Recent studies have revealed that the intestinal microbiota plays essential roles in host energy homeostasis, body adiposity, blood sugar control, insulin sensitivity, hormone secretion, and the pathogenesis of metabolic diseases, such as type 2 diabetes (T2D) and obesity1,2,3. However, most of these studies analyzed stool samples and therefore obtained limited information relative to insights from direct sampling of the intestinal mucosa, which is not possible in most cases. In addition, the composition of microbial communities in stool samples is greatly affected by the specific compartment in which they reside, such as the mucous membrane4. Microbial communities also differ based on their source, ranging from the intestines, skin, and airways, which are frequently studied, to urine and blood, which are generally sterile environments5. Therefore, it is important not only to understand the role of the intestinal microbiota but also to consider the function and combined contribution of the all microbiota throughout the body.

Extracellular vesicles (EVs) have been recently suggested to function as the main messengers between intestinal microorganisms and the host. EVs travel long distances within and between body tissues6 and have been used as biomarkers of atopic dermatitis, alcoholic hepatitis, and asthma7,8,9,10. Microbiota-derived EVs can enter the circulatory system through the intestinal barrier. They are suspected to play a key role in the development of insulin resistance, potentially providing important clues into the pathogenesis of T2D. For example, EVs derived from Pseudomonas panacis are present in the stool samples of high-fat diet-fed mice. They can infiltrated the gut barrier and block the insulin pathway in skeletal muscle and adipose tissue, inducing the development of insulin resistance and glucose intolerance11. However, microbiota-derived EVs are highly variable, as they are modulated by different factors, such as age and sex. Therefore, caution should be exercised when inferring causal relationships based on the statistical analysis of microbiota-derived data. Furthermore, longitudinal microbiota studies may allow for stronger inferences than cross-sectional studies12 and may allow for the detection of microorganisms related to the progression of T2D in healthy subjects. However, existing studies have been predominantly cross-sectional in nature and are based on correlation analyses. As a result, these studies are unable to comprehensively provide an understanding of the exact roles of the intestinal microbiota and EVs in metabolic disease development.

Therefore, in the present study, we investigated the prospective Korean Association REsource project (KARE) cohort13. By tracking changes in microbiota-derived EVs in urine samples from Korean adults collected three times over four years, we explored the potential associations between microorganisms and T2D progression. Furthermore, using genomic and metabolite data from the KARE cohort, we conducted a multiomics analysis to investigate the specific role of microorganisms potentially involved in T2D pathogenesis. We expect our findings to provide information regarding how microbes, the substances they produce, and their byproducts interact with the human body and affect metabolic disease development. In addition, we evaluated causal relationships among microbial organisms, ketone bodies, and clinical measurements, with the aim of further elucidating the relationship between T2D and the microbiota.

Materials and methods

Cohort and study design

The KARE cohort is a prospective study cohort involving subjects from the rural community of Ansung and the urban community of Ansan in South Korea. The KARE project began in 2001 as part of the Korean Genome Epidemiology Study14. We used data from urine samples taken from subjects in 2013, 2015, and 2017, which we refer to as phases 1, 2, and 3 in this study. After collection, the urine was stored at –80 °C. For the 1,891 subjects whose urine samples were available, age, sex, and body mass index (BMI) were matched via 2:1:1 propensity score matching. As a result, a healthy group (healthy in all phases, N = 328), a T2D-at-risk group (T2D-at-risk in all phases, N = 164), and a T2D group (T2D in any of the three phases, N = 164) were selected. From the remaining unmatched subjects, 35 T2D subjects were also included. Consequently, 691 subjects were finally included, and their 2,072 urine samples were subjected to microbiota analysis. Metagenomic, metabolite, clinical, and genomic data were subjected to comprehensive analyses (Fig. 1).

Fig. 1: Summary chart of data analysis.
figure 1

Datasets and analyses employed in the current study are described. Sample sizes for each analysis are presented.

