Immune–metabolic mediation of depression risk: Causal evidence from antibody and metabolite GWAS

Abstract

Depression is a major contributor to global disability, yet its underlying biological mechanisms remain incompletely understood. Immune activation and metabolic alterations, particularly involving lipid metabolism, have been implicated, but their roles in depression remain unclear. We used large-scale genome-wide association study (GWAS) summary statistics to examine the relationships among antibody responses, plasma metabolites, and depression. Mendelian randomization (MR) was applied to evaluate genetically predicted associations of antibody traits and metabolites with depression risk. Generalized summary-data-based Mendelian randomization (GSMR) was used as a complementary analysis, and two-step mediation analyses were performed to assess whether selected metabolites were compatible with mediating immune-related effects on depression. Higher Epstein-Barr virus (EBV) ZEBRA antibody levels were associated with increased depression risk, whereas arachidonate-enriched metabolites tended to show inverse associations with depression and several linoleoyl-related metabolites showed positive associations. Mediation analyses suggested that 1,2-dilinoleoyl-GPC showed a positive indirect effect, whereas 1-arachidonylglycerol showed a negative indirect effect. Two glycerolipid ratios also showed indirect effects consistent with partial mediation. Complementary GSMR analyses showed directionally concordant results for the principal associations. Overall, these findings support an immunometabolic interpretation in which EBV-related immune responses and plasma metabolite remodeling may be relevant to depression vulnerability, and they prioritize candidate metabolites and pathways for future mechanistic and clinical investigation.

Introduction

Depression is among the most prevalent and disabling psychiatric disorders, affecting hundreds of millions worldwide and contributing substantially to the global burden of disease [1, 2]. It is characterized by persistent low mood, cognitive and emotional disturbances, and abnormalities in sleep and appetite, often resulting in marked functional impairment and increased mortality [3]. Epidemiological surveys estimate lifetime prevalence rates of approximately 20% in women and 12% in men, with depression remaining a leading cause of suicide worldwide [4–6]. Beyond its psychiatric manifestations, depression is also associated with an increased risk of medical comorbidities, including cardiovascular disease, autoimmune disorders, diabetes, and certain cancers, further amplifying its public health impact [7–9]. Although pharmacological and psychotherapeutic treatments are available, therapeutic responses remain heterogeneous and a substantial proportion of patients do not achieve remission [10, 11]. These challenges highlight the need to better understand the biological mechanisms underlying depression and to identify biomarkers that may support earlier detection, improved stratification, and more precise intervention.

A growing body of evidence implicates immune dysregulation and metabolic disturbance in the pathophysiology of depression. In particular, immune responses to infectious agents may provide a biologically informative window into long-term host–pathogen interactions, chronic immune activation, and downstream neuroimmune signaling [12, 13]. Serological traits can therefore be viewed not merely as markers of prior exposure, but also as genetically influenced indicators of inter-individual variation in immune responsiveness [14]. In parallel, high-throughput metabolomics has revealed widespread alterations in circulating lipids, amino acids, and other small molecules that may affect inflammation, membrane biology, neurotransmission, and mood-related pathways [15, 16]. However, because most existing evidence remains observational, it is difficult to determine whether these immune and metabolic alterations represent causes, correlates, or consequences of depression, and findings across studies have therefore remained inconsistent [15, 17].

Mendelian randomization (MR) offers a useful framework for investigating whether genetically predicted variation in these traits is associated with depression risk while reducing confounding and reverse causation [18, 19]. In the present study, we adopted a biologically and temporally informed framework in which genetically predicted antibody responses were modeled as immune-related exposures, plasma metabolites as candidate downstream intermediates, and depression as the clinical outcome. This structure was based on the premise that infection-related immune activation may influence systemic metabolism, including lipid remodeling, and that such metabolic alterations may in turn contribute to depression susceptibility [20, 21]. Accordingly, our analyses were designed to evaluate a hypothesized immunometabolic pathway rather than to establish a uniquely identified causal chain [22]. This framework therefore provides a biologically motivated basis for examining how immune responses and downstream metabolic changes may relate to depression vulnerability.

In this study, we systematically evaluated the relationships among antibody immune responses, plasma metabolites, and depression using large-scale genome-wide association study (GWAS) summary statistics from publicly available biobanks. We first screened antibody-depression and metabolite-depression associations using MR, and then applied two-step MR-based mediation analyses to assess the extent to which metabolic alterations may transmit immune-related effects on depression risk. In addition, we used generalized summary-data-based Mendelian randomization (GSMR) as a complementary analysis to examine the consistency of the identified associations under an alternative summary-data-based framework. In the primary analysis, depression was modeled using a prespecified main outcome dataset. We then performed external replication analyses in an independent large-scale depression GWAS to further examine the generalizability of the main findings. Finally, we examined additional depression-related phenotypes in a secondary validation stage to assess whether the identified immunometabolic signals were directionally consistent across related but non-identical clinical definitions of depressive illness. By integrating genetic, immunological, and metabolomic data, this study aims to support an immunometabolic perspective on depression vulnerability and to prioritize candidate biomarkers and pathways for future mechanistic and clinical investigation.

Methods

Study design

This study was designed to investigate the relationships among antibody immune responses, plasma metabolites, and depression, and to explore potential immunometabolic mediation pathways (Fig. 1A). We integrated publicly available GWAS summary statistics for antibody traits, plasma metabolites, and depression-related outcomes. Antibody immune responses were modeled as upstream immune-related exposures, plasma metabolites were considered candidate downstream mediators, and depression was treated as the clinical outcome of interest (Fig. 1B). In the primary analysis, we first performed MR to evaluate the associations of antibody immune responses with the main depression outcome. We then conducted MR analyses to assess the associations of plasma metabolites with the same depression outcome. For associations that were statistically supported in these primary screening analyses, we subsequently applied two-step MR-based mediation analyses to examine whether specific metabolites were compatible with mediating the relationship between antibody immune responses and depression risk, and to estimate the corresponding indirect effects. We further applied GSMR as a complementary analysis to re-examine the significant antibody-depression, metabolite-depression, and antibody-metabolite associations identified in the primary MR analyses. We also examined the main immunometabolic signals in an independent large-scale depression GWAS to assess whether similar association patterns were observed beyond the primary outcome resource. Finally, we evaluated the significant antibody and metabolite traits across additional depression-related phenotypes, including antidepressant medication use, depression or dysthymia, psychotic depression, and recurrent or chronic depression. These analyses were conducted as a secondary stage to examine whether similar immunometabolic patterns were observed across related but non-identical clinical definitions of depressive illness, rather than as co-primary discovery analyses.

Fig. 1. All analyses illustrated in this figure were independently designed and performed by the authors using publicly available summary-level datasets.

A Workflow of the Mendelian randomization analysis. B Schematic diagram of mediation analysis. No pre-performed or third-party analyses were used.

Data sources

A detailed overview of all GWAS data sources used in this study, including phenotype definitions, sample sizes, and ancestry information, is provided in Supplementary Table S1. Genetic association data for antibody immune responses were obtained from the UK Biobank (UKB), a large prospective cohort that recruited more than 500,000 participants from the United Kingdom between 2006 and 2010 [23]. Within this cohort, a subsample of 8735 individuals provided serum samples for high-throughput serological profiling. Antibody responses to 13 pathogens were assessed, yielding 46 immune-related phenotypes, including 15 binary serostatus traits and 31 quantitative antibody measurements across multiple antigens [23]. Cohort-level genotyping quality control and covariate adjustment were conducted in the original GWAS analyses, including the exclusion of low-quality genotypes and adjustment for age, sex, and principal components [23]. GWAS summary statistics were obtained from the GWAS Catalog hosted by the European Bioinformatics Institute (EBI, https://www.ebi.ac.uk/) under the study accessions GCST90006884 to GCST90006929. Detailed dataset-level summary information, including sample size and variant counts, is provided in Supplementary Table S2.

