Gut microbiota composition outperforms chronological age in predicting systemic inflammation and incident disease risk

Abstract

Low-grade systemic inflammation increases with age and accompanies many age-associated lifestyle-related diseases, yet substantial inter-individual variability exists beyond chronological age. Given the gut microbiota’s role in immune regulation, we analyzed paired fecal microbiota profiles, 30 plasma cytokines, and 19 biochemical and physiological measures in 1,199 adults (aged 20-72) from the population-based DanFunD cohort to assess the relative associations of microbiota composition and age with systemic inflammation and subsequent disease risk. Here we show that despite an association between age and 40% of inflammatory cytokines and most physiological variables, gut microbiota composition explained more variance than age for 97% of cytokines and 84% of physiological variables. Individuals with the Bacteroides 2 enterotype exhibited elevated pro-inflammatory cytokines and adverse metabolic profiles already in early adulthood, displaying an inflammaging-like phenotype up to 37 years earlier than other enterotypes. This enterotype was associated with increased risk of future incident disease across multiple organ systems (hazard ratio 1.21), with mediation analyses suggesting that inflammatory cytokines accounted for part of this association. Conversely, a high-diversity enterotype was linked to lower inflammation and reduced disease risk. These findings position gut microbiota composition as a stronger correlate of systemic inflammatory tone than chronological age and a potentially modifiable determinant of inflammation-related disease trajectories across adulthood.

Here, characterizing 1,199 adults, the authors find that gut bacteria patterns predict inflammation and future disease risk better than age, with an identified microbiota type showing signs of “inflammaging” decades earlier, while a diverse microbiota is linked to better health.

Introduction

Lifestyle-related diseases such as obesity, type 2 diabetes, and cardiovascular disease increase markedly with age^1,2 and share a common feature of persistent low‑grade systemic inflammation³. Low-grade systemic inflammation has repeatedly been associated with elevated circulating levels of a few markers including C-reactive protein (CRP), interleukin (IL)-6 and tumor necrosis factor (TNF)-α. Other inflammatory cytokines are also reported to increase by age, including IL-1β, IL-15, IL-18, chemokine (C-X-C motif) ligand (CXCL)8, C-C motif chemokine (CCL)11, and CCL2^4–6, while IL-4, IL-10, and transforming growth factor (TGF)-β decrease by age⁷. Inflammaging has been used to describe the shift toward immune imbalance, often precipitating at the point where ageing and age-related diseases converge⁸. Inflammaging is also assumed to reflect poorly balanced interactions between pro- and anti-inflammatory cytokines, resulting in a systemic state of immune dysregulation associated with increased risk of disease and mortality^4,9.

Multiple factors have been proposed to drive inflammaging, including genetics, lifestyle, metabolic alterations, and gut microbiota composition^9,10, as well as processes such as impaired clearance of cellular debris and accumulation of misfolded or oxidized proteins^11,12. Because the gut microbiota modulates host metabolism and immunity^13,14, and reflects both life‑history exposures¹⁵ and longevity¹⁶, it has been suggested as a potentially central contributor in inflammaging. Little is, however, known about its actual contribution to inflammaging, and whether age itself, gut microbiota, or both play a major role in inflammaging.

Here, using samples from 1199 deeply phenotyped adults from a population-based cohort of 7493 individuals¹⁷, we investigated how age and gut microbiota composition relate to 30 plasma cytokines and 19 biochemical and physiological variables relevant to the development of lifestyle-associated diseases. Through integrative analyses, we assessed whether specific gut microbiota configurations corresponded to inflammaging‑like inflammatory profiles independent of chronological age. By linking microbiota enterotypes to International Classification of Diseases (ICD)‑10-coded hospital diagnoses during periods of up to 9 years of follow‑up, we evaluated how inflammatory and microbial characteristics related to subsequent disease risk and to what extent our findings were reproduced in the independent MetaCardis cohort.

Results

Comparative analysis of age- and microbiota-linked determinants of systemic inflammatory cytokines

Data were collected from 7493 adults aged 18–76 years enrolled in the population-based DanFunD cohort from the greater Copenhagen area¹⁸. On the examination day, participants underwent physiological testing, including a step test for estimating respiratory fitness (VO₂ max)¹⁹, and provided blood samples for measurement of biochemical markers relevant to cardiometabolic, renal, and hepatic function (“Methods”; Supplementary Tables 1 and 2), yielding 19 variables for analysis. Plasma levels of 30 immune‑related cytokines were quantified in a subsample of 4101 participants, and fecal samples for gut microbiota profiling were obtained from 2,460 participants, resulting in 1199 individuals with paired microbiota and cytokine data (Fig. 1a and Supplementary Tables 3 and 4, Supplementary Fig. 1).

Fig. 1. Gut microbiota composition explains more variance in systemic inflammatory cytokine levels than age.