Operational definition of T2D and related phenotypes

Study participants were categorized into control individuals, T2D-at-risk patients, and T2D patients. T2D and T2D-at-risk patients were diagnosed on the basis of the American Diabetes Association criteria, which are provided in Supplementary Table 1. T2D status was then stratified into T2D-at-risk/T2D (0 for healthy; 1 for T2D-at-risk and T2D) and binary_T2D (0 for healthy and T2D-at-risk; 1 for T2D). In addition, we considered other T2D-related indicators, such as BMI, HbA1c levels, fasting glucose and insulin levels, 60- and 120-min plasma glucose levels, and insulin levels after a 75 g oral glucose tolerance test in our analysis. Age, the levels of total cholesterol, high-density lipoprotein (HDL) cholesterol, triglycerides, kidney- and liver-related disease indicators (blood urea nitrogen (BUN), creatinine, aspartate aminotransferase (AST), and alanine aminotransferase (ALT) C-reactive protein (CRP), white blood cell (WBC) count, red blood cell (RBC) count, hemoglobin, hematocrit, and platelet count) were also collected. The homeostatic model assessment for insulin resistance (HOMA-IR) was calculated using fasting glucose and fasting insulin levels15. Descriptive statistics for all variables were generated using Rex software (RexSoft Inc., Seoul, Korea) (Supplementary Table 2)16.

EV isolation and DNA extraction

For EV isolation, urine samples were subjected to differential centrifugation at 10,000 × g and 4 °C for 10 min using a microcentrifuge (Labogene 1730R; Bio-Medical Science, Seoul, Korea)17. To remove bacteria, foreign particles, and waste, the supernatant was filtered through a 0.22-µm filter (Inchpor2 Syringe Filter; Inchemtec, Seoul, Korea). The isolated EVs were boiled at 100 °C for 40 min and centrifuged at 18,214 × g and 4 °C for 30 min to eliminate floating particles and impurities. The supernatant was collected and subjected to DNA extraction using a PowerSoil® DNA Isolation Kit (MO BIO Laboratories, Carlsbad, CA, USA) according to the manufacturer’s protocol. DNA was quantified using the QIAxpert system (Qiagen, Hilden, Germany).

16 S rRNA sequence data processing

Paired-end sequencing of the V3-V4 region of the bacterial 16 S rRNA gene was conducted at MD Health care (Seoul, Korea) with the MiSeq Reagent Kit v3 (600 cycles, Illumina, San Diego, CA, USA) using the widely used primers 16S_V3_F (5′-TCGTCGGCAGCGTCAGATGTGTATAAGAGACA-GCCTACGGGNGGCWGCAG-3′) and 16S_V4_R (5′-GTCTCGTGGGCTCGGAGATGTGTATA-AGAGACAGGACTACHVGGGTATCTAATCC-3′). Adaptor sequences were detected and removed using CUTADAPT software ( with a minimum overlap of 11, a maximum error rate of 10%, and a minimum length of 1018. Sequences were merged using CASPER ( with a mismatch ratio of 0.27 and filtered based on the Phred (Q) score, resulting in sequences of 350–550 bp in length19,20. After the merged sequences were dereplicated, chimeric sequences were detected and removed using VSEARCH ( and the Silva Gold reference database for chimeras21. Open-reference operational taxonomic unit (OTU) picking was conducted based on the EzTaxon database using UCLUST (,23. For each OTU, we calculated its proportion among all OTUs and determined the mean value across all subjects. If the resulting value was <0.001, the OTU was excluded24. Among the 691 subjects, those with a read count <3000 or whose genomic data were not available in any phase were excluded. As a result, 1179 samples from 393 subjects, including 70 genera, were used for subsequent analyses.

Prediction of functional profiles from 16 S rRNA metagenomic data

The functional potential of microbial communities can be predicted from their phylogeny. Tax4fun uses evolutionary modeling to predict metagenomes based on 16 S data from the SILVA reference genome database. The SILVA-based 16 S rRNA profile was used to estimate a taxonomic profile of prokaryotic Kyoto Encyclopedia of Genes and Genomes (KEGG) organisms. The estimated abundances of KEGG organisms were normalized using the 16 S rRNA copy number obtained from the National Center for Biotechnology Information (NCBI) genome annotations. Finally, the normalized taxonomic abundances were used to linearly combine the precomputed functional profiles of KEGG organisms to predict the functional profile of the microbial community25. Similar to the analysis of OTUs, we calculated the mean of the relative proportions across all subjects for each functional profile. If the resulting value was <0.001, the functional profile was excluded from the analysis. As a result, 238 functional profiles were retained for analysis.

Metabolite analysis of ketone bodies

Serum metabolites were analyzed using the Agilent 1290 Infinity LC and Agilent 6490 Triple Quadrupole MS systems (Agilent Technologies, Palo Alto, CA, USA). The levels of acetoacetate and 3-hydroxybutyrate from subjects included in the metagenomics dataset were determined in the multiple reaction monitoring mode. A batch normalizer was used to correct for possible batch effects26.