Plasma metabolomics data were obtained from the Canadian Longitudinal Study on Aging (CLSA), a large population-based cohort that enrolled more than 50,000 Canadians aged 45 to 85 years at baseline [24, 25]. For the present analysis, we focused on 8299 unrelated individuals of European ancestry who underwent genome-wide genotyping and comprehensive plasma metabolite profiling. GWAS summary statistics covering 1091 individual metabolites and 309 metabolite ratios, representing a total of 1400 metabolomic traits, were obtained from the GWAS Catalog hosted by the European Bioinformatics Institute (https://www.ebi.ac.uk/gwas/) under accession numbers GCST90199621 to GCST90201020. Detailed dataset-level summary information for these metabolite GWAS datasets is provided in Supplementary Table S3.

Summary statistics for depression and related clinical phenotypes were obtained from the FinnGen biobank (Version: R12, https://www.finngen.fi/en), a large population-based study integrating nationwide Finnish health registries with genomic data [26]. The primary outcome for the main analyses was the FinnGen depression endpoint, which included 59,333 cases and 434,831 controls. This endpoint was based on registry-derived clinical diagnoses of depressive episodes and recurrent depressive disorder recorded in hospital discharge and cause-of-death registers (ICD-10: F32, F33; ICD-9: 2961, 2968, 3004), and included psychotic depression within the FinnGen endpoint hierarchy. Controls excluded broader mood disorder endpoints. To examine whether similar patterns were observed across related but non-identical clinical definitions of depressive illness, we additionally evaluated several secondary outcomes from the same resource, including depression or dysthymia (72,588 cases and 254,393 controls), psychotic depression (2622 cases and 254,393 controls), recurrent or chronic depression (26,094 cases and 254,393 controls), and antidepressant medication use (149,403 cases and 111,976 controls). Compared with the primary depression endpoint, the depression-or-dysthymia phenotype captures a broader affective diagnostic spectrum by additionally incorporating dysthymia-related diagnoses, including primary healthcare outpatient records, whereas the recurrent-or-chronic depression phenotype represents a narrower subgroup enriched for persistent or relapsing illness and is defined specifically by recurrent depressive disorder diagnoses (ICD-10: F33). The psychotic depression phenotype represents a more clinically specific and severe subgroup, defined by depressive episodes with psychotic symptoms (ICD-10: F32.3 and F33.3; ICD-9: 2961E). By contrast, the antidepressant medication endpoint is treatment-based rather than diagnosis-based and reflects antidepressant use categories within the FinnGen framework rather than a formal depressive disorder diagnosis. Collectively, these additional phenotypes were selected to reflect complementary clinical dimensions of depressive illness, including broader diagnostic breadth, psychotic features, recurrence burden, and treatment-related ascertainment. Because all outcomes were derived from the same biobank framework, partial sample overlap is expected; accordingly, these phenotypes were analyzed as secondary outcomes rather than as co-primary discovery endpoints.

To further examine the main findings beyond the primary outcome resource, we performed external replication analyses using summary statistics from the Psychiatric Genomics Consortium major depression GWAS release (PGC-MDD2025), specifically the European-ancestry dataset excluding both 23andMe and UK Biobank [27] (https://pgc.unc.edu/). According to the consortium documentation, this release represents a QC-filtered European-ancestry major depression meta-analysis and was selected to provide an additional outcome resource while minimizing potential overlap with the UK Biobank-derived antibody GWAS. The accompanying cohort-level metadata indicated a total of 357,636 cases and 1,281,936 controls. This dataset was used as an external outcome resource for the principal associations identified in the primary MR analyses.

This study was reported in accordance with the STROBE-MR guidelines, and the completed checklist is provided in Supplementary File 1.

Selection of instrumental variables

After obtaining the GWAS summary statistics, we applied a standardized workflow for post-download processing and instrumental-variable selection before MR analysis [28]. First, single nucleotide polymorphisms (SNPs) associated with each exposure trait were selected according to prespecified significance thresholds. For antibody immune responses, we used a suggestive threshold of P < 1 × 10⁻⁵, whereas for plasma metabolites, we applied the conventional genome-wide significance threshold of P < 5 × 10⁻⁸ [29, 30]. To ensure adequate allele frequency, only variants with a minor allele frequency (MAF) greater than 0.05 were retained [31]. Second, to reduce redundancy and promote independence among instruments, SNPs were pruned using stringent clumping criteria (r² <0.001) with a physical distance threshold of 10,000 kb based on the 1000 Genomes Project (Phase 3) European reference panel. Third, weak instruments were excluded by removing SNPs with F-statistics below 10, thereby reducing potential bias due to insufficient instrument strength [32, 33]. During harmonization with the outcome datasets, palindromic or otherwise ambiguous variants were excluded to minimize strand-misalignment errors. Detailed summaries of the instrumental variables used in all analyses, including genome-wide significant variants, LD-clumped SNPs, and post-harmonization final SNP sets, are provided in Supplementary Tables S2–S15.

MR analysis and sensitivity tests

We conducted MR analyses using the “TwoSampleMR” R package (version 0.6.4) to evaluate genetically predicted associations of antibody immune responses and plasma metabolites with depression risk. MR analyses were interpreted under the standard instrumental-variable assumptions that the selected genetic instruments are associated with the exposure of interest, are not associated with major confounders of the exposure–outcome relationship, and influence the outcome primarily through the exposure [28]. Because these assumptions cannot be verified directly from summary statistics alone, we combined multiple complementary MR estimators and sensitivity analyses to evaluate the consistency of the findings under different patterns of pleiotropy and instrument validity. For exposures represented by multiple independent variants, the inverse variance weighted (IVW) method was adopted as the primary estimator because of its high statistical efficiency under balanced pleiotropy assumptions [18, 34]. Additional approaches, including MR-Egger regression, weighted median, simple mode, and weighted mode estimators, were implemented as complementary analyses. IVW results served as the main reference, whereas the alternative estimators provided triangulation evidence for interpretation [18]. To address potential inflation of false positives due to multiple testing, we applied Bonferroni correction, with the significance threshold defined as 0.05 divided by the number of tests performed in each analytical stage. We also carried out a series of sensitivity analyses. MR-Egger intercept tests were used to assess directional pleiotropy, with P < 0.05 taken as evidence suggestive of potential pleiotropic bias [35]. Associations showing such evidence were examined further in conjunction with complementary sensitivity analyses, including weighted median and weighted mode estimators, heterogeneity statistics, MR-PRESSO, and leave-one-out analyses. Overall heterogeneity among instrumental variables was examined using Cochran’s Q statistic within the IVW framework and Rücker’s Q statistic within MR-Egger regression, allowing heterogeneity to be assessed under both models [36]. Additionally, we applied the MR Steiger directionality test to evaluate whether the direction of association was more consistent with exposure-to-outcome effects than with the reverse [37]. Variants showing stronger associations with the outcome than with the exposure were excluded, and MR analyses were repeated to reduce potential reverse-direction bias. MR-PRESSO (Mendelian Randomization Pleiotropy Residual Sum and Outlier) was used to evaluate overall distortion and identify potential outlier variants [38]. When heterogeneity was detected, the corresponding associations were further investigated using MR-PRESSO, alternative MR estimators, and leave-one-out analyses. If outlier SNPs were identified, analyses were repeated after excluding those variants; otherwise, the original estimates were retained and interpreted in the context of the heterogeneity results. Finally, leave-one-out analyses were conducted by sequentially omitting each variant and recalculating the effect estimate to identify any single SNP exerting disproportionate influence on the overall result. Effect estimates were reported according to the scale of the corresponding outcome GWAS: binary depression outcomes were presented as odds ratios (ORs) with 95% confidence intervals, whereas analyses based on continuous or linear effect estimates were presented as β coefficients with 95% confidence intervals.