a Study design: a total of 7493 individuals participated in the population-based DanFunD cohort. 2460 provided stool samples for gut microbiota profiling, and plasma levels of 30 cytokines were quantified from 4101 individuals. Paired microbiota-cytokine data were available for 1199 individuals included in the present analyses. Volcano plots showing age associations for 30 systemic cytokines (b) and 19 physiological/biochemical variables (c) using generalized linear models. Adjusted effect sizes represent sex‑, antibiotic‑, and smoking‑adjusted z‑values. Dot size reflects significance (−log₁₀(p_adj = 0.05)), and the horizontal dashed line marks the threshold at −log₁₀(p_adj = 0.05). Labels are shown for all statistically significant associations. All statistical tests were two-sided. d Variance in each plasma cytokine (top) and physiological/biochemical variable (bottom) explained by age (yellow), gut microbiota composition (blue), or both combined (red). Microbiota composition was represented by principal coordinates (PCoA on Bray-Curtis dissimilarity), capturing ~80% of compositional variance to minimize inflation due to correlations among taxa. Variance explained was calculated by summing contributions from significant coordinates (p < 0.05) in permutational ANOVA models with and without age. b–d include data from 1199 individuals with paired microbiota and cytokine measurements. Source data are provided as a source data file. ALT alanine aminotransferase; BDNF brain-derived neurotrophic factor; BP blood pressure; CCL C-C motif chemokine; CRP C-reactive protein; CXCL chemokine (C-X-C motif) ligand; eGFR estimated glomerular filtration rate; HbA1c glycated hemoglobin; HDL high-density lipoprotein; IFN interferon; IL interleukin; LDL low-density lipoprotein; MIF macrophage migration inhibitory factor; TAG triacylglycerol; TNF tumor necrosis factor; VEGF vascular endothelial growth factor.

We first examined how plasma cytokines and biochemical and physiological variables varied with age. Twelve cytokines (CXCL8, IL-6, TNF-α, CCL11, CXCL10, IL-15, CCL2, VEGF, CCL17, CRP, IFN-γ, and IL-5) increased with age, whereas four (IL-22, IFN-β, IFN-α2a, and IL-31) decreased (for all tests, p_adj < 0.05; Fig. 1b and Supplementary Table 5). The remaining 14 (47%) cytokines showed no age association.

Among biochemical and physiological measures, 16 of 19 variables, including VO₂ max, eGFR, albumin, ALT, calcium, carbamide, TAG, HDL‑ and LDL‑cholesterol, total cholesterol, glucose, HbA1c, body mass index (BMI), body fat (%), systolic and diastolic blood pressure (BP) correlated with age, whereas sodium, creatinine, and resting heart rate did not. VO₂ max, eGFR, and serum albumin levels were lower in older individuals, whereas the remaining 13 biochemical and physiological variables increased with age (for all tests, p_adj < 0.05; Fig. 1c and Supplementary Table 5). Overall, age was associated with higher levels of most biochemical and physiological measures, with the exceptions of VO₂ max, eGFR, and albumin, and with increased plasma concentrations of 40% of the cytokines.

To evaluate the relative contributions of age and gut microbiota composition to plasma cytokines and physiological variables, we quantified the variance in each outcome explained by age alone, the microbiota alone, and both factors combined. Microbiota-associated variance was estimated using principal coordinates capturing ~80% of Bray-Curtis compositional variability, thereby limiting inflation caused by correlations among individual bacterial taxa.

For 29 out of the 30 plasma cytokines (97%), gut microbiota composition explained more variance than age (Fig. 1d). Only CXCL8 showed comparable contributions from both factors. Cytokines involved in antimicrobial and acute phase responses (IL-23, IL-6, TNF-α, CRP, and MIF), in addition to IL-13 and VEGF, were most strongly associated with microbiota composition (Fig. 1d).

Age accounted for most of the variance in 3 of 19 (16%) of the biochemical and physiological variables (VO₂ max, systolic BP and diastolic BP), whereas microbiota composition explained more variance in the remaining 16 variables (84%), including BMI, body fat %, LDL-cholesterol, HDL-cholesterol, TAG, total-cholesterol, glucose, and HbA1c (Fig. 1d). These patterns were replicated in the independent MetaCardis cohort²⁰ across all overlapping variables, including BMI, body fat %, systolic BP, diastolic BP, LDL-cholesterol, TAG and HbA1c (Supplementary Fig. 2a).

To assess predictive generalizability, we complemented the variance-partitioning analysis with internal validation using elastic-net regression and 10-fold cross-validation in both cohorts. Cross-validated R² values were comparable to or higher than those obtained from the ANOVA-based variance partitioning, demonstrating that microbiota composition provides robust and generalizable predictive power for most plasma cytokines and physiological variables (Supplementary Fig. 2b). This pattern was likewise reproduced in the MetaCardis cohort for all available variables (Supplementary Fig. 2c).

Elevated early-adulthood low-grade systemic inflammation in individuals with the Bacteroides 2 enterotype

To examine how distinct gut microbial community structures relate to low-grade systemic inflammation, individuals were clustered by fecal microbiota composition using a Dirichlet multinomial mixture model²¹ (Supplementary Fig. 3a). This reproduced previously described enterotypes^22,23: Bacteroides 1 (B1, 34%), Richness (R, 30%), Prevotella (P, 19%), and Bacteroides 2 (B2, 17%) (Fig. 2a; Supplementary Fig. 3b, c; Supplementary Table 6, Supplementary Data 1).

Fig. 2. Early‑adult life elevations in low‑grade systemic inflammation in individuals with the Bacteroides 2 enterotype.

