The gut microbiota dietary index reveals a modifiable diet–depression association varying by BMI and possibly linked to inflammation: a hypothesis-generating modified diet for future intervention research

Abstract

In the context of the microbiota-gut-brain axis, diet represents a critical yet still debated component of translational psychiatry. However, population-based evidence linking microbiota-supportive dietary patterns to depressive symptoms remains limited. This research leverages the Dietary Index for Gut Microbiota (DI-GM) to examine the associations between a gut-microbiota-friendly diet and depression, providing a theoretical framework for future research on practical applications. Using data from 28,884 adults, this cross-sectional study evaluated the association between DI-GM and PHQ-9 using survey-compatible descriptive and conventional regression analyses, ordinary-weighted generalized additive models (GAMs), XGBoost with SHAP for exploratory feature ranking, and an exploratory mediation-type analysis of systemic inflammation. Participants with depression had lower DI-GM scores than those without depression. Model-based analyses suggested an inverse predicted pattern between DI-GM and PHQ-9, especially with the interaction GAM indicating significant BMI-related heterogeneity (a more pronounced inverse association at higher BMI). XGBoost with SHAP analyses identified fiber, refined grains, percent energy from fat and whole grains as dietary features with relatively larger contributions to model predictions. Exploratory mediation-type analysis suggested that inflammation may partially explain the DI-GM–PHQ-9 association. However, a high DI-GM diet lacks depression-related nutrients like vitamin D, omega-3 fatty acids, and zinc. Based on these hypothesis-generating findings, we propose a conceptual framework for a Modified Gut Microbiota-Friendly Diet (MGFD) with high DI-GM score, caloric control, and targeted nutrient supplementation, requiring validation via prospective cohorts, randomized trials, and gut microbiome/metabolomic studies.

Introduction

Depression is a leading global health challenge, with increasing evidence implicating dietary factors in its etiology through nutrient, gut microbiota, and inflammatory pathways [1].

Although the ideal antidepressant diet remains debated, certain dietary patterns and nutrients show promise in managing depression. Mediterranean and DASH diets are linked to reduced depression risk, whereas diets rich in refined sugars and saturated fats are associated with increased risk [2]. Specific nutrients such as omega-3 fatty acids, B vitamins, vitamin D, and zinc exert anti-inflammatory and neuroprotective effects [3–5]. Conversely, pro-inflammatory diets may exacerbate neuroinflammation and major depressive disorder [6, 7]. Nevertheless, the strength and clinical relevance of these associations may vary by population, baseline nutritional status, comorbidity profile, and study design.

The gut microbiome is a particularly crucial mediator of these effects through the gut–brain axis [8–10]. Recently, Kase et al. developed the Dietary Index for Gut Microbiota (DI-GM), which quantifies the extent to which dietary patterns support microbial health [11]. This index provides a dietary proxy for studying microbiota-supportive eating patterns in population-based nutritional epidemiology.

Based on the widely proposed diet-gut microbiota-brain pathway, this cross-sectional study investigated the association between DI-GM and depressive symptoms. It evaluated DI-GM score differences across depression severity levels and its predictive value after adjusting for covariates, assessed dietary components’ association strengths with depression severity via machine learning, examined body mass index as a potential moderator, and explored systemic inflammation as a potential correlational pathway. Based on these exploratory findings, we also discuss a hypothesis-generating modified gut microbiota-friendly dietary framework for future intervention research.

Methods

Study design

This study is a nationwide analysis utilizing publicly available data from the National Health and Nutrition Examination Survey (NHANES). NHANES is publicly available, the survey was conducted at multiple mobile examination centers across the United States. We included data from seven continuous cycles (duration 2005–2018), which initially included 71,830 participants. The study design is schematically illustrated in Fig. 1.

Fig. 1. Schematic diagram of the study design and process.

”Background” illustrating the core question of the relationship between diet and depression, with potential mechanisms involving gut-brain axis, neuroinfl ammation, and gut microbiota, and proposing the Dietary Index for Gut Microbiota (DI-GM) as a novel research indicator. ”Methods” (1) descriptive statistics of key variables in the study population; (2) GAM showing the interaction effect of DI-GM on depression (PHQ-9), stratifi ed by BMI; (3) XGBoost combined with SHAP analysis illustrating the relative importance and directional eff ects of dietary components on depression; (4) exploratory mediation framework suggesting that DI-GM may infl uence depression through INFLA. BMI, body mass index; GAM, Generalized Additive Model; XGBoost, eXtreme Gradient Boosting; SHAP, SHapley Additive exPlanations; PHQ-9, Patient Health Questionnaire-9; INFLA, low-grade systemic infl ammation. ”Results” showed an inverse predicted pattern between DI-GM and PHQ-9 modifi ed by BMI, with dietary components exerting diff erential eff ects; INFLA partially mediated this pathway (mediation proportion: 12.1%). ”Conclusion” Higher DI-GM was associated with lower PHQ-9 scores, while high-fat diet increased and fi ber/whole grains decreased depression risk. Higher BMI amplifi ed this association. A new dietary concept (MGFD, modified gut microbiota-friendly diet) is proposed.