Mediation analysis

To investigate whether plasma metabolites may mediate the relationship between antibody immune responses and depression, we performed a two-step MR-based mediation analysis [39]. Effect estimates for each step were obtained using the IVW method, which was selected as the primary estimator because of its statistical efficiency under standard MR assumptions [40]. In the first step, we estimated the association of antibody immune responses with each prioritized metabolite (β₁). In the second step, we estimated the association of each metabolite with depression risk (β₂). The overall antibody–depression effect was obtained independently from the primary MR analysis and was treated as the total effect (β₃). The indirect effect was calculated as the product of the two step-specific estimates (β_indirect = β₁ × β₂), and the direct effect was derived as the residual effect after subtracting the indirect effect from the total effect (β_direct = β₃ − β_indirect). The mediation proportion was then estimated by dividing the indirect effect by the total effect. Confidence intervals for the indirect effect and mediation proportion were calculated using the Delta method. For the mediation analyses, statistical significance was evaluated after correction for the number of tested mediators.

GSMR

After statistically supported associations had been identified in the primary MR analyses, we employed GSMR as a complementary method to further examine these findings using an alternative summary-data-based framework. Specifically, GSMR was used to re-evaluate the significant antibody–depression, metabolite–depression, and antibody–metabolite associations identified in the primary MR analyses. For GSMR, we used the same LD-clumped instrumental SNP sets as those retained in the primary MR analyses, thereby maintaining a set of near-independent instruments rather than defining a separate instrument-selection procedure. The 1000 Genomes Project (Phase 3) European reference panel was used to estimate linkage disequilibrium patterns, consistent with the primary MR instrument-selection process. To reduce potential bias related to pleiotropy, the heterogeneity in dependent instruments (HEIDI) outlier test was applied. Variants flagged by HEIDI as showing inconsistent effects across exposure and outcome traits were excluded using a P-value cutoff of 0.01 [41]. GSMR analyses were then repeated after excluding these variants to assess the consistency of the effect estimates.

Results

Identification of antibody traits through MR and sensitivity analyses

We evaluated the associations of 46 antibody traits with depression using MR. To correct for multiple comparisons, a Bonferroni-adjusted threshold of P < 0.05/46 (1.09 × 10⁻³) was applied. In the primary analysis, two traits reached statistical significance: higher EBV ZEBRA antibody levels were associated with increased depression risk (IVW: OR = 1.040, 95% CI: 1.018–1.062, P = 2.59 × 10⁻⁴), whereas higher VZV glycoprotein E/I antibody levels were associated with a lower risk of depression (IVW: OR = 0.927, 95% CI: 0.889–0.966, P = 3.40 × 10⁻⁴) (Fig. 2A; Supplementary Table S16).

Fig. 2. Identification of antibody traits through MR and sensitivity analyses.

A Volcano plot showing MR results for 46 antibody traits in relation to depression. B Forest plot of outlier-corrected MR estimates after heterogeneity adjustment. C Diagnostic plots from sensitivity analyses of EBV ZEBRA antibody levels, including MR-Egger tests. IVW: Inverse variance weighted; CI Confidence interval, OR Odds ratio.

Evaluation of pleiotropy and heterogeneity showed that both associations were affected by residual heterogeneity, particularly for VZV glycoprotein E/I (Cochran’s Q P = 2.50 × 10^⁻26; Rücker’s Q P = 4.77 × 10^⁻26; MR-PRESSO global test P < 0.001; Supplementary Table S17). After excluding outlier variants, only the association with EBV ZEBRA antibody levels remained significant (IVW: OR = 1.035, 95% CI: 1.016–1.053, P = 1.91 × 10⁻⁴), whereas the VZV glycoprotein E/I signal was attenuated and no longer met the significance threshold (Fig. 2B; Supplementary Tables S4, S5, and S18). Additional sensitivity analyses supported the EBV ZEBRA result: the MR-Egger intercept did not indicate directional pleiotropy, alternative MR estimators showed similar effect estimates, and leave-one-out analysis did not suggest disproportionate influence from any single SNP (Cochran’s Q P = 0.043; Rücker’s Q P = 0.054; MR-PRESSO global test P = 0.044) (Fig. 2C; Supplementary Table S18; Supplementary Figure S1). Taken together, these findings support a genetically predicted association between higher EBV ZEBRA antibody levels and increased depression risk, whereas the VZV glycoprotein E/I signal appeared less stable and was likely influenced by heterogeneity.

Screening of candidate metabolites associated with depression

From the initial panel of 1400 metabolites, 571 provided valid instruments for MR analysis with depression (Supplementary Tables S3 and S6). In the primary IVW screening analyses, applying a Bonferroni-adjusted significance threshold of P < 0.05/571 (8.75 × 10⁻⁵), 17 metabolites showed significant associations with depression (Fig. 3A; Supplementary Table S19).

Fig. 3. Screening of candidate metabolites associated with depression.

A Volcano plot showing MR results for 571 plasma metabolites in relation to depression, with Bonferroni-significant associations highlighted. B Forest plot showing heterogeneity-adjusted MR estimates. IVW: Inverse variance weighted; CI Confidence interval, OR Odds ratio.

Several of these 17 metabolites showed evidence of heterogeneity (Supplementary Table S20). Notably, 1-arachidonoyl-GPC (20:4n6), 1-myristoyl-2-arachidonoyl-GPC (14:0/20:4), 1-stearoyl-2-arachidonoyl-GPC (18:0/20:4), and the oleoyl-linoleoyl-glycerol (18:1 to 18:2) [2] to linoleoyl-arachidonoyl-glycerol (18:2 to 20:4) [1] ratio exhibited clear evidence of instrument heterogeneity. Two additional traits, 1-arachidonoyl-GPE (20:4n6) and 1-(1-enyl-stearoyl)-2-arachidonoyl-GPE (p-18:0/20:4), showed nominal heterogeneity (P < 0.05) in Q-based tests, although MR-PRESSO did not identify specific outlier variants. MR-Egger intercepts did not indicate directional pleiotropy across these metabolites. After heterogeneity assessment and correction where applicable, similar overall patterns were observed (Fig. 3B; Supplementary Tables S7 and S21; Supplementary Figure S2). Most significant metabolites showed inverse associations with depression, particularly those enriched in arachidonoyl-containing phospholipids and arachidonate-related traits. For example, higher levels of 1-arachidonoyl-GPE (20:4n6) (IVW OR = 0.931, 95% CI: 0.910–0.952, P = 7.91 × 10⁻¹⁰), 1-arachidonylglycerol (20:4) (OR = 0.890, 95% CI: 0.863–0.917, P = 4.46 × 10⁻¹⁴), and arachidonate (20:4n6) (OR = 0.903, 95% CI: 0.880–0.926, P = 5.77 × 10⁻¹⁵) were all inversely associated with depression. Similar inverse associations were observed for ratio traits involving arachidonate, such as the arachidonate-to-pyruvate ratio (OR = 0.886, 95% CI: 0.856–0.918, P = 1.03 × 10⁻¹¹). In contrast, a smaller group of metabolites was positively associated with depression risk. These included 1,2-dilinoleoyl-GPC (18:2/18:2) (OR = 1.121, 95% CI: 1.083–1.160, P = 8.80 × 10⁻¹¹), 1-(1-enyl-palmitoyl)-2-linoleoyl-GPE (p-16:0/18:2) (OR = 1.149, 95% CI: 1.088–1.213, P = 5.23 × 10⁻⁷), and multiple oleoyl-linoleoyl-glycerol to linoleoyl-arachidonoyl-glycerol ratios (OR range 1.056–1.068, P < 1.00 × 10⁻⁹). Overall, these screening analyses suggested a broadly coherent metabolite pattern in which arachidonate-enriched traits tended to show inverse associations with depression, whereas a more limited number of linoleoyl-related traits showed positive associations.