a NMDS plot based on Bray-Curtis dissimilarity of gut microbiota composition. Points are colored by enterotype (dark green: Bacteroides 1, red: Bacteroides 2, yellow: Prevotella and light green: Richness. Colors are used throughout the figure), with confidence ellipses indicating within‑enterotype covariance. Point size reflects participant age. b Prevalence of each enterotype across age groups. c Alpha diversity (Inverse Simpson) across enterotypes and age groups. Boxplot centers represent the median, and whiskers extend to values within 1.58 × the interquartile range; points represent individual observations. d Heatmaps of z‑scored, age‑adjusted plasma cytokine levels (left) and biochemical/physiological variables (right) across enterotypes. Rows and columns are ordered using complete hierarchical clustering. Asterisks indicate statistically significant pairwise differences (two-sided Wilcoxon rank‑sum test), with colors corresponding to enterotype labels. e Infographic summary of age‑related trajectories for cytokines and biochemical/physiological markers across enterotypes. Top left: markers with parallel age‑related trajectories; middle left: markers with differential trajectories. Bottom left: example trajectory for TNF‑α against age, shown with a solid line representing the parametric linear fits and shaded areas representing 95% confidence intervals. The horizontal line marks the average TNF‑α level of a 20‑year‑old B2 individual, and vertical dashed lines denote the ages at which other enterotypes reach that level. Right: The dot on top of each bar represents differences in years relative to a 20‑year‑old B2 individual for markers with differential trajectories. N = 1199 paired samples. p values in bold indicate statistically significant associations adjusted for multiple testing (p_adj < 0.05). All statistical tests were two-sided. Source data are provided in the Source data file. ALT alanine aminotransferase; BDNF brain-derived neurotrophic factor; BP blood pressure; CCL C-C motif chemokine; CRP C-reactive protein; CXCL chemokine (C-X-C motif) ligand; eGFR estimated glomerular filtration rate; HbA1c glycated hemoglobin; HDL high-density lipoprotein; IFN interferon; IL interleukin; LDL low-density lipoprotein; MIF macrophage migration inhibitory factor; TAG triacylglycerol; TNF tumor necrosis factor; VEGF vascular endothelial growth factor.

Enterotype distributions varied with age (Fig. 2b), with males showing a more pronounced shift and a convergence between sexes after ~60 years (Supplementary Fig. 3c). Alpha diversity was stable across age but differed between enterotypes (for all tests, p_adj < 0.05; Fig. 2c and Supplementary Tables 7 and 8). Within-enterotype Bray-Curtis dissimilarity was highest in the B2 enterotype (Supplementary Fig. 3d), consistent with an earlier report²⁴, while alpha diversity (inverse Simpson) was highest in the R enterotype; both findings were replicated in MetaCardis (Supplementary Fig. 3e).

After adjusting for age, B2 individuals had significantly elevated levels of 8 of 30 cytokines (26.7%), including IL-6, TNF-α, CRP, IFN-γ, IL-18, MIF, VEGF, and CCL22 relative to the other enterotypes (for all tests, p_adj < 0.05; Fig. 2d and Supplementary Table 9). In contrast, R individuals exhibited the lowest CRP levels and higher levels of the antiviral defense cytokine IL-12p70 than B1 individuals. B2 individuals also had higher BMI and TAG, and lower HDL-cholesterol and VO₂ max than the other enterotypes (for all tests, p_adj < 0.05; Fig. 2d and Supplementary Table 10). R individuals showed lower ALT, BMI, and TAG than other enterotypes, and P individuals had higher diastolic blood pressure than B1 individuals. Although biochemical and physiological measures were within standard clinical reference ranges, B2 individuals consistently had average values closer to upper or lower clinical limits (Supplementary Table 11).

B2 enterotype individuals exhibited significantly different early-adult life offsets for five of the 16 age-associated cytokines (IFN-γ, IL-6, TNF-α, CRP, and VEGF) and four of the 17 age-associated biochemical/physiological variables (HDL- and LDL-cholesterol, TAG, and VO₂ max) (for all tests, p_adj < 0.05; Fig. 2e and Supplementary Fig. 4a). Individuals with the B1, P, and R enterotypes reached cytokine levels comparable to those of 20-year-old B2 individuals only after an additional ~37 years (Fig. 2e and Supplementary Table 12). Although B2 individuals showed earlier elevations in hallmark low-grade inflammatory cytokines (CRP, IL-6, TNF-α), cytokine levels converged across enterotypes by approximately 65 years of age (Supplementary Fig. 4a). TAG levels observed in 20-year-old B2 individuals were not reached by B1, P, and R individuals until 37–41 years later. B2 individuals also displayed consistently lower VO₂ max and HDL-cholesterol across the age range (Fig. 2e, Supplementary Fig. 4a and Supplementary Table 12). These patterns were reproduced for overlapping markers in the independent MetaCardis cohort (Supplementary Fig. 4b, c).

Age-related plasma variables showed strong interconnectedness across enterotypes (Supplementary Fig. 4d and Supplementary Table 13). IFN-γ, IL-6, TNF-α, CRP, VEGF, TAG, and LDL-cholesterol were positively associated, with inverse associations to HDL-cholesterol and VO₂ max (for all tests, p_adj < 0.05). Together, these linked markers displayed earlier adult-life offsets in B2 individuals, consistent with an early-onset inflammaging-like profile.

Pro-inflammatory cytokines account for part of the association between the Bacteroides 2 enterotype and future disease risk

To assess whether the B2 enterotype is linked to future incident disease development during an up to 9-year follow-up (average 8 years), we examined incident hospital-associated diagnoses in the DanFunD cohort using the Danish National Patient Register (DNPR). B2 enterotype individuals showed a significantly higher risk of future incident hospital diagnoses, with an overall hazard ratio (HR) of 1.21 (p_adj = 0.009) in Cox proportional hazards models using age at diagnosis as the time scale (Fig. 3a and Supplementary Table 14). This elevated risk extended across 7 of 13 disease chapters, including nervous, circulatory, digestive, respiratory, musculoskeletal, and ocular (eyes and adnexa) disorders, as well as infections (HR: 1.40, p_adj = 0.026; 1.39, p_adj = 0.009; 1.44, p_adj = 0.007; 1.46, p_adj = 0.012; 1.40, p_adj = 0.0008; 1.36, p_adj = 0.020; and 1.53, p_adj = 0.015, respectively). In contrast, individuals with the R enterotype exhibited a reduced overall hazard of future disease (HR: 0.79, p_adj = 0.0001), with particularly lower risks for digestive, musculoskeletal, genitourinary, and ocular diagnoses, as well as infections (HR: 0.72, p_adj = 0.008; 0.75, p_adj = 0.0008; 0.78, p_adj = 0.025; 0.75, p_adj = 0.025; and 0.68, p_adj = 0.025, respectively).