Participants

To ensure data integrity and relevance to the study objectives, we excluded participants with (1) incomplete Patient Health Questionnaire-9 (PHQ-9) for depression screening or missing items; (2) missed 24-h dietary recall data for either day 1 or day 2 in the questionnaire; (3) age <20 years; (4) reported implausible daily energy intake ( < 500 kcal or >6000 kcal); (5) had extreme body mass index values (BMI < 15 or > 65 kg/m²);

A final analytical sample of 28,884 adults was retained for statistical analysis, eFigure 1(in the Appendix) outlines the flowchart of Participant Screening.

Participants were categorized into depression (PHQ-9 > 9) and non-depression (PHQ-9 ≤ 9) groups. Following established PHQ-9 cutoffs, participants were categorized into five depression-severity groups: no/minimal (0–4), mild (5–9), moderate (10–14), moderately severe (15–19), and severe (20–27) [12].

Procedures

The primary independent variable in this study is the Dietary Index for Gut Microbiota (DI-GM), a literature-based dietary index developed by Kase et al. through a systematic review of 106 longitudinal studies (102 of which were interventional studies), which quantifies the quality of a diet supportive of gut microbiota [11]. The DI-GM score was significantly positively correlated with urinary enterodiol and enterolactone levels, which are indirect biomarkers of gut microbiota diversity (r = 0.22, p < 0.0001) [11]. Based on these developmental and validation data, the present study employed the DI-GM as an exposure variable to assess the potential impact of diet on gut microbiota. Nevertheless, it is important to note that the DI-GM is a proxy indicator rather than a direct measurement of the gut microbiome.

The index includes 14 food components: Favorable components included fermented dairy, chickpeas, soy products, whole grains, fiber, cranberries, avocado, broccoli, coffee, and green tea; unfavorable components included red meat, processed meat, refined grains, and high-fat diets (40% of total energy derived from fat). Each variable was dichotomized based on sex-specific median intakes. For favorable components, an intake above the median was assigned a score of 1, and 0 otherwise, while unfavorable components were scored inversely. As data on green tea were unavailable, the total possible score ranged from 0 to 13.

Dietary intake was assessed using two interviewer-administered 24-h dietary recalls from the What We Eat in America component of NHANES. The recalls were conducted by trained dietary interviewers using the USDA Automated Multiple-Pass Method (AMPM), a standardized five-step, computer-assisted approach designed to improve completeness and accuracy of 24-h dietary recall data. In the NHANES cycles included in this study, the first dietary recall was collected in person in the Mobile Examination Center, and the second recall was collected by telephone 3–10 days later on a nonconsecutive day. NHANES dietary recalls were collected across weekdays and weekend days, and recall quality was monitored using dietary recall status codes indicating completeness and reliability. Nutrient and food group assignments were based on the USDA Food and Nutrient Database for Dietary Studies (FNDDS). We used the average intake from the two recalls to construct the DI-GM score.

Depression severity was assessed using the Patient Health Questionnaire-9 (PHQ-9), which is a validated screening tool consistent with the DSM-5 criteria. The total score ranges from 0 to 27, with higher scores indicating more severe symptoms.

Based on literature and subject-matter relevance, we included the following covariates to adjust for potential confounding: sex (male/female), age (years), education level (less than high school, high school, some college, college or above), marital status (married, widowed, divorced, separated, never married, Living with partner), body mass index (BMI in kg/m²), ethnicity (Mexican American, other Hispanic, non-Hispanic White, non-Hispanic Black, non-Hispanic Asian, other race/multiracial), poverty income ratio, smoking status (never, former, current), current drinking (yes/no), diabetes (yes/no), cardiovascular disease history (yes/no), supplement use (yes/no), and antidepressant use in the past 30 days (yes/no).

Missing values in covariates were addressed using multiple imputation. Analyses were performed separately in each imputed dataset, and model estimates were combined using Rubin’s rules. Unless otherwise specified, all reported statistical results, including model estimates, standard errors, 95% confidence intervals, and P values, are pooled estimates based on the multiply imputed datasets.

Statistical analysis

All analyses were conducted in Python 3.11. Survey-compatible descriptive and conventional regression analyses incorporated NHANES dietary weights, strata, and primary sampling units, whereas more complex models, including GAM, nutrient GAM, XGBoost/SHAP, and mediation analyses, used ordinary weighted approaches for model stability. This strategy was adopted to reduce instability caused by applying full complex survey design estimation to highly parameterized nonlinear, interaction, or machine-learning models. These models therefore used dietary sampling weights as analytical weights and adjusted for the same core covariates.