Identification of candidate metabolite mediators linked to antibody response

We next examined whether EBV ZEBRA antibody levels were associated with the 17 metabolites previously identified in the metabolite-to-depression screening analyses. Using the IVW method as the primary estimator and applying a Bonferroni threshold of P < 0.05/17 (2.94 × 10⁻³), four metabolites showed significant associations (Fig. 4; Supplementary Tables S8, S9, and S22). Higher EBV ZEBRA antibody levels were associated with higher levels of 1-arachidonylglycerol (20:4) (β = 0.057, 95% CI: 0.023–0.090, P = 9.09 × 10⁻⁴), 1,2-dilinoleoyl-GPC (18:2/18:2) (β = 0.105, 95% CI: 0.065–0.144, P = 1.95 × 10⁻⁷), the oleoyl-linoleoyl-glycerol [2] to linoleoyl-arachidonoyl-glycerol [1] ratio (β = 0.077, 95% CI: 0.039–0.115, P = 7.28 × 10⁻⁵), and the oleoyl-linoleoyl-glycerol [2] to linoleoyl-arachidonoyl-glycerol [2] ratio (β = 0.060, 95% CI: 0.024–0.096, P = 0.001) (Fig. 4). Sensitivity analyses further contextualized these results (Supplementary Table S23; Supplementary Figure S3). The MR-Egger intercepts did not indicate directional pleiotropy for any of the four metabolites. For 1-arachidonylglycerol (20:4) and both oleoyl-linoleoyl ratios, neither significant heterogeneity nor MR-PRESSO outliers were detected. For 1,2-dilinoleoyl-GPC (18:2/18:2), Q statistics suggested moderate heterogeneity; however, MR-PRESSO did not identify specific outlier variants, suggesting that the observed association was unlikely to be driven by a single influential SNP. To further assess the possibility of downstream selection bias, we additionally examined associations between EBV ZEBRA antibody levels and the broader set of 571 metabolites with valid instruments; the corresponding MR results and sensitivity analyses for this broader screen are provided in Supplementary Tables S22 and S23.

Fig. 4. Forest plots showing MR analyses of EBV ZEBRA antibody levels with depression-associated metabolites.

Four metabolites reached Bonferroni significance, supported by sensitivity analyses. IVW Inverse variance weighted, CI Confidence interval.

Metabolic mediation of the EBV ZEBRA antibody effect on depression

To further evaluate potential metabolic mediators, we conducted MR-based mediation analyses on the four metabolites that were Bonferroni-significant in both the antibody and depression association analyses. Applying a multiple-testing threshold (P < 0.05/4 = 0.0125), all four metabolites showed significant indirect effects (Table 1; Supplementary Table S24).

Table 1.

Mediation analysis of prioritized metabolites for the EBV ZEBRA antibody-depression association.

Mediator	Antibody	Outcome	Indirect effect	Direct effect	Z statistic	P value	Mediation proportion	Lower 95% CI	Upper 95% CI
1-arachidonylglycerol (20:4) levels	Epstein-Barr virus ZEBRA antibody	Depression	−0.007	0.041	−3.230	0.001	−0.196	−0.314	−0.077
1,2-dilinoleoyl-GPC (18:2/18:2) levels			0.012	0.022	4.106	4.03 × 10⁻⁵	0.351	0.183	0.519
Oleoyl-linoleoyl-glycerol (18:1 to 18:2) [2] to linoleoyl-arachidonoyl-glycerol (18:2 to 20:4) [1] ratio			0.005	0.029	2.987	0.003	0.141	0.049	0.234
Oleoyl-linoleoyl-glycerol (18:1 to 18:2) [2] to linoleoyl-arachidonoyl-glycerol (18:2 to 20:4) [2] ratio			0.004	0.030	2.996	0.003	0.115	0.040	0.191

Indirect effect: mediated effect estimate (β₁ × β₂); Direct effect: direct effect estimate (β₃ − β_indirect); Z statistic: Z value for the indirect effect; Mediation proportion: proportion mediated; Lower 95% CI: lower 95% confidence interval for the mediation proportion; Upper 95% CI: upper 95% confidence interval for the mediation proportion.

Among them, 1,2-dilinoleoyl-GPC (18:2/18:2) showed the largest positive indirect effect (β_indirect = 0.012, 95% CI: 0.006–0.018, P = 4.03 × 10⁻⁵), accounting for an estimated 0.351 (95% CI: 0.183–0.519) of the total effect. Two glycerolipid ratios also showed consistent positive mediation: the oleoyl-linoleoyl-glycerol (18:1 to 18:2) [2] to linoleoyl-arachidonoyl-glycerol (18:2 to 20:4) [1] ratio (β_indirect = 0.005, 95% CI: 0.002–0.008, P = 0.003; mediation proportion = 0.141, 95% CI: 0.049–0.234), and the oleoyl-linoleoyl-glycerol (18:1 to 18:2) [2] to linoleoyl-arachidonoyl-glycerol (18:2 to 20:4) [2] ratio (β_indirect = 0.004, 95% CI: 0.001–0.006, P = 0.003; mediation proportion = 0.115, 95% CI: 0.040–0.191). In contrast, 1-arachidonylglycerol (20:4) showed a significant negative indirect effect (β_indirect = –0.007, 95% CI: –0.011 to –0.003, P = 0.001), corresponding to an estimated mediation proportion of –0.196 (95% CI: –0.314 to –0.077). This pattern is consistent with a suppressive indirect effect, in which the direct effect estimate was larger after accounting for the negative indirect pathway. Taken together, these findings suggest that multiple metabolites may contribute to the association between EBV ZEBRA antibody levels and depression risk through indirect pathways, with 1,2-dilinoleoyl-GPC (18:2/18:2) showing the largest positive indirect effect in the present analyses.