Fig. 3. Increased future incident disease risk in individuals with a Bacteroides 2 gut microbiota enterotype.

a Forest plot of adjusted hazard ratios with 95% confidence intervals (CI) represented as dots and horizontal lines for future incident disease across 13 ICD-10 chapters for each enterotype (n = 2259). Each enterotype is represented by the colors: dark green; Bacteroides 1, red: Bacteroides 2, yellow: Prevotella and light green: Richness. Colors are used throughout the figure. Cox proportional hazards models were adjusted for sex and smoking status, using age as the underlying time scale. p values in bold indicate statistically significant associations adjusted for multiple testing (p_adj < 0.05). All statistical tests were two-sided. b Schematic overview of the mediation analysis, with B2 enterotype status as the exposure, the cytokine profile as the mediator, and incident disease risk as the outcome. c Proportion of the elevated disease risk in B2 individuals (relative to the other three enterotypes) that is mediated through cytokine levels (n = 1199). Bars represent only statistically significant mediation effects. Source data are provided in the Source data file.

Focusing on current disease at the time of fecal sampling, we confirmed earlier reports^20,25 of a higher cross-sectional prevalence of metabolic diseases in B2 enterotype individuals (Odds ratio (OR): 2.39, p_adj = 1.81 × 10⁻⁸, Supplementary Fig. 5a, b and Supplementary Table 15). B2 enterotype individuals also showed higher prevalence across most other disease categories, with an overall risk OR of 2.16 (p_adj = 0.0002, Supplementary Fig. 5b and Supplementary Table 15).

We next evaluated whether the elevated future incident disease risk among B2 enterotype individuals was mediated by systemic inflammatory cytokines. Mediation analyses were performed using B2 enterotype status as the exposure, each disease chapter with a significantly increased HR as the outcome, and plasma cytokines as potential mediators (Fig. 3b). To account for correlations among cytokines, all 30 markers were combined into a single individual-level score using dbRDA (“Methods”; Supplementary Fig. 6). Cytokines mediated 4.1% of the overall disease risk, with contributions varying across disease chapters (4–25%). Mediation was strongest for respiratory (25%) and infectious diseases (24%), followed by musculoskeletal (16%), digestive (13%), and circulatory diseases (4%) (Fig. 3c and Supplementary Table 16). Together, these findings suggest that the early-adult pro-inflammatory phenotype observed in B2 individuals may partly explain their increased incident disease risk up to 9 years later.

Associations between Bacteroides 2-enriched taxa and early-onset inflammaging markers

We next examined associations between individual bacterial taxa and the early-onset inflammaging markers, elevated IFN-γ, IL-6, TNF-α, CRP, VEGF, TAG, and LDL-cholesterol, and reduced HDL-cholesterol and VO₂ max, while adjusting for age and relevant covariates. Among the 351 detected genera, 30 were associated with both age and the inflammaging-related variables. Genera characteristic of the B2 enterotype were more frequently positively associated with age and early-onset inflammaging markers than those characteristic of the B1 and P enterotypes, whereas taxa enriched in the R enterotype showed predominantly inverse associations (Fig. 4a and Supplementary Table 17). Eight B2-abundant taxa, including Escherichia/Shigella, Lactobacillus, and Ruminococcus gnavus, were positively associated with IFN-γ, IL-6, TNF-α, CRP, VEGF, and TAG, and inversely associated with HDL-cholesterol and VO₂ max. An exception was Lactococcus, which showed an inverse association with CRP.

Fig. 4. Gut bacteria elevated in the Bacteroides 2 enterotype strongly correlate with early-onset inflammaging markers.

a Network of associations between gut bacterial genera, plasma cytokines; vascular endothelial growth factor (VEGF), interleukins (IL), C-reactive protein (CRP), tumor necrosis factor-α (TNF-α) and interferon-ɣ (IFN-ɣ), and biochemical/physiological variables; estimated respiratory fitness (VO₂ max), low-density lipoprotein (LDL), high-density lipoprotein (HDL), and triacylglycerol (TAG). Bacterial node colors indicate enterotype‑associated enrichment based on relative abundance. Each enterotype is represented by the colors: dark green: Bacteroides 1 (B1), red: Bacteroides 2 (B2), yellow: Prevotella (P), and light green: Richness (R). Colors are used throughout the figure. Edges represent significant associations after adjustment for multiple testing (p_adj < 0.05), derived from adjusted z‑values in generalized linear models. All statistical tests were two-sided. Edge colors denote directionality, and node size reflects the number of significant connections. Exact p values are provided in Supplementary Table 17. Taxonomic levels above genus are indicated as follows: family (_f), order (_o), class (_c). b, c Combined relative abundance of B2‑ or R-associated taxa (identified in a) plotted against age within each enterotype. Enterotype‑specific trajectories are shown as a solid line representing parametric linear fits to the relative abundances with 95% confidence intervals indicated by the shaded area. All panels include data from 1,199 individuals with paired gut microbiota and cytokine measurements. Source data are provided in the Source data file.

In contrast, 12 R-enriched taxa, including the Eubacterium siraeum group, Lachnotalea, Paludicola, and members of the Ruminococcaceae family, were inversely associated with one or more of IFN-γ, IL-6, TNF-α, CRP, and TAG and positively associated with HDL-cholesterol and VO₂ max, consistent with a profile linked to healthier aging.