GAMs were used to examine the nonlinear association between DI-GM and PHQ-9 scores:

$\mathit{PHQ}-9_i={\beta }_0+f(\mathit{DI}-G{M}_i\mathit{df}=5)+{\gamma }^T{C}_i+{\epsilon }_i$, where $C_i$ represents the vector of core covariates. PHQ-9 was treated as a continuous outcome, and the GAM was fitted under a Gaussian error distribution with an identity link.The 5-df spline was the default spline specification used in our GAM implementation.

To explore the relative model-based associations of dietary components and potential subgroups, we performed interaction analysis using GAM with a DI-GM×BMI tensor product interaction term and feature importance analysis using XGBoost with SHAP interpretation on an 80/20 stratified split. For the XGBoost analysis, the predictor set included the 13 dietary components constituting the DI-GM together with the covariates used in the main model. The dataset was split into training and test sets at an 80:20 ratio using stratification by depression status. The NHANES two-day dietary recall weight was incorporated as an analytical sample weight during model fitting. Model performance was evaluated on the held-out test set using the weighted test-set R², which was 0.1509. SHAP values were calculated to provide an exploratory ranking of the dietary components. Because the model was intended for exploratory interpretation rather than clinical prediction, SHAP results were interpreted cautiously as model-based association patterns.

The interaction model extended the primary GAM specification:

$$\mathit{PHQ}-9_i={\beta }_0+f_1(\mathit{DI}-G{M}_i)+f_2\left(\mathit{BM}{I}_i\right)+\mathit{te}(\mathit{DI}-G{M}_i,\mathit{BM}{I}_i)+{\gamma }^T{C}_i+{\epsilon }_i$$

where f₁ and f₂ denote univariate smooth terms for DI-GM and BMI respectively, while te is a tensor product smooth capturing their interaction.

In addition, we conducted an exploratory mediation-type analysis using an ordinary-weighted three-model framework to examine whether INFLA-score represented a possible inflammation-related indirect association between DI-GM and PHQ-9. Models were fitted separately in each imputed dataset and pooled using Rubin’s rules. Bootstrap P values were obtained using 1000 resamples within each imputation. The INFLA-score [13] is a composite indicator of low-grade systemic inflammation, constructed from four circulating biomarkers: C-reactive protein (CRP), white blood cell count (WBC), platelet count (PLT), and neutrophil-to-lymphocyte ratio (NLR). The construction method is described in eMethods 5 in the Appendix. A schematic of the exploratory ordinary-weighted three-model framework is available in eFig. 2 in the Appendix.

Fig. 2. Predicted PHQ-9 scores from the ordinary-weighted GAM across DI-GM scores.

Note.Covariates were held at weighted reference values. The shaded area indicates the 95% confidence interval of the model-based prediction.

Results

Baseline characteristics

Weighted analysis of 28,884 NHANES participants represented 214,889,161 U.S. adults, with 16,300,326 (7.59%) classified as having depression and 198,588,834 (92.41%) as non-depressed.

Baseline characteristics by depression status are shown in Table 1.

Table 1.

Baseline Characteristics by Depression Status.

Characteristic	Non-depression	Depression	P value
DI-GM score	4.52 (4.49, 4.56)	4.15 (4.06, 4.25)	<0.001
PHQ-9 score	1.99 (1.94, 2.03)	13.97 (13.74, 14.20)	<0.001
Age, years	47.13 (46.82, 47.44)	46.02 (45.07, 46.97)	0.003
Poverty income ratio	3.11 (3.08, 3.14)	2.12 (2.02, 2.21)	<0.001
BMI, kg/m2	28.18 (28.06, 28.29)	29.99 (29.53, 30.45)	<0.001
Sex
Male	49.0 (48.1, 50.0)	38.0 (35.0, 41.0)	<0.001
Female	51.0 (50.0, 51.9)	62.0 (59.0, 65.0)
Education
<9th grade	3.9 (3.5, 4.2)	7.0 (5.4, 8.6)	<0.001
9–11th grade	9.4 (8.9, 10.0)	16.1 (13.8, 18.3)
High school/GED	22.8 (22.0, 23.6)	28.7 (25.9, 31.5)
Some college	31.7 (30.9, 32.6)	34.6 (31.6, 37.5)
College graduate or above	32.2 (31.4, 33.1)	13.7 (11.6, 15.8)
Marital status
Married	57.2 (56.3, 58.1)	38.2 (35.2, 41.2)	<0.001
Widowed	4.9 (4.5, 5.3)	6.5 (5.0, 8.0)
Divorced	9.8 (9.2, 10.3)	17.8 (15.4, 20.2)
Separated	2.00 (1.70, 2.20)	6.00 (4.90, 7.00)
Never married	18.30 (17.10, 19.50)	22.30 (19.90, 24.70)
Living with partner	7.90 (7.20, 8.50)	9.30 (7.70, 10.90)
Race/ethnicity
Mexican American	8.2 (7.7, 8.7)	8.6 (6.8, 10.3)	<0.001
Other Hispanic	5.3 (4.9, 5.7)	8.0 (6.3, 9.7)
Non-Hispanic White	68.9 (68.0, 69.8)	64.5 (61.5, 67.4)
Non-Hispanic Black	11.7 (11.1, 12.3)	15.1 (12.9, 17.3)
Non-Hispanic Asian	2.4 (2.0, 2.8)	2.3 (1.2, 3.5)
Other/multiracial	3.5 (3.0, 4.1)	1.5 (1.0, 2.0)
Smoking status
Never	57.1 (56.2, 58.0)	37.2 (34.2, 40.2)	<0.001
Former	25.0 (24.2, 25.8)	21.1 (18.6, 23.6)
Current	18.0 (17.1, 18.8)	41.7 (38.2, 45.2)
Current drinking
No	22.3 (21.6, 23.1)	19.6 (17.2, 22.1)	<0.001
Yes	77.7 (76.9, 78.4)	80.4 (77.9, 82.8)
Diabetes
No	91.1 (90.6, 91.6)	85.4 (83.2, 87.6)	<0.001
Yes	8.9 (8.4, 9.4)	14.6 (12.4, 16.8)
Cardiovascular disease
No	92.5 (92.0, 92.9)	84.1 (81.8, 86.4)	<0.001
Yes	7.5 (7.1, 8.0)	15.9 (13.6, 18.2)
Supplement use
No	44.5 (43.6, 45.5)	49.2 (46.1, 52.3)	<0.001
Yes	55.5 (54.5, 56.4)	50.8 (47.7, 53.9)
Antidepressant use in past 30 days
No	88.8 (88.3, 89.4)	63.5 (60.5, 66.5)	<0.001
Yes	11.2 (10.6, 11.7)	36.5 (33.5, 39.5)