Complementary GSMR analysis of antibody and metabolite associations with depression

To further examine the main MR findings using an alternative summary-data-based framework, we conducted GSMR analyses for the identified traits (Fig. 5; Supplementary Table S25). Consistent with the primary MR results, higher EBV ZEBRA antibody levels were associated with an increased risk of depression (β = 0.032, 95% CI: 0.015–0.049, P = 1.74 × 10⁻⁴). Among the metabolites, 1-arachidonylglycerol (20:4) showed an inverse association with depression (β = –0.115, 95% CI: –0.147 to –0.083, P = 2.11 × 10⁻¹²), whereas 1,2-dilinoleoyl-GPC (18:2/18:2) (β = 0.112, 95% CI: 0.076–0.149, P = 1.16 × 10⁻⁹), the oleoyl-linoleoyl-glycerol (18:1 to 18:2 [2]) to linoleoyl-arachidonoyl-glycerol (18:2 to 20:4 [1]) ratio (β = 0.065, 95% CI: 0.048–0.083, P = 1.13 × 10⁻¹³), and the corresponding [2] ratio (β = 0.064, 95% CI: 0.044–0.084, P = 1.32 × 10⁻¹⁰) showed positive associations with depression. Reciprocal GSMR analyses also showed directionally consistent associations of EBV ZEBRA antibody levels with these metabolites, including 1-arachidonylglycerol (β = 0.057, 95% CI: 0.021–0.093, P = 0.002), 1,2-dilinoleoyl-GPC (β = 0.111, 95% CI: 0.075–0.146, P = 1.34 × 10⁻⁹), the oleoyl-linoleoyl-glycerol (18:1 to 18:2 [2]) to linoleoyl-arachidonoyl-glycerol (18:2 to 20:4 [1]) ratio (β = 0.090, 95% CI: 0.052–0.128, P = 3.19 × 10⁻⁶), and the corresponding [2] ratio (β = 0.076, 95% CI: 0.039–0.114, P = 7.17 × 10⁻⁵). Overall, the GSMR analyses showed effect directions consistent with those observed in the primary MR analyses.

Fig. 5. Complementary GSMR analysis of antibody and metabolite associations with depression.

Forest plots showing GSMR analyses for EBV ZEBRA antibody levels and prioritized metabolites in relation to depression. Reciprocal analyses of antibody–metabolite associations are also presented. CI Confidence interval.

External replication of the main associations in PGC-MDD2025

To further examine the main findings beyond the primary outcome resource, we analyzed the independent PGC-MDD2025 (EUR, noUKB) dataset. In the initial MR analyses, higher EBV ZEBRA antibody levels remained associated with increased depression risk, and the four prioritized metabolites showed directionally consistent associations with those observed in the primary analyses (Supplementary Table S26). However, sensitivity analyses indicated residual heterogeneity for the associations involving EBV ZEBRA antibody levels and the oleoyl-linoleoyl-glycerol to linoleoyl-arachidonoyl-glycerol [1] ratio. For EBV ZEBRA antibody levels, MR-PRESSO identified outlier variants, and the outlier-corrected estimate remained statistically significant, although residual heterogeneity persisted (Cochran’s Q P = 0.048; Rücker’s Q P = 0.037; MR-PRESSO global test P = 0.092; Supplementary Tables S11 and S26). In contrast, although heterogeneity signals were also observed for the ratio [1] metabolite, MR-PRESSO did not detect any outlier SNPs that could be removed, and the original estimate was therefore retained for interpretation (Cochran’s Q P = 2.62 × 10^⁻5; Rücker’s Q P = 7.90 × 10^⁻5; MR-PRESSO global test P < 2 × 10^⁻4; Supplementary Tables S10, S12, and S27). After outlier removal, higher EBV ZEBRA antibody levels remained significantly associated with increased depression risk (IVW: OR = 1.025, 95% CI: 1.013–1.037, P = 3.96 × 10⁻⁵), with concordant support from GSMR (OR = 1.026, 95% CI: 1.016–1.037, P = 1.76 × 10⁻⁷) (Fig. 6; Supplementary Tables S27 and S28). As shown in Fig. 6, 1-arachidonylglycerol (20:4) remained inversely associated with depression, whereas 1,2-dilinoleoyl-GPC (18:2/18:2) and both oleoyl-linoleoyl-glycerol to linoleoyl-arachidonoyl-glycerol ratios remained positively associated with risk, with closely concordant effect estimates across IVW and GSMR analyses. We next examined whether the four prioritized metabolites were also compatible with a mediating role in the EBV ZEBRA antibody–depression association in the external PGC dataset (Table 2). Consistent with the primary analyses, 1-arachidonylglycerol (20:4) showed a significant negative indirect effect (β_indirect = –0.004, 95% CI: –0.006 to –0.001, P = 0.002), whereas 1,2-dilinoleoyl-GPC (18:2/18:2) showed a significant positive indirect effect (β_indirect = 0.008, 95% CI: 0.003–0.012, P = 0.001) (Supplementary Table S29). By contrast, the two glycerolipid ratios showed directionally similar but weaker mediation signals, with the ratio [1] mediator not reaching statistical significance and the ratio [2] mediator remaining only nominally significant. Taken together, these analyses indicated that the main antibody and metabolite associations were observed in an independent depression GWAS, whereas external support for the mediation results was strongest for 1-arachidonylglycerol and 1,2-dilinoleoyl-GPC. In addition, the full set of 17 prioritized depression-associated metabolites was systematically re-evaluated in the independent PGC dataset using standardized MR and GSMR analyses, together with tests for heterogeneity, pleiotropy, and outlier effects; the complete results, including heterogeneity-corrected estimates where applicable, are presented in Supplementary Tables S11, S12, S26, S27, and S28.

Fig. 6. Forest plots showing IVW MR and GSMR estimates for EBV ZEBRA antibody levels and the four prioritized metabolites in relation to depression in the independent PGC-MDD2025 (EUR, noUKB) dataset.

IVW Inverse variance weighted, CI Confidence interval, OR Odds ratio.

Table 2.

External mediation analysis of prioritized metabolites for the EBV ZEBRA antibody-depression association in PGC-MDD2025.

Mediator	Outcome	Antibody	Indirect effect	Direct effect	Z statistic	P value	Mediation proportion	Lower 95% CI	Upper 95% CI
1-arachidonylglycerol (20:4) levels	PGC-MDD2025 (EUR, noUKB)	Epstein-Barr virus ZEBRA antibody levels	−0.004	0.028	−3.174	0.002	−0.158	−0.256	−0.061
1,2-dilinoleoyl-GPC (18:2/18:2) levels			0.008	0.017	3.244	0.001	0.306	0.121	0.490
Oleoyl-linoleoyl-glycerol (18:1 to 18:2) [2] to linoleoyl-arachidonoyl-glycerol (18:2 to 20:4) [1] ratio			0.003	0.022	1.842	0.065	0.114	−0.007	0.234
Oleoyl-linoleoyl-glycerol (18:1 to 18:2) [2] to linoleoyl-arachidonoyl-glycerol (18:2 to 20:4) [2] ratio			0.003	0.022	2.232	0.026	0.103	0.013	0.194

Indirect effect: mediated effect estimate (β₁ × β₂); Direct effect: direct effect estimate (β₃ − β_indirect); Z statistic: Z value for the indirect effect; Mediation proportion: proportion mediated; Lower 95% CI: lower 95% confidence interval for the mediation proportion; Upper 95% CI: upper 95% confidence interval for the mediation proportion.

MR and GSMR analyses across additional depression-related phenotypes

To further examine whether the main findings extended across related clinical definitions of depressive illness, we conducted additional analyses using both heterogeneity-filtered IVW MR and GSMR approaches, focusing on one immune trait (EBV ZEBRA antibody levels) and four prioritized metabolites across additional depression-related phenotypes, including antidepressant medication use, depression/dysthymia, psychotic depression, and recurrent or chronic depression (Fig. 7; Supplementary Tables S13, S14, S15, S30, S31, and S32).

Fig. 7. Forest plots showing heterogeneity-filtered IVW MR and GSMR estimates for EBV ZEBRA antibody levels and prioritized metabolites across additional depression-related phenotypes, including antidepressant medication use, depression/dysthymia, psychotic depression, and recurrent or chronic depression.

IVW Inverse variance weighted, CI Confidence interval, OR, Odds ratio.