Within the P enterotype, Haemophilus correlated positively with CRP and inversely with HDL-cholesterol, whereas Oribacterium showed inverse associations with CRP and TAG. The B1-enriched Ruminococcus gauvreauii group exhibited positive associations with CRP, VEGF, and TAG and inverse associations with HDL-cholesterol (Supplementary Table 17).

Although taxa assigned to specific enterotypes were detectable across all enterotypes, their relative abundances differed markedly. B2-abundant bacteria were most prevalent in older B2 individuals (Fig. 4b), and while B1, P, and R individuals also harbored these taxa, the levels observed in 20-year-old B2 individuals were not reached until after ~50 years of age or later in the other enterotypes. Conversely, B1 and P individuals exhibited higher abundances of R-enriched taxa than B2 individuals (Fig. 4c).

Discussion

Age is widely considered the dominant driver of inflammaging, and indeed, we reconciled this here, observing strong age-dependencies for more than half of the plasma cytokines and most biochemical and physiological variables. However, gut microbiota composition showed even stronger associations with systemic inflammatory tone, accounting for more variance than chronological age for nearly all cytokines. This greater explained variance likely reflects the capacity of a multidimensional microbiota profile to capture inter-individual differences in inflammatory tone beyond chronological age. This pattern was most pronounced in individuals with the Bacteroides 2 (B2) enterotype, who displayed an inflammaging-like profile already in early adulthood and a higher incidence of subsequent disease. Mediation analyses further indicated that part of this increased disease risk was statistically attributable to elevated pro-inflammatory cytokine levels detectable up to 9 years before diagnosis. Together, these findings demonstrate that microbiota-associated inflammatory profiles can be present well in advance of clinical disease manifestation.

The inflammaging-like profile observed in B2 individuals was characterized by higher relative abundances of specific bacterial taxa, including Escherichia/Shigella, Lactobacillus, and Ruminococcus gnavus. These taxa were positively associated with elevated IFN-γ, IL-6, TNF-α, CRP, VEGF, and TAG, and inversely associated with HDL-cholesterol and VO₂ max, particularly in younger adults. Together, these associations identify specific microbiota configurations that co-occur with systemic inflammatory and metabolic features linked to increased disease risk, without implying directionality.

In contrast, taxa characteristic of the high-diversity R enterotype, including the Eubacterium siraeum group, Lachnotalea, Paludicola, and members of the Ruminococcaceae family, showed inverse associations with inflammatory cytokines and more favorable physiological profiles. Individuals with an R enterotype exhibited lower systemic inflammation across adulthood and a reduced overall incidence of disease. These findings are consistent with previous studies linking high gut microbial diversity or “uniqueness” to healthier aging²⁶. The observed associations for R-enriched taxa may relate to their metabolic capacity for short-chain fatty acid production, particularly butyrate^27,28, a metabolite known to reinforce intestinal barrier integrity and mitigate intestinal inflammation²⁹. These processes may help delay or even prevent inflammation-driven acceleration of local and systemic senescence³⁰, thereby reducing the risk of age-related lifestyle-related diseases like type-2 diabetes³¹. Such microbial functions provide a plausible biological context for the healthier inflammatory profiles associated with this enterotype, although causal relationships cannot be inferred from the present data.

The higher prevalence of the R enterotype among older participants in both the DanFunD and MetaCardis cohorts warrants consideration. One potential explanation is selective participation, whereby individuals reaching older age with lower systemic inflammation and reduced morbidity may be more likely to enroll in cohort studies, possibly leading to underrepresentation of less healthy individuals. This represents a general limitation of population-based research. As in any observational microbiome study, residual confounding cannot be excluded, particularly from dietary factors that may influence both gut microbiota composition and systemic inflammation. Broad dietary information was available in DanFunD, but a more detailed dietary assessment would be needed to fully address this potential source of confounding. Key strengths of the present study included the large number of participants with paired fecal and blood samples, enabling integrated microbiota-immune analyses, linkage to national health registers for prospective disease outcomes, and replication of major findings in an independent cohort.

We further observed that taxa enriched in the B2 enterotype increased in relative abundance with age across all enterotypes, indicating that enterotype classification does not fully capture inter-individual variation in inflammatory profiles. Individuals assigned to other enterotypes may still harbor B2-associated taxa at varying levels, potentially influencing their systemic inflammatory signatures. For example, Escherichia/Shigella was frequently detected outside the B2 enterotype, albeit at lower abundances. Thus, if a person accumulates higher abundances of the inflammaging-linked gut bacteria Escherichia/Shigella, Lactobacillus, or Ruminococcus gnavus, independent of their assigned enterotype, our findings suggest an increased likelihood of developing an inflammaging-like phenotype. On the other hand, individuals with higher abundances of R-associated taxa, including the Eubacterium siraeum group, Lachnotalea, Paludicola, and members of the Ruminococcaceae family, exhibit lower systemic inflammation, a healthier physiological profile, and a reduced overall disease risk. These findings underscore that enterotypes represent dominant community patterns rather than discrete biological states and that relative abundances of specific taxa may provide complementary information.

In summary, gut microbiota composition showed consistently stronger associations than chronological age with low-grade systemic inflammatory markers linked to future disease risk. Individuals with the B2 enterotype exhibited inflammatory and physiological profiles resembling those typically observed later in adulthood, with average offsets of up to 37 years compared with other enterotypes. Together, these results highlight robust associations between gut microbiota composition, systemic inflammation, host physiology, and subsequent disease risk across the adult lifespan, positioning the gut microbiota as an important correlate of inflammation-related health trajectories while underscoring the need for longitudinal and interventional studies to establish causality.