Values are survey-weighted means or percentages with 95% confidence intervals. P values were obtained from design-adjusted tests accounting for NHANES dietary weights, strata, and primary sampling units.

Compared with participants without depression, participants with depression had lower DI-GM scores (4.15 vs 4.52), higher PHQ-9 scores (13.97 vs 1.99), higher BMI (29.99 vs 28.18 kg/m2), and lower poverty income ratio (2.12 vs 3.11), Across PHQ-9 severity groups, survey-weighted DI-GM means were 4.55, 4.40, 4.21, 4.02, and 4.11 from no/mild through severe depression, as illustrated in eFigure 3and eTable 1(in the Appendix).

Nonlinear association between the DI-GM and depressive symptoms

The ordinary-weighted GAM showed an inverse fitted pattern between DI-GM and PHQ-9. However, the overall DI-GM spline block was not statistically significant in the MI D1 test (F = 1.743, df = 5 and 268467.6, P = 0.121), and model fit was modest (pseudo R² = 0.1515). Thus, this result should be interpreted as a descriptive model-based trend rather than a significant nonlinear association. Complete coefficients and overall tests are shown in eTable 2 and eTable 3. Several covariates were independently associated with PHQ-9 scores. Female sex, higher BMI, current drinking, diabetes, cardiovascular disease, smoking status, antidepressant use, and several non-married marital-status categories were associated with higher PHQ-9 scores, whereas older age, higher poverty income ratio, and the highest educational attainment category were associated with lower PHQ-9 scores.

Figure 2 shows predicted PHQ-9 values across the DI-GM range, holding other covariates constant at their weighted mean (continuous variables) or weighted mode (categorical variables) values. The shaded area represents the 95% confidence interval of the model-based prediction. A general downward trend was observed, suggesting that higher DI-GM scores (i.e., more microbiota-friendly dietary patterns) correspond to lower depressive symptom burden. However, the effect magnitude was relatively small: the change in predicted PHQ-9 scores across the full DI-GM range was about 1.12 points. Because the overall DI-GM spline block was not statistically significant, Fig. 2 should be interpreted as a descriptive model-based visualization rather than confirmatory evidence of a significant DI-GM association.

Joint association of DI-GM and BMI on depression severity

We used an interaction GAM to examine whether the association between DI-GM and PHQ-9 varied across BMI levels. The interaction model included smooth terms for DI-GM and BMI and a tensor-product smooth term for their interaction. The smooth-term statistics indicated significant nonlinear smooth effects of DI-GM(edf = 11.87, test statistic = 71.07, P < 0.001), BMI (edf = 15.33, test statistic = 93.71, P < 0.001), and the DI-GM × BMI interaction term (edf = 20.05, test statistic = 196.54, P < 0.001). These results suggest BMI-related heterogeneity in the model-based association between DI-GM and PHQ-9. The corresponding smooth-term statistics and BMI-stratified predicted contrasts are provided in eTable 4 and eTable 5.