In the initial IVW MR analyses, significant heterogeneity was observed in five exposure–outcome pairs: EBV ZEBRA antibody levels with antidepressant medication use, depression/dysthymia, psychotic depression, and recurrent or chronic depression, as well as the oleoyl-linoleoyl-glycerol ratio with depression/dysthymia (Supplementary Table S30). Subsequent MR-PRESSO analyses indicated that no outlier SNPs were present for the associations of EBV ZEBRA with psychotic depression or for the oleoyl-linoleoyl-glycerol ratio with depression/dysthymia; these results were therefore retained without correction. For the remaining associations, detected outliers were removed, and the corresponding estimates were re-evaluated (Supplementary Tables S14, S15, and S31). All final estimates, including those corrected via MR-PRESSO and those unaffected by heterogeneity, were subsequently examined using GSMR, which showed generally similar patterns across the depression-related phenotypes (Supplementary Table S32).

Immune trait. EBV ZEBRA antibody levels were positively associated with multiple phenotypes: depression/dysthymia (IVW: OR = 1.057, 95% CI: 1.037–1.077, P = 6.01 × 10⁻⁹; GSMR: OR = 1.061, 95% CI: 1.043–1.078, P = 2.16 × 10⁻¹²), psychotic depression (IVW: OR = 1.120, 95% CI: 1.028–1.220, P = 0.010; GSMR: OR = 1.149, 95% CI: 1.065–1.239, P = 3.42 × 10⁻⁴), and recurrent or chronic depression (IVW: OR = 1.048, 95% CI: 1.019–1.078, P = 0.001; GSMR: OR = 1.049, 95% CI: 1.023–1.076, P = 1.58 × 10⁻⁴), but not antidepressant medication use (IVW: OR = 0.997, 95% CI: 0.976–1.019, P = 0.798; GSMR: OR = 1.010, 95% CI: 0.993–1.027, P = 0.246).

Inverse-associated metabolite. 1-arachidonylglycerol (20:4) showed inverse associations with antidepressant medication use (IVW: OR = 0.941, 95% CI: 0.913–0.969, P = 5.50 × 10⁻⁵; GSMR: OR = 0.942, 95% CI: 0.914–0.971, P = 1.16 × 10⁻⁴), depression/dysthymia (IVW: OR = 0.904, 95% CI: 0.872–0.938, P = 5.81 × 10⁻⁸; GSMR: OR = 0.908, 95% CI: 0.881–0.937, P = 1.18 × 10⁻⁹), and recurrent or chronic depression (IVW: OR = 0.870, 95% CI: 0.832–0.911, P = 2.09 × 10⁻⁹; GSMR: OR = 0.873, 95% CI: 0.833–0.915, P = 1.78 × 10⁻⁸). For psychotic depression, the effect estimates were directionally similar but did not reach conventional significance (IVW: OR = 0.870, 95% CI: 0.758–1.000, P = 0.050; GSMR: OR = 0.871, 95% CI: 0.757–1.001, P = 0.052).

Positive-associated metabolite. In contrast, 1,2-dilinoleoyl-GPC (18:2/18:2) was positively associated with antidepressant medication use (IVW: OR = 1.057, 95% CI: 1.022–1.093, P = 0.001; GSMR: OR = 1.056, 95% CI: 1.021–1.093, P = 0.002), depression/dysthymia (IVW: OR = 1.093, 95% CI: 1.055–1.131, P = 5.13 × 10⁻⁷; GSMR: OR = 1.089, 95% CI: 1.052–1.128, P = 1.72 × 10⁻⁶), psychotic depression (IVW: OR = 1.231, 95% CI: 1.052–1.441, P = 0.010; GSMR: OR = 1.228, 95% CI: 1.048–1.440, P = 0.011), and recurrent or chronic depression (IVW: OR = 1.132, 95% CI: 1.075–1.191, P = 2.62 × 10⁻⁶; GSMR: OR = 1.128, 95% CI: 1.070–1.190, P = 9.11 × 10⁻⁶).

Metabolite ratios. Two ratios of oleoyl-linoleoyl-glycerol (18:1 to 18:2) to linoleoyl-arachidonoyl-glycerol (18:2 to 20:4) were positively associated with multiple depression-related phenotypes. For Ratio [1], significant positive associations were observed for antidepressant medication use (IVW: OR = 1.034, 95% CI: 1.018–1.051, P = 4.86 × 10⁻⁵; GSMR: OR = 1.034, 95% CI: 1.017–1.050, P = 6.97 × 10⁻⁵), depression/dysthymia (IVW: OR = 1.048, 95% CI: 1.025–1.071, P = 4.21 × 10⁻⁵; GSMR: OR = 1.046, 95% CI: 1.029–1.063, P = 5.47 × 10⁻⁸), psychotic depression (IVW: OR = 1.090, 95% CI: 1.011–1.174, P = 0.024; GSMR: OR = 1.089, 95% CI: 1.010–1.174, P = 0.027), and recurrent or chronic depression (IVW: OR = 1.071, 95% CI: 1.045–1.098, P = 3.58 × 10⁻⁸; GSMR: OR = 1.070, 95% CI: 1.044–1.097, P = 1.09 × 10⁻⁷). Ratio [2], derived from an alternative positional isomer, showed similar associations across these phenotypes, including antidepressant medication use (IVW: OR = 1.037, 95% CI: 1.018–1.057, P = 1.24 × 10⁻⁴; GSMR: OR = 1.037, 95% CI: 1.017–1.056, P = 1.84 × 10⁻⁴), depression/dysthymia (IVW: OR = 1.058, 95% CI: 1.035–1.082, P = 4.34 × 10⁻⁷; GSMR: OR = 1.056, 95% CI: 1.037–1.077, P = 1.49 × 10⁻⁸), psychotic depression (IVW: OR = 1.098, 95% CI: 1.007–1.198, P = 0.035; GSMR: OR = 1.097, 95% CI: 1.005–1.197, P = 0.038), and recurrent or chronic depression (IVW: OR = 1.083, 95% CI: 1.053–1.114, P = 4.02 × 10⁻⁸; GSMR: OR = 1.082, 95% CI: 1.051–1.113, P = 1.15 × 10⁻⁷). Taken together, the heterogeneity-filtered MR and GSMR analyses showed broadly similar patterns across the additional depression-related phenotypes.

Candidate metabolite mediation patterns across additional depression-related phenotypes

The broadly similar associations observed across IVW MR and GSMR analyses suggested that the main immunometabolic signals extended across additional depression-related phenotypes (Table 3; Supplementary Table S33). We next performed mediation analyses to examine whether the prioritized metabolites were compatible with a mediating role in the association between EBV ZEBRA antibody levels and these depression-related outcomes.

Table 3.

Mediation analysis of prioritized metabolites for the EBV ZEBRA antibody association across additional depression-related phenotypes.