Methods

Cohort characteristics and physiological measures

This study involved participants enrolled in the population-based DanFunD cohort¹⁸. The study was approved by the Ethical Committee of Copenhagen County (Ethics Committee approvals H‑3‑2011‑081 and H‑3‑2012‑0015) and the Danish Data Protection Agency. All participants provided written informed consent. A total of 25,368 men and women aged 18–72 years living in the western part of the greater Copenhagen area were invited to participate. Of these, 7493 participants (29.5%) were enrolled in the period 2013–2015. Exclusion criteria were being born outside Denmark, not holding Danish citizenship, or pregnancy. At the study visit, participants underwent a general health examination and completed a questionnaire beforehand. All examinations were conducted between 8:00 and 15:00. Participants arrived after at least 6 h of fasting (including abstaining from coffee) and were instructed to refrain from smoking for at least 1 h prior to their appointment.

The health examination included assessments of cardiopulmonary fitness (blood pressure, resting heart rate, step test performance) and anthropometric characteristics (weight, height, and body fat percentage). Blood pressure and resting heart rate were measured after 5 min of rest. Cardiorespiratory fitness (VO₂max, mL/kg/min) was estimated using the Danish step test¹⁹. In brief, participants stepped at an initial frequency of 0.2 steps/s, gradually increasing to 0.8 steps/s at 6 min. Four bench heights (20, 25, 30, or 35 cm) were used depending on the participant’s age, general fitness, and baseline physical activity level¹⁹. Questionnaire data captured lifestyle factors such as smoking, alcohol consumption, and dietary habits. Smoking was categorized as a daily/occasional smoker versus an ex‑smoker/never smoker. Alcohol consumption was recorded in self‑reported weekly units, and dietary habits were assessed using food‑frequency questionnaires and categorized as poor, average, or healthy as previously described³².

Systemic physiological markers of metabolic health were measured through analysis of blood samples. Plasma was analyzed for glucose; glycated hemoglobin (HbA1c, IFCC method); total, high- and low-density lipoprotein (HDL‑ and LDL)‑cholesterol; triacylglycerol (TAG); sodium; calcium; alanine aminotransferase (ALT); albumin; carbamide; creatinine; and estimated glomerular filtration rate (eGFR). Total cholesterol, HDL‑cholesterol, and TAG were measured from freshly collected fasting plasma using colorimetric slide methods on the Vitros 4600/5600 Ortho Clinical Diagnostics platform. Non‑HDL cholesterol was calculated as total cholesterol minus HDL‑cholesterol. Glucose was measured using the hexokinase/glucose‑6‑phosphate dehydrogenase assay (Roche Diagnostics, formerly Boehringer Mannheim), and HbA1c was determined using high-performance liquid chromatography (TOSOH, Minato, Japan). Remaining biochemical markers were analyzed at the Clinical Biochemistry Department at Glostrup Hospital using standardized diagnostic methods.

Determination of systemic inflammatory cytokines

Plasma from 4101 individuals was collected during the health examination and stored at −80 °C until analysis. Samples were obtained as part of a nested case-control design and included individuals who reported having a functional disorder (n = 1190) along with randomly selected controls (n = 2775). Among the 1199 individuals included in the present study, 345 reported having a functional disorder.

Plasma was quantified for inflammatory cytokine levels using the following pre-coated single and multiplex immunoassay kits from MesoScale Discovery (Rockville, MD, USA): V-PLEX CRP, R-PLEX human brain-derived neurotrophic factor (BDNF), U-PLEX Biomarker Group1 (CCL24, interferon (IFN)-α2a, IFN-β, IL-17E/IL-25, IL-18, IL-22, IL-23, IL-31, IL-33, macrophage migration inhibitory factor (MIF)), V-PLEX Cytokine Panel 1 (IL-α, IL-5, IL-15, IL-17A, vascular endothelial growth factor (VEGF)), V-PLEX Chemokine Panel 1 (CCL11, CCL17, CXCL-10, CCL2, CCL22), V-PLEX Proinflammatory Panel 1 (IFN-γ, IL-1β, IL-2, IL-4, IL-6, IL-8, IL-10, IL-12p70, IL-13, TNF-α). In each plate, an internal standard was included in duplicate, and wells with buffer for calculation of lower limit of detection (LLOD) were included. Samples were analyzed according to the manufacturer’s instructions, except for overnight incubation of samples/standards on a shaker at 4 °C to increase sensitivity. Plates were read using a Sector Imager 6000 (MesoScale Discovery®), and concentrations were estimated using an eight-point standard curve. For IL-1α, 78% of the samples were below the detection level, and for IL-1β, it was 69% of the samples. Levels for IL-1α and IL-1β were therefore removed before the subsequent data analysis.

Collection of fecal samples and 16S rRNA gene library preparation

As part of the initial visit, all participating individuals were invited to take part in the microbiota analysis, and 2460 individuals agreed to donate a fecal sample. Samples were collected by participants at home, stored at −20 °C for a minimum of 6 h, and subsequently transported to the laboratory on dry ice. Microbial DNA was extracted using the NucleoSpin Soil kit (Macherey‑Nagel, Düren, Germany). Libraries were sequenced on the Illumina HiSeq 2500 platform using paired‑end reads (2 × 250 bp).

16S rRNA gene data processing

Sequencing reads were processed and denoised using DADA2 v1.20.0³³. Resulting amplicon sequence variants (ASVs) were compared to the 99% identity clustered SILVA database v138.1³⁴ using a naïve Bayesian classifier³⁵ implemented in DADA2, which was trained on the amplified variable region. Across all samples, we identified 13,078 unique bacterial ASVs belonging to 595 genera or above if not defined at the genus level. Genera detected in fewer than 30 individuals (2.5% of the cohort) or represented by fewer than 100 total reads were excluded, yielding 351 genera for downstream analyses.