Figure 3 presents the predicted PHQ-9 surface, contour plot, and BMI-stratified prediction curves. For visual clarity, confidence bands are not displayed in the BMI-stratified curves in Fig. 3C; the corresponding BMI-specific uncertainty estimates are reported in eTable 5. To quantify the BMI-specific patterns, we estimated model-predicted PHQ-9 contrasts between DI-GM = 0 and DI-GM = 10 at fixed BMI levels of 25, 30, 35, and 40 kg/m². The model-predicted PHQ-9 contrasts were 1.594 points at BMI 25, 1.164 points at BMI 30, 1.610 points at BMI 35, and 2.325 points at BMI 40. Thus, the inverse association between DI-GM and PHQ-9 varied across BMI levels and showed the largest model-predicted contrast at BMI 40.

Fig. 3. Joint association of DI-GM and BMI with predicted PHQ-9 scores.

Predicted PHQ-9 scores were generated from the ordinary-weighted generalized additive model with a DI-GM × BMI interaction term. A Three-dimensional surface showing predicted PHQ-9 values as a function of DI-GM and BMI. B Contour plot of predicted PHQ-9 values from the DI-GM × BMI interaction model. C Predicted PHQ-9 scores plotted against DI-GM at fixed BMI levels of 25, 30, 35, and 40 kg/m². Note. For visual clarity, confidence bands are not shown in panel C;BMI-specific uncertainty estimates are reported in eTable 5 in .Supplementary data.

However, the BMI-specific pattern was not strictly monotonic, because the predicted contrast at BMI 30 was smaller than those at BMI 25 and BMI 35. In addition, the confidence intervals for the fixed-BMI predicted contrasts were relatively wide. Therefore, these results should be interpreted as model-based evidence of BMI-related heterogeneity rather than as definitive subgroup effects.

Exploratory SHAP-based feature ranking of dietary components

To evaluate the relative contribution of DI-GM dietary components while accounting for the expanded covariate set, we fitted a weighted XGBoost model including the 13 food items constituting the DI-GM and the covariates used in the main model, with PHQ-9 score as the outcome variable. The model achieved a modest weighted test-set performance (R² = 0.1509), indicating that the model explained a limited but non-negligible proportion of the variation in PHQ-9 scores. SHAP-based feature ranking identified dietary fiber, refined grains, percent energy from fat, and whole grains as the dietary features most strongly associated with PHQ-9 scores in this exploratory model. Higher intake of fiber and whole grains was associated with lower predicted PHQ-9 scores, whereas refined grains and a higher percent of energy from fat were associated with higher predicted PHQ-9 scores (Fig. 4). These findings should be interpreted as SHAP-based model-explanation patterns in cross-sectional data and should not be interpreted as causal effects of dietary components.

Fig. 4. Exploratory SHAP-based feature ranking and associations of dietary components with predicted PHQ-9 scores in the weighted XGBoost model.

A SHAP feature importance ranking showing the mean absolute SHAP values of each dietary component, withhigher values indicating greater contribution to the model prediction. B SHAP summary (beeswarm) plot illustratingthe directional relationship between each dietary feature and its impact on predicted PHQ-9 scores; the color scalerepresents the feature value from low (blue) to high (red), with positive SHAP values indicating increased depressionrisk and negative values indicating protective effects.

Exploratory mediation-type analysis of systemic inflammation between DI-GM and PHQ-9

Exploratory mediation-type analysis suggested a possible inflammation-related indirect association between DI-GM and PHQ-9 through INFLA-score. As shown in eTable 6 in the Appendix, the estimated possible indirect association was −0.0031 (95% CI: −0.0053 to −0.0009; Rubin P = 0.005), with an exploratory indirect-to-total ratio of 12.1%. However, the bootstrap P value for this indirect association was 0.058, and the total association was not statistically significant. Therefore, these findings should be interpreted cautiously as exploratory evidence of a possible indirect association through systemic inflammation, rather than as confirmatory evidence of causal mediation.

Discussion

The optimal dietary pattern for depression prevention remains debated. While some cohort studies have reported no significant association between Mediterranean diet adherence and depression risk [14], the “SMILES” trial found that a modified Mediterranean diet improved MADRS depression scores [15]. A 2024 study linked adherence to the EAT-Lancet diet with reduced risks of depression, anxiety, and comorbidities [16]. Additionally, intake of viable dietary microorganisms has been associated with lower depression risk [17]. Recent work on the gut–brain axis [18] further supports diet-induced microbiota modulation as a promising strategy for mood disorder management [19]. Some studies have demonstrated a close bidirectional relationship between gut microbiota and depression, primarily mediated by the microbiota-gut-brain axis; Patients with depression commonly exhibit reduced gut microbiota diversity and alterations in the abundance of specific microbial taxa, while microbiota-targeted interventions (e.g., probiotics and synbiotics) may ameliorate depressive symptoms [20, 21]. Although the DI-GM used in this study is only an indirect indicator of gut microbiota health, this study extends this body of evidence by examining the relationship between DI-GM, a dietary proxy of microbiota-supportive eating, and depression severity (PHQ-9), yielding the following main findings.

