The gut microbiota and depressive symptoms across ethnic groups

Depressive disorders are highly prevalent and disabling, and current treatments are suboptimal. Emerging evidence suggests the gut microbiome may influence affect and cognition via the gut–brain axis, raising the possibility of microbiome-targeted interventions for depression. However, human data have often been limited to small, inconsistently adjusted studies, and prior large-scale analyses were conducted in relatively ethnically homogeneous European-ancestry cohorts. Given that both gut microbiota composition and depression burden vary across ethnic groups, it is crucial to examine whether microbiome–depression associations generalize across ethnicities and whether microbiota differences contribute to ethnic disparities in depressive symptoms. This study investigates associations between gut microbiota features and depressive symptom levels in a large, multi-ethnic urban cohort, tests generalizability across six ethnic groups, and assesses whether microbiota differences statistically account for ethnic disparities in depression.

Two large population studies have linked depression to gut microbiota composition. The LifeLines Study (N=1135) found depression associated with β-diversity, and the Flemish Gut Flora Project (N=1068) replicated this while adjusting for age, sex, BMI, and GI parameters; it also identified lower abundances of Dialister and Coprococcus among depressed individuals, after excluding those on antidepressants and cross-validating in LifeLines. These studies, however, used limited confounder adjustments and were conducted in ethnically homogeneous populations of North-European ancestry. Demographic, lifestyle, and health factors are major determinants of microbiome variation, with prior HELIUS analyses indicating ethnicity explains more variance in gut microbiota than other collected measures. The extent to which microbiome–depression associations generalize across ethnic groups and whether microbiota differences help explain ethnic disparities in depression remained unclear, motivating the present work.

Design and cohort: Cross-sectional analyses within the HELIUS cohort, a multi-ethnic population study in Amsterdam, The Netherlands. Participants aged 18–70 were sampled via municipal registry stratified by ethnicity. At the time of analysis, 16S rRNA fecal data were available for 3,343 participants from 8 groups; those identifying as Indonesian-Surinamese (N=46) and other/unknown (N=63) and those without PHQ-9 data (N=93) were excluded. The final analytic sample included six groups: Dutch (N=769), African Surinamese (N=767), South-Asian Surinamese (N=527), Turkish (N=349), Moroccan (N=473), and Ghanaian (N=458). Depending on covariate availability, N ranged from 3211 (Model 1) to 3088 (Model 3). Ethics approvals were obtained; informed consent was collected. Depression assessment: Depressive symptoms were measured with the Patient Health Questionnaire-9 (PHQ-9; range 0–27), invariant across the six ethnic groups. Single missing items were imputed using the mean of available items; >1 missing rendered the PHQ-9 missing. Covariates: Sociodemographics (ethnicity, age, sex, education), behaviors (smoking yes/no; alcohol via AUDIT; physical activity as minutes×intensity based on MET categories; BMI), and medical variables (self-reported GI disorder, diabetes via composite algorithm of self-report or fasting glucose ≥7 mmol/L or HbA1c ≥48 mmol or glucose-lowering medication; proton pump inhibitor use; antibiotics past 2 weeks; diarrhea past week). Antidepressant users (N=132) were excluded from main analyses; probiotics use recorded. Auxiliary analyses considered parental history of depression, number of prior depressive episodes, neuroticism; C-reactive protein was available in a subset (N=975). Stool collection and processing: Participants provided fresh stool samples within 6 h or stored at home freezer overnight; samples were frozen at −20 °C at assessment centers and transferred to −80 °C within 1–4 weeks (local −20 °C storage duration not logged). Standardized SOPs and quality checks were used. Sequencing and bioinformatics: The V4 region of 16S rRNA was sequenced on Illumina MiSeq (2×250 bp; 515F/806R primers). PCR duplicates were combined and purified; amplicons pooled and cleaned. Reads were processed with USEARCH: merging, quality filtering (≤1 expected error), dereplication; ASVs inferred via UNOISE3 (a=2.0), with sequences occurring ≥8 times retained. All reads mapped back to ASVs to generate an ASV table. Taxonomy assigned via SINTAX using Greengenes v13.5 and SILVA 132. The ASV table was rarefied to 14,942 counts/sample. Multiple sequence alignment via MAFFT and phylogeny via IQ-TREE; tree midpoint-rooted. Integration via phyloseq. Of 1,438 ASVs identified, 418 had non-trivial counts (>0.02%); a core subset was defined as ASVs present in ≥75% of the cohort (N=109). Diversity metrics: α-diversity primarily via Shannon index (also Simpson, Chao1, ACE Observed, Phylogenetic Diversity); β-diversity via Bray–Curtis and weighted UniFrac distances followed by PCoA. The first 20 coordinates underwent forward selection; selected coordinates were used as predictors. Statistical analyses: Multiple linear regression modeled PHQ-9 sum scores as the dependent variable with three adjustment tiers: Model 1 (age, sex, ethnicity, education), Model 2 (Model 1 + smoking, alcohol, physical activity, BMI), Model 3 (Model 2 + GI disorder, diabetes, PPI use, recent antibiotics, diarrhea). For ASV-level analyses, both PHQ-9 and relative abundances were rank-transformed. Multiple testing at ASV-level controlled via FDR (Benjamini–Hochberg), significance at corrected p<0.05. Ethnic heterogeneity was assessed via (1) GLM interaction tests (ethnicity×ASV; FDR-adjusted) and (2) ethnicity-stratified associations summarized with heterogeneity f² (≥30% moderate; ≥50% high). Mediation analyses evaluated whether β-diversity attenuated the association between ethnicity and PHQ-9 following Baron & Kenny.