Statistics and reproducibility

All analyses were performed in R Statistical Software v4.2.0³⁶. Generalized linear modeling (GLM) of age, microbiota features, cytokines, and physiological parameters was carried out using the glmmTMB package v1.1.8³⁷. Prior to correlation analyses, cytokines and physiological variables were log10‑transformed, and a small number of missing physiological values (n = 1–15) were imputed using the mean of individuals matched on sex and age group. Models with age, cytokines, or physiological parameters as outcomes were fitted assuming a Gaussian distribution, whereas microbiota features were modeled as raw read counts using a negative binomial distribution (nbinom2 in glmmTMB). All models were adjusted for age, sex, antibiotic use, and smoking status. An offset term of logged sequencing depth was included in models with microbiota features as outcomes to account for variation in sequencing coverage across samples. Group differences were assessed using the Wilcoxon rank‑sum test. All statistical tests were two‑sided unless specified otherwise. p values were adjusted for multiple testing using the Benjamini–Hochberg procedure and are reported as false‑discovery‑rate-adjusted values (p_adj). An adjusted significance threshold of 0.05 was applied whenever multiple variables were tested (e.g., cytokine-taxa, physiology-taxa, or age-cytokine/physiology associations). No statistical method was used to predetermine sample size.

Identification of gut microbiota enterotypes

Gut microbiota enterotypes were identified through unsupervised clustering of individuals based on raw genus‑level counts using a Dirichlet multinomial mixture model, as implemented by Holmes et al. in the DirichletMultinomial package v1.38.0²¹. The optimal number of mixture components (enterotypes, n = 4) was determined by evaluating model fit using the Bayesian Information Criterion (BIC), as shown in Supplementary Fig. 3a.

Gut microbiota composition, diversity, and within-enterotype variation

Unconstrained ordination was applied to visualize gut microbiota enterotypes using non-metric multidimensional scaling (NMDS) of a Bray-Curtis dissimilarity matrix, computed with the vegdist function from the vegan package v2.6-4³⁸. Ordination was performed in four dimensions using the metaMDS function from the same package. The resulting Bray-Curtis matrix was also used to calculate average within-enterotype dissimilarity for each individual across age groups. Microbial diversity was quantified using the Inverse Simpson and Shannon indices, computed with the diversity function in the vegan package.

Variance in plasma cytokines, biochemistry, and physiology is explained by age and the gut microbiota

For each cytokine and each biochemical or physiological variable, we quantified the proportion of variance explained by age and by the genus‑level composition of the gut microbiota. To avoid inflating the contribution of the microbiota due to co‑correlations among genera, microbiota composition was represented using principal coordinates that together explained 80% of the total variance. These coordinates were computed on square‑root-transformed Bray–Curtis dissimilarity distances using the ape package v5.6‑2³⁹. Permutational analysis of variance (ANOVA), implemented in the lmperm package v2.1.0, was used to assess the significance of age and/or the principal coordinates for each log‑transformed cytokine or physiological measure. Only principal coordinates with p < 0.05 were included in the calculation of the total variance explained.

Modeling of disease risk using the Danish National Patient Registers

Using the Danish National Patient Registers, we identified all hospital‑associated contacts for each participant within the Danish healthcare system. Individuals were followed from the time of sampling (2013–2015) until emigration, death, or 31 August 2022, whichever came first, yielding a maximum follow‑up period of 9 years and an average of 8 years. Diagnosis outcomes were defined as the first occurrence of any three‑letter ICD‑10 code within a given chapter that had not previously appeared in an individual’s medical record. We investigated ICD‑10 chapters I, II, III, IV, VI, VII, VIII, IX, X, XI, XII, XIII, and XIV. Analyses were conducted separately for each chapter, with participants grouped according to their enterotype.

Among individuals who provided a fecal sample, only those with available smoking information were included (n = 2259) to enable adjustment for smoking status. Each enterotype was iteratively compared using the remaining three enterotypes as the reference group. Adjusted hazard ratios were estimated using Cox proportional hazards regression models with age as the underlying time scale. Analyses were performed using the survival package v.3.5‑7 and adjusted for smoking status as described above, with sex included as a stratum to account for diagnoses with strong sex‑specific prevalence.

To assess differential diagnosis prevalence at the time of microbiota sampling across enterotypes, we applied a similar framework using logistic regression. Here, the outcome was defined as the first ICD‑10 code for each chapter recorded prior to sampling. These models were implemented using the glm function and adjusted for age and smoking, with sex again included as a stratum.

Mediation analyses

To perform a causal mediation analysis with B2 microbiota enterotype status as the exposure, disease risk as the outcome, and the cytokine landscape as the mediator, we first reduced the 30 cytokines into a single multivariate cytokine score per individual to account for inter‑cytokine correlations. This was achieved using distance‑based redundancy analysis (dbRDA) on a dissimilarity matrix (Manhattan distance) computed from log10‑transformed and scaled cytokine values. The dbRDA was conditioned on age, sex, and smoking status to adjust for confounding and constrained on enterotype. The first dbRDA axis, representing the major gradient in cytokine variation, was used as the cytokine score in the mediation framework. Mediation analysis was conducted under the counterfactual framework⁴⁰ using the Mediator package v.0.0.9 (https://github.com/GerkeLab/mediator). The effect of exposure on the outcome was modeled using a Cox proportional hazards model as described above, including the cytokine score as a covariate, while the mediator model was specified as a linear regression of the form using the lm function, modeled as: cytokine score ~ microbiota enterotype status + sex + smoking status.