Main findings

Although the overall DI-GM spline term did not reach statistical significance, univariate analysis showed that participants with depression had lower DI-GM scores and a negative trend was observed between DI-GM scores and the severity of depressive symptoms. The absolute reduction in PHQ-9 scores across the full DI-GM range was only 1.12 points, yet this still corresponded to a meaningful 8.0–22.4% relative reduction among individuals with mild-to-moderate depression (PHQ-9 scores ranged from 5–14). Given the relatively low model fit (pseudo R² = 0.1515), we are more inclined to interpret this result as an exploratory descriptive trend. Future studies with larger sample sizes or more precise methods for nutritional exposure assessment are needed to verify whether a high DI-GM diet truly exerts linear or nonlinear effects on depression independent of other factors.

Compared with the model that only focused on the main effect of predicted PHQ-9 scores by DI-GM (GAM), the introduction of the BMI interaction term revealed a pattern with research relevance for future longitudinal or interventional studies (compared with other socioeconomic covariates, BMI is more modifiable). As shown in Fig. 3, the inverse relationship between DI-GM and PHQ-9 varied across BMI levels and showed the largest model-predicted contrast at BMI = 40. The interaction GAM suggested that the DI-GM–PHQ-9 pattern varied by BMI. BMI-stratified analysis (eTable 5) showed that increasing DI-GM from 0 to 10 was associated with reduced PHQ-9 across all BMI levels, with absolute reductions ranging from 1.16 (BMI = 30) to 2.33 (BMI = 40). Average marginal effects per unit DI-GM ranged from –0.116 to –0.233 (relative: –3.21––4.96%). The largest model-predicted contrast was observed at BMI = 40 (PHQ-9 reduction: 5.27–2.94). Taken together, these results suggest BMI-related heterogeneity in the inverse model-based association. This suggests that the inverse association between a microbiota-supportive diet and depressive symptoms may be more pronounced in individuals with higher BMI, particularly at BMI = 40 in the present model-based analysis.The bidirectional relationship between obesity and depression may also explain this finding: obesity confers a 55% increased risk of incident depression, and depression confers a 71% increased risk of incident obesity [22]. Further analysis (Fig. 3C) showed that increasing DI-GM from 0 to 10 was associated with a model-predicted reduction in PHQ-9 scores from approximately 5.27 at BMI = 40 to below 2.36 at BMI = 25, a total visual model-based difference of approximately 2.9 points ( > 50%), numerically larger than some associations reported in other similar cross-sectional studies [23–25]. This margin also exceeds the main model’s 1.12 points, although it should be interpreted cautiously as an exploratory model-based contrast rather than evidence of clinical benefit. This also may inform the design of future interventional studies examining depressive symptom severity by concurrently optimizing gut-microbiota-friendly diet and considering energy-balance or weight-management components.

Explainable contributions of dietary features to depression

To further elucidate the role of individual food components within the DI-GM, we conducted further investigations using XGBoost-based SHAP feature attribution analysis. The results indicated that the model explained a limited but non-negligible proportion of the variation in PHQ-9 scores. This modest performance is also consistent with the multifactorial nature of depressive symptoms, which are influenced by psychosocial, socioeconomic, behavioral, clinical, and biological factors beyond diet. Therefore, the following results are interpreted as exploratory feature-ranking results: dietary fiber, refined grains, percent energy from fat, and whole grains were the dietary features most strongly associated with PHQ-9 scores. As shown in Fig. 4 and Table 2, dietary fiber and whole grains showed negative associations, whereas refined grains and percent energy from fat showed positive associations. Broccoli intake demonstrated moderate negative associations, whereas refined grains, processed meat, and red meats emerged as moderate positive association. Although cross-sectional studies cannot elucidate mechanisms, numerous existing reports in the literature have clearly demonstrated the effects of the above dietary components: Dietary fiber and resistant starch act as prebiotics that promote short-chain fatty acid (SCFA) production, mitigating depressive behaviors by inhibiting pro-inflammatory cytokines and modulating immune function [26]. Energy intake from fat is a more complex issue. While n-3 PUFAs (e.g., EPA, DHA) have strong anti-inflammatory properties [27], n-6 PUFAs (e.g., linoleic acid) are pro-inflammatory [28]. The typical Western diet exhibits elevated n-6/n-3 ratios (≥ 15:1),far exceeding recommended levels (<4:1) [29], thus creating a pro-inflammatory internal milieu, which is closely associated with the development of depression. In addition, excess saturated fats induce gut dysbiosis, further exacerbating inflammation and potentially worsening depression [30].

Table 2.

Exploratory SHAP-based associations for dietary features in the weighted XGBoost model.

