.58 (P < 0.05) was partially mediated by VLDL particles, (8.9%; β_mediation = 0.05, P < 0.01). “muesli” showed a negative mediation via PC aa C36:0, accounting for −13.6% of the total effect, indicating that this lipid may attenuate the protective effect of muesli (Table 5).

Table 5 Mediation Analysis of Dietary Exposures on AD Risk via Lipid Traits

All instrumental variables were robust (F-statistic > 10), with no evidence of heterogeneity or pleiotropy detected. Figure 5 illustrates the causal pathways through which dietary habits influence AD risk via specific lipid profiles (Spread → SM C20:2, Bread → VLDL traits, Cereal → PC aa C36:0), with key mediation statistics—β_mediation, proportion mediated, and P value—provided. To validate the causal direction, reverse MR analysis was conducted. No significant causal effects of AD on either dietary habits or lipid traits were observed (all P > 0.05; Supplementary Table 2), supporting the proposed directionality in the primary model.

Figure 5 Mediation pathways from dietary habits to atopic dermatitis. Abbreviations: L-VLDL, large very-low-density lipoprotein; PC, phosphatidylcholine; TL, total lipids; VL-VLDL-P, very large very-low-density lipoprotein particles.

Discussion

Employing a two-step Mendelian randomization framework, this study establishes causal links between dietary patterns and AD through distinct pathways of food processing and lipid metabolism. Our genetic analysis revealed that ultra-processed brown bread (OR = 1.78) and wholemeal bread (OR = 1.23) increased AD risk, whereas minimally processed muesli, lamb, and other oil-based spreads conferred protection. By providing novel genetic evidence and mechanistic clarification, these findings advance nutritional epidemiology. They not only support the known principle that processed grains elevate systemic inflammation while nutrient-dense foods attenuate it,^25,26 but further identify specific lipid mediators—such as VLDL particles and sphingomyelins—which underlie these causal relationships.

Our MR results both align with and extend prior evidence. The protective association of unsaturated fat-rich spreads aligns with dietary guidelines and observational studies linking n-3 PUFAs to lower inflammation. However, our finding that genetic predisposition to higher brown and wholemeal bread intake increases AD risk appears counterintuitive and contrasts with some epidemiological studies promoting whole grains. This discrepancy may be reconciled by considering food processing. Most MR and large cohort studies classify “whole grain” as a beneficial category. Our instrument, derived from UK Biobank’s specific “bread type” questions, likely captures consumption of commercially produced breads, which even when labeled “brown” or “wholemeal,” can undergo significant processing, containing additives, high glycemic index flours, and added sugars—factors linked to inflammation. This distinction highlights a key novelty of our work: moving beyond broad food groups to reveal how commercial processing may alter the health effects of nominally healthy foods, a nuance crucial for dermatological nutrition. This interpretation is reinforced by the mediation analysis. The link between brown bread and AD was partly mediated (8.9%) by VLDL traits. This likely reflects the pro-inflammatory impact of processing-associated components like added sugars and high glycemic index flours—which promote hepatic VLDL secretion—rather than the intrinsic properties of whole-grain fiber.

Mechanistically, food processing alters the food matrix, affecting bioactive components and gut microbiota. High-temperature baking and mechanical grinding degrade compounds, such as β-glucan and polyphenols, promote pro-inflammatory bacteria (eg., Prevotella spp.), and reduce short-chain fatty acid (SCFA) production.^27,28 These alterations activate TLR4/NF-κB signaling via LPS, downregulate Claudin-1, increase intestinal permeability, and impair skin barrier function.^29,30 In contrast, minimally processed whole grains and n-3 polyunsaturated fatty acids (n-3 PUFAs) in lamb enrich beneficial gut microbiota, elevate SCFAs, suppress NF-κB activation, and reinforce skin barrier integrity.^31,32 Specifically for highly processed grains, this gut dysbiosis and increased permeability can stimulate hepatic secretion of large, triglyceride-rich VLDL particles. These VLDL particles, in turn, can activate inflammatory pathways such as TLR2/4-NF-κB, releasing IL-6 and IL-1β, thereby creating a direct “gut-liver-skin” axis that propagates dietary inflammation to AD pathogenesis.¹⁹

These findings outline a mechanistic model where the gut-skin axis and neuro-skin axis jointly regulate AD pathogenesis. Intestinal pro-inflammatory mediators, such as IL-6, activate vagal afferent fibers, prompting sensory neurons to release calcitonin gene-related peptide (CGRP) and substance P (SP).³³ These neuropeptides stimulate keratinocytes to secrete thymic stromal lymphopoietin (TSLP), which drives a Th2/Th22-dominant immune response and perpetuates the itch–scratch cycle. This integrated cascade illustrates how dietary modulation of lipid metabolism can influence AD progression, highlighting the combined systemic and neural contributions to disease pathophysiology.

Beyond the damaging VLDL pathway, mediation analysis revealed protective and antagonistic lipid mechanisms.

1. Sphingomyelin-PPARγ Protective Pathway: The observed protective effect of “Other oil-based spread” (OR = 0.56) likely reflects the predominance of unsaturated fats in generic vegetable oil spreads (eg., sunflower, rapeseed, or olive oil blends), as opposed to the saturated fats in butter, aligning with dietary guidelines to mitigate inflammation. Mechanistically, these unsaturated fats promote PPARγ activation and upregulate barrier-associated proteins, attenuating inflammation and reducing AD risk.³⁴ This pathway mediated 15.9% of the protective effect (β = −0.09, P = 0.003), underscoring the role of optimized lipid composition in barrier repair.

2. Phosphatidylcholine Antagonistic Network: Conversely, certain lipids such as oat-derived phosphatidylcholine PC aa C36:0 may partially counteract protective effects by modulating keratinocyte metabolism and immune responses,³⁴ revealing complex interactions within the lipid network. Collectively, these pathways delineate a multi-faceted “diet-lipid metabolism-AD” axis.

In summary, this study establishes lipid metabolism as a central mediator linking diet to atopic dermatitis, delineating specific pro-inflammatory (VLDL particles), protective (sphingomyelin-PPARγ), and antagonistic (phosphatidylcholine) pathways within the “diet→lipid metabolism→AD” axis. This framework provides a mechanistic basis for dietary interventions: reducing highly processed grains to lower inflammatory triggers, while increasing intake of n-3 PUFA-rich foods and minimally processed whole grains to support skin barrier function.

The key strength of this work is its two-step Mendelian randomization mediation design, providing novel genetic evidence for these causal pathways beyond observational associations. Limitations include potential residual pleiotropy, limited generalizability from European-centric data, and an inability to capture life-stage-specific dietary effects. Future studies should validate these findings in trans-ethnic populations and employ more complex models to further elucidate dietary impacts on AD.

Conclusion

This Mendelian randomization study establishes lipid metabolism as a key mediator linking dietary factors to atopic dermatitis. A lipid regulatory network, including ceramide homeostasis, pro-inflammatory VLDL signaling, and barrier-protective sphingolipid pathways, underlies this relationship. The identified lipid species represent promising candidate biomarkers that mechanistically link diet to AD pathogenesis. Prioritizing these lipid pathways in future epidemiological studies could pave the way for validating targeted nutritional strategies.