Sample: N ranged from 3211 (Model 1) to 3088 (Model 3) across analyses; six ethnic groups residing in the same urban area. Alpha-diversity: The Shannon index predicted PHQ-9 depressive symptom scores. Unadjusted standardized coefficient for Shannon was −0.120 (t=−6.66, p<0.0001). After adjustment: Model 1 β≈−0.073 (t=−3.84, p=0.0001); Model 2 β≈−0.060 (t=−3.15, p=0.0016); Model 3 β≈−0.042 (t=−2.22, p=0.026). Model fit improved with covariates: R² rose from 0.0144 (unadjusted) to 0.0754 (Model 1), 0.0841 (Model 2), 0.1109 (Model 3); ethnicity contributed most to Model 1 (ΔR²=0.0431 after age/sex). No significant ethnicity×α-diversity interactions (p≥0.134 across models); stratified analyses showed coefficients near zero across groups. Results replicated with the Simpson index. Specificity: Adjusting for neuroticism abolished the α-diversity–depression association; α-diversity remained significantly associated with neuroticism when adjusting for depressive symptoms, suggesting neuroticism is the stronger predictor. Adjusting for parental history and number of prior episodes only minimally attenuated associations (Model 3 standardized β>−0.0384, p<0.033). Reverse models: With Shannon as outcome, Model 3 explained ~18% of variance, largely due to ethnicity (ΔR²=0.1143 after age/sex); PHQ-9 remained a significant negative predictor of Shannon (β≈−0.039, p≈0.026). Beta-diversity: Forward-selected PCoA components derived from Bray–Curtis and weighted UniFrac predicted PHQ-9. Six coordinates were retained; PCoA#2 explained 6.5% (Bray–Curtis) and 9.73% (UniFrac) of microbiome variation and correlated strongly with Shannon (r=0.83), indicating α-diversity is integral to β-diversity in this framework. The selected components jointly explained modest increments in PHQ-9 variance (ΔR² ~1.5% in Model 1 to ~0.5% in Model 3). Mediation of ethnic disparities: β-diversity statistically attenuated the ethnicity–PHQ-9 association by 29.2% (Model 1a: age, sex), 23.0% (Model 1b: +education), 22.7% (Model 2: +behaviors), and 18.1% (Model 3: +medical factors). All AR² p≤0.001 except ethnicity-inclusive Model 3 (p=0.023). Taxa-level associations: Of 416 non-trivial ASVs, 117 showed significant unadjusted correlations with PHQ-9 (FDR<0.05), 99 negatively correlated, ~65% within Firmicutes (76 ASVs). After adjustment for age, sex, ethnicity (Model 1), 70 ASVs remained significant; after adding behaviors (Model 2), 48 remained; after full medical adjustment (Model 3), 23 ASVs (≥15 genera) remained significant. Prominent negatively associated taxa included Coprococcus (GCA-900066575; butyrate-producer), Bacteroides, Ruminococcaceae UCG-005, Ruminococcus 1, Peptococcus, Holdemanella (H. biformis), Roseburia (R. inulinivorans), and several Lachnospiraceae genera/groups (e.g., FCS020, NK4A136, Marvinbryantia including M. formatexigens, Blautia including B. obeum). Desulfovibrio was also negatively associated. Positively associated taxa included Blautia (B. caecimuris, B. producta), Lachnoclostridium, Oscillibacter, and Dialister (mapped to D. invisus). Christensenellaceae (R-7 group) showed negative associations that attenuated after medical adjustment (linked to BMI, diabetes, GI disease). Some ASVs within the same genus showed opposite directions (e.g., Blautia, Bacteroides, Oscillospira), highlighting intra-genus heterogeneity and taxonomy database differences (Greengenes vs SILVA). Across-ethnicity generalization: Ethnicity×ASV interactions were rare (<6% across models; unadjusted for multiple testing), approximating Type I error; 81% of ethnicity-stratified coefficients differed by <30%, with only 3.6% showing high heterogeneity (f>50%). Core taxa (≥75% prevalence) showed similar association patterns, effect sizes, robustness, and heterogeneity as non-core taxa.