Feature	Importance Level	Direction of Association
Fiber	High (mean |SHAP | = 0.1043)	Negative
Refined grains	High (mean |SHAP | = 0.0709)	Positive
Percent energy from fat	High (mean |SHAP | = 0.0651)	Positive
Whole grains	High (mean |SHAP | = 0.0634)	Negative
Processed meat	Medium (mean |SHAP | = 0.0509)	Positive
Broccoli	Medium (mean |SHAP | = 0.0425)	Negative
Fermented dairy products	Medium (mean |SHAP | = 0.0332)	Mixed
Soy products	Medium-low (mean |SHAP | = 0.0286)	Negative
Red meat	Low (mean |SHAP | = 0.0242)	Positive
Coffee	Low (mean |SHAP | = 0.0228)	Mixed
Avocado	Low (mean |SHAP | = 0.0110)	Mixed
Cranberry	Low (mean |SHAP | = 0.0025)	Mixed
Chickpeas	Low (mean |SHAP | = 0.0005)	Mixed

Importance levels are based on mean absolute SHAP values from the weighted XGBoost model. Direction of association refers to the direction of the SHAP-based model-explanation pattern with predicted PHQ-9 scores.

The moderate protective effect of broccoli may be attributed not only to its fiber content but also to bioactive compounds such as sulforaphane, which alleviates depression-like behaviors by regulating the composition of gut microbiota, enhancing the expression of intestinal tight junction proteins, and inhibiting the inflammatory response mediated by the NF-κB/NLRP3 pathway, while also upregulating multiple antioxidants and protecting neurons against various oxidative damages, thereby increasing brain levels of brain-derived neurotrophic factor, serotonin, and dopamine [31, 32]; By contrast, High-GI foods (such as sugary snacks and refined grain) trigger dramatic fluctuations in insulin levels, increasing physical stress, and research shows that they are associated with an elevated risk of depression [33]; Additionally, refined grains lose key nutrients such as dietary fiber, phenolic compounds, and B vitamins, which may deprive beneficial gut bacteria of the “fuel” they need, leading to gut dysbiosis and consequently affecting mood [34]. Processed and red meats contain nitrates, nitrites, and AGEs, all of which are pro-inflammatory [35]. Additionally, red meat-derived TMAO has been linked to systemic inflammation and depression [36]. Meta-analysis (including 17 original studies), based on data from 241,738 participants, showed that red and processed meat intake was significantly associated with the risk of depression (OR = 1.08, 95% CI: 1.04–1.12, p < 0.001), corresponding to an approximately 8% increased risk [37]. These in-depth analyses provide theoretical support for optimizing DI-GM-related dietary patterns and future diet-depression studies.

Inflammatory mediation and exploratory analyses

To explore whether systemic inflammation may represent a possible indirect associational link between DI-GM and PHQ-9 scores, we conducted an exploratory mediation-type analysis using INFLA-score. The INFLA-score is a validated composite indicator of low-grade systemic inflammation that has been associated with depression-related outcomes and shown to mediate the relationship between diet and depressive symptoms in large population-based studies [38]. Exploratory mediation-type analysis suggested a possible inflammation-related indirect association. The estimated indirect-to-total ratio was 12.1%, but the bootstrap P value for the indirect association was 0.058 and the total association was not statistically significant. Meanwhile, given the limitations of cross-sectional mediation analysis and the lack of temporal sequence rather than a causal pathway [39], inflammation should be interpreted as a potential associational link. Future longitudinal studies with repeated measures of diet, inflammatory markers, microbiome, and depressive symptoms are needed to assess whether inflammation may play an indirect or mediating role in longitudinal settings. The remaining unexplained effects may involve endocrine, immune, psychological, or behavioral factors, which could not be fully examined due to database limitations and warrant investigation in future cohort studies.

DI-GM and antidepressant nutrients

Despite the observed association between DI-GM and PHQ-9, the research potential, limitations, and need for further exploration of a high DI-GM diet remain to be clarified. We also considered the potential relevance of key nutrients to depression-related nutritional research. According to the 2021 clinical nutrition guidelines [40], four nutrients are at least weakly recommended[+] for depression prevention or treatment, supported by meta-analyses or ≥2 randomized controlled trials: omega-3, vitamin D, folate, and zinc. In this study, we also investigated the relationship between these nutrients and DI-GM. GAM-fitted curves (eFigure 4in the Appendix) showed that as DI-GM increased, intake of vitamin D and folate also rose, suggesting that a high-DI-GM dietary pattern may be accompanied by higher intake of these depression-related vitamins. However, even among individuals in the high DI-GM group≥6, mean vitamin D intake remained below 10 μg (400 IU)—well under the Dietary Guidelines for Americans (DGA) daily nutritional goal of 15 μg (600 IU) [41]. Thus, before a high-DI-GM dietary pattern can be further evaluated as a candidate dietary framework for depression-related intervention research, its shortfall in vitamin D should be considered and addressed. Omega-3 fatty acids, particularly EPA and DHA, carry the strongest antidepressant recommendation (+++) [40]. As DI-GM increased from 0 to 10, EPA and DHA intake increased but remained far below the FAO’s lower adequacy limit (250 mg/day) [42]. By contrast, alpha-linolenic acid (ALA) intake fluctuated but stayed above the DGA recommendation (1.3 g/day).However, in terms of total omega-3 fatty acids, those with moderately high scores may still have inadequate intake, underscoring the need for further dietary optimization. Zinc intake exhibited a different trend: as DI-GM increased, dietary zinc fluctuated and decreased slightly. The predicted zinc intake in eFigure 4,B in the Appendix mostly failed to meet the DGA recommendation of 11 mg. This likely reflects the plant-centered nature of a high DI-GM diet, which may reduce the intake of highly bioavailable zinc from animal sources. Therefore, while increasing DI-GM, the potential for inadequate zinc intake should also be considered. These observations indicate that in research associating a microbiota-supportive dietary pattern with depression, evaluation should not rely solely on the DI-GM score, but should also integrate more comprehensive indicators of nutritional adequacy for a holistic judgment.