Mediator	Outcome	Antibody	Indirect effect	Direct effect	Z statistic	P value	Mediation proportion	Lower 95% CI	Upper 95% CI
1-arachidonylglycerol (20:4) levels	Depression or dysthymia	Epstein-Barr virus ZEBRA antibody	−0.006	0.061	−2.717	0.007	−0.103	−0.178	−0.029
1,2-dilinoleoyl-GPC (18:2/18:2) levels			0.009	0.046	3.537	4.00 × 10⁻⁴	0.167	0.075	0.26
Oleoyl-linoleoyl-glycerol (18:1 to 18:2) [2] to linoleoyl-arachidonoyl-glycerol (18:2 to 20:4) [1] ratio			0.004	0.052	2.258	0.024	0.065	0.009	0.121
Oleoyl-linoleoyl-glycerol (18:1 to 18:2) [2] to linoleoyl-arachidonoyl-glycerol (18:2 to 20:4) [2] ratio			0.003	0.052	2.661	0.008	0.061	0.016	0.107
1-arachidonylglycerol (20:4) levels	Psychotic depression		−0.008	0.121	−0.799	0.424	−0.07	−0.24	0.101
1,2-dilinoleoyl-GPC (18:2/18:2) levels			0.022	0.092	1.291	0.197	0.192	−0.099	0.483
Oleoyl-linoleoyl-glycerol (18:1 to 18:2) [2] to linoleoyl-arachidonoyl-glycerol (18:2 to 20:4) [1] ratio			0.007	0.107	1.839	0.066	0.058	−0.004	0.121
Oleoyl-linoleoyl-glycerol (18:1 to 18:2) [2] to linoleoyl-arachidonoyl-glycerol (18:2 to 20:4) [2] ratio			0.006	0.108	1.303	0.192	0.049	−0.025	0.124
1-arachidonylglycerol (20:4) levels	Recurrent or chronic depression		−0.008	0.055	−2.344	0.019	−0.169	−0.31	−0.028
1,2-dilinoleoyl-GPC (18:2/18:2) levels			0.013	0.034	3.338	0.001	0.276	0.114	0.438
Oleoyl-linoleoyl-glycerol (18:1 to 18:2) [2] to linoleoyl-arachidonoyl-glycerol (18:2 to 20:4) [1] ratio			0.005	0.042	3.073	0.002	0.114	0.041	0.186
Oleoyl-linoleoyl-glycerol (18:1 to 18:2) [2] to linoleoyl-arachidonoyl-glycerol (18:2 to 20:4) [2] ratio			0.005	0.042	2.979	0.003	0.102	0.035	0.169

Indirect effect: mediated effect estimate (β₁ × β₂); Direct effect: direct effect estimate (β₃ − β_indirect); Z statistic: Z value for the indirect effect; Mediation proportion: proportion mediated; Lower 95% CI: lower 95% confidence interval for the mediation proportion; Upper 95% CI: upper 95% confidence interval for the mediation proportion.

Several metabolites showed significant indirect effects. 1-arachidonylglycerol (20:4) showed significant negative indirect effects for both depression or dysthymia (β_indirect = –0.006, 95% CI: –0.010 to –0.002, P = 6.58 × 10⁻³; mediation proportion = –0.103) and recurrent or chronic depression (β_indirect = –0.008, 95% CI: –0.014 to –0.001, P = 1.91 × 10⁻²; mediation proportion = –0.169). In contrast, 1,2-dilinoleoyl-GPC (18:2/18:2) showed significant positive indirect effects for depression or dysthymia (β_indirect = 0.009, 95% CI: 0.004 to 0.014, P = 4.04 × 10⁻⁴; mediation proportion = 0.167) and recurrent or chronic depression (β_indirect = 0.013, 95% CI: 0.005 to 0.021, P = 8.42 × 10⁻⁴; mediation proportion = 0.276).

For the oleoyl-linoleoyl-glycerol (18:1 to 18:2) [2] to linoleoyl-arachidonoyl-glycerol (18:2 to 20:4) [1] ratio, significant indirect effects were observed for depression or dysthymia (β_indirect = 0.004, 95% CI: 0.0005 to 0.007, P = 0.024; mediation proportion = 0.065) and recurrent or chronic depression (β_indirect = 0.005, 95% CI: 0.002 to 0.009, P = 2.12 × 10⁻³; mediation proportion = 0.114). Similarly, the corresponding [2] ratio showed significant indirect effects for depression or dysthymia (β_indirect = 0.003, 95% CI: 0.001 to 0.006, P = 7.80 × 10⁻³; mediation proportion = 0.061) and recurrent or chronic depression (β_indirect = 0.005, 95% CI: 0.002 to 0.008, P = 2.90 × 10⁻³; mediation proportion = 0.102). No significant indirect effects were detected for psychotic depression. Taken together, these results suggested that several prioritized metabolites were compatible with a mediating role in the association between EBV ZEBRA antibody levels and selected depression-related phenotypes.

Discussion

Our MR analyses showed that higher EBV ZEBRA antibody levels were associated with multiple depression-related phenotypes, particularly depression/dysthymia and recurrent or chronic depression, whereas the evidence for psychotic depression was weaker. Previous studies have mainly focused on antibodies against viral capsid antigen (VCA) or Epstein–Barr nuclear antigen-1 (EBNA-1), but the reported associations have been inconsistent, possibly because these markers primarily reflect prior exposure rather than lytic viral activity [20, 42, 43]. Accordingly, ZEBRA-specific antibody responses may more closely reflect processes related to viral reactivation [44, 45]. EBV reactivation has been linked to higher circulating CCL2 levels, a chemokine associated with chronic inflammation and immune aging [46, 47]. Persistent immune activation and altered B-cell metabolism may in turn contribute to broader inflammatory states, which have increasingly been implicated in depression pathophysiology [48, 49]. Within this framework, our findings are consistent with the possibility that EBV-related immune dysregulation may contribute to depression risk through immunometabolic pathways, although this interpretation should remain cautious.

Among the four prioritized metabolites, 1,2-dilinoleoyl-GPC (18:2/18:2), a polyunsaturated phosphatidylcholine (PC) species, showed the largest positive indirect effect in our mediation analyses. PCs are major components of neuronal membranes and are involved in synaptic plasticity, neurotransmitter cycling, and neuroinflammatory regulation [50–52]. Previous studies have often reported reduced total PC levels in individuals with depression, although these analyses have rarely distinguished among molecular subtypes [53]. Our findings suggest that PC-related signals may not be uniform across species, and that specific unsaturated molecules such as 1,2-dilinoleoyl-GPC, which contains two linoleic acid chains, may be linked to depression through distinct immunometabolic processes. Immune activation, particularly in the setting of viral reactivation, can perturb host lipid homeostasis [54]. Experimental studies have shown that EBV-infected cells, especially under hypoxic or stress-related conditions, may exhibit altered phospholipid synthesis and lipid allocation that support cellular transformation and viral replication [55, 56]. Such changes could plausibly contribute to the relative accumulation or altered turnover of specific PC species. In this context, unsaturated PCs such as 1,2-dilinoleoyl-GPC may reflect lipid states associated with altered membrane dynamics or pro-inflammatory signaling, thereby influencing neuroimmune communication. Evidence from animal studies also supports a role for glycerophospholipid dysregulation, including perturbations in PC pathways, in depression-related phenotypes [57, 58]. In addition, disturbances in the gut-brain axis have been linked to altered phosphatidylcholine metabolism, further supporting the relevance of lipid homeostasis to emotional regulation [58, 59]. Taken together, these observations are consistent with the possibility that 1,2-dilinoleoyl-GPC represents a candidate metabolite mediator linking EBV-related immune activity and depression vulnerability, while also highlighting the broader importance of lipid remodeling in immunometabolic models of depression.