This large, multi-ethnic cohort demonstrates robust associations between gut microbiota and depressive symptom severity across multiple levels (α-diversity, β-diversity, and specific taxa), persisting after extensive sociodemographic, behavioral, and medical adjustments. The associations generalized across six ethnic groups despite substantial ethnic differences in both depression levels and microbiota composition, supporting the broader applicability of microbiome-based assessments or interventions. β-diversity partially mediated ethnic disparities in depressive symptoms, suggesting that microbiota variation contributes to observed differences in depression across groups. The strong linkage of PCoA#2 with α-diversity, and the overlap of taxa related to α-diversity, depressive symptoms, and known depression risk factors (e.g., BMI, metabolic and inflammatory markers), align with the view of α-diversity as a generic biomarker of health and vulnerability. Adjustment for neuroticism eliminated the α-diversity association, implying that personality-linked vulnerability may account for a substantial portion of microbiome–depression covariance; other β-diversity components remained independent of neuroticism and may capture depression-specific microbiota features. The identification of butyrate-producing genera (e.g., Coprococcus, Roseburia) with negative associations to depressive symptoms is consistent with anti-inflammatory mechanisms. Heterogeneous ASV-level effects within genera caution against genus-level aggregation and underscore the value of ASV-resolution analyses and careful taxonomy curation (differences between Greengenes and SILVA assignments affected interpretation for Oscillospira). Overall, the findings strengthen the evidence connecting the gut microbiome with depressive symptoms and point to potential microbial targets for psychobiotic strategies.

In an ethnically diverse urban cohort, gut microbiota diversity and composition were consistently associated with depressive symptom levels, and these associations largely generalized across ethnic groups. β-diversity partly accounted for ethnic disparities in depression. Multiple taxa, predominantly within Firmicutes (including Christensenellaceae, Lachnospiraceae, Ruminococcaceae), were linked to symptoms even after extensive adjustment, highlighting candidate targets (e.g., Coprococcus, Roseburia, Dialister) for future psychobiotic interventions. The study advances the field by applying comprehensive confounder adjustments, cross-validating diversity metrics, and probing ethnic heterogeneity. Future research should incorporate longitudinal designs to infer causality, refine taxonomic/strain-level resolution (e.g., metagenomics), integrate dietary data and other environmental exposures, consider depression subtypes and neuroticism, and perform multi-cohort analyses to improve power and generalizability.

The cross-sectional design precludes causal inference. Depression was assessed via PHQ-9 symptom scores rather than clinical diagnoses, although PHQ-9 correlates strongly with diagnostic outcomes. Diet and certain environmental exposures were not included in the main models and could confound or mediate associations. Potential overadjustment is possible if some covariates lie on the causal pathway. The duration of local −20 °C storage of stool samples was not logged, which could introduce technical variability. Taxonomic assignment differences between databases (Greengenes vs SILVA) and genus-level aggregation can obscure ASV-level heterogeneity. Although antidepressant users were excluded and extensive covariates applied, residual confounding cannot be ruled out. The cohort is from a single urban setting, which may limit generalizability beyond similar environments.