Shortcomings of high DI-GM diet and Recommendations for Improvement

Although increasing DI-GM correlated with reduced PHQ-9 scores, the maximum observed reduction was approximately 1.12 PHQ-9 points. By contrast, concurrent reductions in BMI yielded a substantially greater decrease, suggesting that calorie restriction should accompany high DI-GM dietary practices. The high DI-GM diet still has the problem of inadequate intake of vitamin D, omega-3 fatty acids, and zinc. Meanwhile, high DI-GM tends to involve a higher proportion of plant-based diets, and high levels of phytic acid may hinder zinc absorption—all of which could undermine its potential to be considered an antidepressant diet worthy of further study. Therefore, our findings emphasize that when evaluating an antidepressant dietary pattern with research potential through cross-sectional studies, it is necessary not only to consider the direct association between the dietary pattern and depressive symptoms, but also to pay attention to the comprehensiveness of nutrients and strategies to mitigate potential limitations. Based on the core content of the discussion section, to enable the DI-GM diet to be included in the research scope of antidepressant diets and to better realize its potential, we propose a modified model, namely the modified gut microbiota-friendly diet (MGFD) model. This model is characterized by a high DI-GM score, calorie control (to maintain a normal body mass index), vitamin D supplementation, and supplementation of zinc and omega-3 fatty acids (or increased intake of foods rich in both, such as fatty fish, shellfish).

Limitations and future directions

First, despite being based on the established “diet-gut microbiota-brain” framework and utilizing advanced statistics, the cross-sectional design limits causal conclusions. Second, the DI-GM is a proxy indicator, not a direct measure of the gut microbiome, which introduces limitations. Moreover, its component scores were dichotomized based on NHANES-derived sex-specific medians, making the scoring criteria population-dependent and limiting cross-population comparability. Future studies should recalibrate these medians within target populations or adopt continuous scoring methods (e.g., quartiles, z-scores, energy-adjusted residuals) or absolute thresholds to enhance broad applicability. Third, dietary assessment was based on two 24-h recalls, which may introduce recall bias and cannot fully capture usual intake, day-to-day variation, weekday/weekend differences, or seasonal dietary variation. Fourth, unmeasured confounders such as genetic susceptibility and psychosocial stress may have influenced the findings. Nonetheless, these findings establish a theoretical framework for future interventional studies, such as randomized controlled trials. Future research may further deepen the mechanistic understanding of relevant biological pathways by integrating multi-omics data, including BMI, gut microbiome sequencing, metabolomics, and inflammatory proteomics.

Conclusion

This study found that higher DI-GM showed an inverse but modest model-based pattern with PHQ-9 depressive symptom scores, particularly among individuals with higher BMI. Exploratory analyses identified several dietary components with relatively larger contributions to model predictions and suggested a small possible inflammation-related indirect association. Accordingly, we propose a modified gut microbiota-friendly dietary framework emphasizing a higher DI-GM pattern, energy balance, and nutrient adequacy, which may help guide future intervention research.

Author contributions

Lin Kong and Yingting Luo: Conceptualization, Study design, Writing-Original Draft. Lin Kong:Funding acquisition. Yingting Luo: Supervision, Project administration. Wenyue Ma and Jiani Wang: Formal analysis, Data Curation, Software. Tingting Liu, Jingqi Yang, and Qinmiao Yang: Investigation, Resources, Methodology. Ziyi Li: Validation, Visualization. Yongfang Liu: Writing-Review & Editing. All authors have read and approved the final version of the manuscript.

Funding

This work was supported by Opening Project of Key Laboratory of Public Nutrition and Health, National Health Commission of the People’s Republic of China (No. WLKFZZ202408).

Data availability

Data described in the manuscript are publicly available from the National Health and Nutrition Examination Survey (NHANES) at https://www.cdc.gov/nchs/nhanes/.Analytic code will be made available by contacting Prof. Luo (e-mail: lyt83@163.com).

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

Ethical approval for NHANES was granted by the institutional review board and the NCHS Research Ethics Review Board, and all participants gave prior informed consent. Since the authors utilized only anonymized, publicly accessible data, direct consent was not required. All methods were implemented in full compliance with relevant ethical regulations.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41398-026-04346-2.