Two additional ratio-based candidate mediators, namely the ratios of oleoyl-linoleoyl-glycerol (18:1/18:2) to linoleoyl-arachidonoyl-glycerol (18:2/20:4) ([1] and [2]), may reflect shifts in specific diacylglycerol (DAG) subclasses within the glycerolipid family. DAGs serve both as precursors for phospholipid synthesis and as second messengers involved in signaling pathways such as protein kinase C (PKC) activation [60]. In this context, the observed elevation of these ratios is consistent with the possibility that EBV-related immune activity is associated with a relative enrichment of oleoyl-linoleoyl species and a relative depletion of arachidonoyl-containing DAG species. Experimental studies have suggested that EBV infection can influence phosphatidylinositol-related lipid metabolism, including phosphatidylinositol 4-kinase and phosphatidylinositol 4-phosphate kinase activity, thereby potentially affecting phosphatidylinositol 4,5-bisphosphate (PIP₂) turnover and downstream DAG/IP₃ signaling [61, 62]. Such changes could plausibly favor altered accumulation of particular DAG species and perturb PKC-related signaling pathways [62]. Arachidonoyl-rich DAGs are also important substrates for the synthesis of 2-arachidonoylglycerol (2-AG), the major endocannabinoid regulating CB₁ receptor activity across serotonergic, GABAergic, and noradrenergic circuits [63–65]. Reduced availability of arachidonoyl-containing DAG species could therefore be relevant to endocannabinoid dysregulation, which has been implicated in depression vulnerability [66]. In addition, diacylglycerol kinase-mediated phosphorylation of DAG has been linked to PKCα localization and long-term synaptic plasticity, further connecting lipid remodeling to neuronal signaling processes. Nevertheless, these ratio-based findings should be interpreted with particular caution, because such traits are more likely to reflect broader compositional shifts in lipid metabolism than uniquely attributable mediator-specific pathways. In this sense, the two DAG ratios may be better viewed as composite metabolic signatures of immunometabolic remodeling rather than as isolated molecular mediators.

The remaining prioritized metabolite, 1-arachidonylglycerol (1-AG, 20:4), belongs to the monoacylglycerol (MAG) subclass of endocannabinoid-related lipids. Unlike the positive indirect effects observed for the other metabolites, 1-AG showed a negative indirect effect, a pattern consistent with a possible counterbalancing or protective role. Together with 2-arachidonoylglycerol (2-AG), 1-AG contributes to the broader pool of arachidonoyl-containing MAGs involved in endocannabinoid signaling [67]. 2-AG is a major ligand of CB₁ receptors and modulates neurotransmission across serotonergic, GABAergic, and noradrenergic pathways [63–65]. Within this framework, lower 1-AG levels may reflect reduced availability of arachidonoyl-related substrates relevant to endocannabinoid homeostasis, which could in turn be associated with impaired CB₁ signaling and greater vulnerability to depression. Accordingly, maintenance of arachidonoyl-containing MAG levels may be relevant to neuroimmune balance. Our findings are therefore compatible with the possibility that 1-AG marks a lipid state linked to relative protection in the context of EBV-associated depression, although this interpretation requires further validation.

The generally similar directions of effect across the additional depression-related phenotypes suggest that the identified immunometabolic signals may not be restricted to a single operational definition of depression, but may instead reflect biological processes relevant across a broader depressive spectrum. In particular, the consistent positive association of EBV ZEBRA antibody levels and the parallel metabolite-related patterns across depression/dysthymia and recurrent or chronic depression may indicate that EBV-related immune activation and downstream lipid remodeling are more closely linked to persistent or cumulative dimensions of depressive illness, such as chronicity, recurrence, and long-term inflammatory burden [68, 69]. By contrast, the weaker or less consistent findings for psychotic depression may reflect the greater clinical heterogeneity of this phenotype, its smaller sample size, and the possibility that psychotic features involve additional neurobiological mechanisms not fully captured by the present immunometabolic framework. The antidepressant medication phenotype should also be interpreted differently from diagnosis-based outcomes, because it reflects treatment-related ascertainment rather than depressive disorder itself and may therefore capture a broader mix of indication, severity, healthcare access, and prescribing behavior [68]. Taken together, these patterns suggest that the EBV-metabolite-depression axis may be more relevant to common and recurrent forms of depressive illness than to all phenotype definitions equally.

Overall, our findings support an immunometabolic interpretation in which EBV-related immune activity is associated with metabolic remodeling patterns that may be relevant to depression. Within this framework, 1,2-dilinoleoyl-GPC and the two DAG-related ratio traits were compatible with positive indirect effects, whereas 1-arachidonylglycerol showed a negative indirect effect consistent with a possible counterbalancing pattern. These observations suggest that integrating immune-related traits with metabolomic data may help to refine the characterization of EBV-associated depression risk. At the same time, the identified metabolites should be interpreted as candidate signals rather than definitive mechanistic intermediates, and any translational implications remain to be established through future experimental and longitudinal studies.

Several limitations should be acknowledged. First, the antibody traits were instrumented using a suggestive significance threshold (P < 1 × 10⁻⁵) rather than the conventional genome-wide significance threshold, because the available antibody GWAS yielded relatively few strongly associated variants. Although we attempted to mitigate weak-instrument bias through F-statistic filtering, complementary sensitivity analyses, GSMR, and external analyses, the antibody-related findings should still be interpreted with caution. Second, the metabolites taken forward to the antibody-metabolite and mediation analyses were prioritized on the basis of their metabolite-to-depression associations. Because the antibody exposure was itself associated with depression, this strategy may introduce downstream selection bias, including the possibility of collider bias, and may inflate subsequent EBV-to-metabolite or mediation effect estimates. To further assess this concern, we additionally examined associations between EBV ZEBRA antibody levels and the broader set of 571 metabolites with valid instruments. Under Bonferroni correction across all 571 tests, not all previously prioritized metabolite signals remained statistically significant. We therefore interpret these findings in a hierarchical manner, with the strongest emphasis placed on signals supported across multiple analytical layers, whereas metabolites showing weaker evidence under the broader 571-metabolite framework are regarded as more tentative and require additional validation. Third, the two-step MR mediation framework provides evidence compatible with a mediating role, but does not constitute definitive proof of a uniquely identified causal decomposition of direct and indirect effects. Fourth, two of the prioritized metabolites were ratio-based measures, which are more difficult to interpret causally than single-metabolite traits and may reflect broader compositional shifts in lipid metabolism rather than uniquely attributable molecular pathways. Fifth, although FinnGen participants are of European ancestry, the Finnish population represents a relatively distinct subpopulation within Europe; therefore, the findings may not be fully generalizable to broader European populations or to other ancestries. Sixth, the additional depression-related phenotypes were derived from the same FinnGen resource as the primary outcome, so partial sample overlap is expected and these analyses should be viewed as secondary rather than fully independent validation. Finally, the metabolites were measured in plasma rather than in central nervous system tissue, and EBV-related immune activity was inferred from genetically predicted ZEBRA antibody levels rather than direct measures of viral reactivation. Further longitudinal, multi-omic, and experimental studies will be needed to clarify the biological mechanisms and clinical relevance of these associations.

Author contributions

YZ, JH, and HW conceived and designed the study. YZ developed the methodology, performed the investigation, managed the resources, conducted data curation, prepared the original draft, and generated the visualizations. YZ and LZ performed software development and validation. YR conducted the formal analyses. FC and GS contributed to data collection, validation, and interpretation of the results. HW supervised the study and critically reviewed and edited the manuscript. HW and JH administered the project. LZ acquired funding for the study. All authors reviewed the manuscript and approved the final version for publication.

Funding

This work was supported by the Wujiang Science, Education and Health Project (WWK202122, WWK202503), the Suzhou Clinical Key Disease Diagnosis and Treatment Technology Program (LCZX202343), and the Suzhou Science and Technology Development Guidance Project (SYWD2025133).

Data availability

All relevant data from this study are included in the article. For additional information or inquiries, please contact the corresponding authors.

Ethics

All data used in this study were obtained from publicly available GWAS datasets. No new data collection was conducted, and ethical approval was not required.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41398-026-04236-7.