Estimation of enterotype-specific trajectories for systemic cytokine and physiological variables

The relationship between microbiota enterotypes and age for each cytokine and physiological variable was modeled using the formula variable ~ age × enterotype + covariates, with the B2 enterotype as the reference group and sex, antibiotic use, and smoking included as covariates. Features showing a significant interaction between the variable and microbiota enterotype were used to estimate differential age‑trajectories. For B2 individuals, covariate‑adjusted estimates and intercepts were extracted and used to calculate the predicted value for a 20‑year‑old B2 subject using linear modeling (feature level at age 20 = intercept + estimate × 20 y). To determine the corresponding age for the other enterotypes, estimates and intercepts were taken from linear models fitted within each respective enterotype: estimated age for a non‑B2 enterotype = (feature level of a 20‑year‑old B2 − intercept) / estimate. The difference in years relative to a 20‑year‑old B2 individual was then obtained by subtracting 20 years.

Network analysis

Networks were constructed based on correlations between age, microbial taxa, and cytokine or physiological variables, following the procedures detailed under Statistics and reproducibility. Network structures were generated using the igraph package v.1.4.2⁴¹ and visualized in Cytoscape v.3.10.0⁴². The interactions underlying each network, such as cytokine-to-taxa or physiology-to-taxa links, were represented by z-values (estimate/standard error) derived from GLMs performed as described above. Only statistically significant associations (p_adj < 0.05) were retained to define interactions.

Elastic net model

Elastic net regression was applied using the glmnet package v.4.1-10 package⁴³ with 10-fold cross-validation to assess predictive generalization of the analyses shown in Fig. 1d. We quantified how much of the variance in each cytokine and physiological parameter could be explained by (1) chronological age, (2) gut microbiota composition, and (3) the combined contribution, following the structure of the original analysis. Microbiota composition was represented by Principal coordinate analysis (PCoA) components as previously described. For each outcome, elastic net models were fitted using 10‑fold cross‑validation. Age‑only models were constrained to a ridge penalty (α = 0), whereas microbiota‑based models were tuned across α ∈ {0, 0.1, …, 1.0} to identify the optimal ridge-lasso balance. In all models, the regularization parameter λ was selected by minimizing the cross‑validated mean absolute error. Predictors and outcomes were z‑scored within each training fold to avoid data leakage. Model performance was evaluated using the cross‑validated coefficient of determination (R²), expressed as the percentage of variance explained, calculated as the squared correlation between observed and predicted values for each of the three predictor sets.

Replication using the independent MetaCardis cohort

Technical procedures for sample collection, DNA extraction, metagenomic sequencing, and initial quality control are described in detail in Fromentin et al.²⁰. Processed metagenomic data from the MetaCardis study were obtained through the curatedMetagenomicData package⁴⁴. Metadata on age, sex, BMI, TAG, HbA1c, diastolic and systolic blood pressure, body fat percentage, and LDL cholesterol were retrieved from publicly available data repositories^20,45,46. The original MetaCardis cohort (n = 1831) included participants from Denmark, France, and Germany; Danish participants were excluded to avoid a within‑country replication, resulting in a final sample size of n = 1317. All analyses were replicated using the same methodological framework and analytical procedures applied to the original cohort.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Conceptualization: C.E., K.K., S.B.; Data curation: T.M.D., A.L., T.Jø., C.E.; Formal analysis: C.E.; Funding acquisition: T.Je., T.Jø., T.M.D., S.B.; Investigation: C.E., K.H.A., T.Je., O.P., T.H., T.M.D., T.Jø., S.B.; Project administration: T.Je., K.H.A., K.K., S.B.; Resources: T.M.D., O.P., T.H., A.L., T.Jø., S.B.; Software: C.E., I.C., R.K.D., P.L.J.; Supervision: T.Je., K.H.A., K.K., S.B.; Validation: C.E.; Visualization: C.E., S.B.; Writing—original draft: C.E., S.B.; Writing—review & editing: C.E., I.C., R.K.D., P.L.J., T.M.D., O.P., T.H., A.L., T.Je., K.H.A., T.Jø., K.K., S.B.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Funding

The DanFunD cohort was supported by TrygFonden (7-11-0213), the Lundbeck Foundation (R155-2013-14070), Novo Nordisk Foundation (NNF15OC0015896), and RCPH. The current study was supported by Læge Sofus Carl Emil Friis og Hustru Olga Doris Friis’ legat to T.Jø., and the Danish National Research Foundation (grant number DNRF148) to T.Je. T.H. and O.P. are employed at The Novo Nordisk Foundation Center for Basic Metabolic Research, which is an independent research institution at the University of Copenhagen, partially funded by an unrestricted donation from the Novo Nordisk Foundation. S.B. was the incumbent of the FII institute Research Chair at DTU in Immune-based Prediction of Disease (2021-2023).

Data availability

The sequencing data generated in this study have been deposited in the ENA database under accession code PRJEB123666. The metadata from the DanFunD cohort is available under restricted access due to national data privacy laws. Access can be obtained by following the procedure described at: https://www.frederiksberghospital.dk/ckff/omCKFF/sektioner/sfe/tidligere-undersoegelser/danfund/Sider/default.aspx. In short, access can be obtained upon reasonable request to line.lund.kaarhus@regionh.dk and upon filing of a written agreement that ensures compliance with local legal and ethical regulations. Disease history data were obtained from the Danish nationwide health registers (https://sundhedsdatastyrelsen.dk). Registered data are protected under the Danish Act on Processing of Personal Data. Access can be obtained by applying to both the Danish Data Protection Agency and the Danish Health Data Authority. Source data are provided with this paper.

Code availability

No custom algorithms were developed for this study. Details on all software and analytical tools used are provided in the “Methods” section.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77499-9.