Analysis of bacterial composition and microbial variance

We calculated alpha- and beta-diversity indices using R (v3.6.2) after read number normalization with the Rarefy function in the R package GUniFrac (v1.1). The R package Fossil (v0.4.0) was used to obtain Chao1 and ACE diversity indices. The Shannon index and Simpson’s diversity index were calculated using the Vegan package in R (v2.5.6). Taxonomy-based ring charts were created using the Krona tool27. PERMANOVA is a nonparametric multivariate analysis of variance test based on pairwise distances28. The R package pldist was used to obtain the microbial variance for individuals in repeated measurements of microbial profiles. pldist summarizes within-individual shifts in the microbiome composition and compares these across individuals. pldist also calculates dissimilarities based on a novel transformation of relative abundances, which are then extended to more than two time points. They are then incorporated into a chosen beta-diversity, which, in our case, was Bray–Curtis dissimilarity. PERMANOVA was performed for biochemistry-related KARE phenotypes using the adonis function in R. PERMANOVA can be applied to the cross-sectional data, and thus, the phenotypes were averaged for phases 1, 2, and 3.

Statistical analysis of the effect of the microbiome on T2D and diabetes risk indicators

For each taxon and functional profile, a generalized linear mixed model (LMM) with the logit link function was used to find associations with binary_T2D and T2D-at-risk/T2D, whereas an LMM was used for log-transformed diabetes risk indicators. A random effect with a compound symmetry structure for each time point was incorporated to adjust the similarity of T2D status for the same subject at different time points, and the sandwich estimator was used to find a robust estimate against the misspecified covariance matrix. To accommodate the multiple testing problem, p values were adjusted for the false discovery rate (FDR) using the Benjamini–Hochberg method29.

Network analysis of a T2D-related taxon based on multiomics data

To assess overall associations using repeatedly measured multiomics data, we first modeled an LMM using log-transformed diabetes risk indicators as response variables and age in phase 1 as well as sex as explanatory variables with a compound symmetry structure for its covariance. We modeled an LMM with a T2D-related taxon as the response variable with the same covariates and covariance structure. For each combination of diabetes risk indicators and a T2D-related taxon, two different sets of residuals were obtained, and Spearman correlations between the residuals were calculated. Similarly, the association between a chosen microbial marker and the levels of ketone bodies was analyzed.

Network analysis was conducted to calculate simple correlations among diabetes risk indicators, a chosen taxon, and ketone bodies. Edge width was calculated as –log10 of the p value. The network was visualized using the R package visNetwork (v2.0.8).

Genotyping, imputation, and quality control

Quality control and genotype imputation were performed according to the standard quality control and imputation protocols for the genotypes of 8842 KARE cohort participants30. After quality control, 8216 subjects with 17,716,215 single-nucleotide polymorphisms (SNPs) were included in the analysis. In total, the data of 351 subjects with a read count <3,000 and nonmissing T2D status for all phases were used for a genome-wide association study (GWAS) of metagenomic data. A total of 574 subjects who had no missing metabolite levels and T2D status for all three phases were selected for a GWAS of metabolite levels. Among the subjects not included in the metabolite or metagenome GWAS, 3542 subjects had KARE phenotypes for the three phases and were thus included in a GWAS of KARE phenotypes. We excluded subjects in the metabolite or metagenome GWAS for the purposes of a two-sample Mendelian randomization (MR) study. Details are provided in Supplementary Fig. 1, and all the associated SNPs from each GWAS are listed in Supplementary Table 3.

MR analysis

MR uses genetic variants that are not associated with conventional confounders of observational studies and is therefore considered analogous to randomized controlled trials31. Randomly selected alleles are transmitted from parents, and genotypes can be assumed to be independent, with many potential confounders. This randomization produces unbiased estimates for the associations between the main exposures and outcomes. Thus, genetic variants associated with the main exposure were used as instrumental variables. There are two types of MR, namely, two-sample MR and one-sample MR. The former uses two independent datasets with nonoverlapping samples for the association of SNP exposure and SNP outcome (as opposed to one-sample MR) It is less likely to lead to inflated type 1 error rates and false-positive findings when compared to one-sample MR. Two-sample MR was conducted to identify the effect of a microbial taxon or each ketone body on KARE phenotypes by using no overlapping samples. One-sample MR was conducted to estimate the effect of a chosen taxon on each ketone body.

For one-sample MR, we conducted two-stage least-squares regression. The first stage consisted of a regression for SNP exposure, and the second stage consisted of a regression for the outcome of interest on the fitted values from the first-stage regression. The estimator of the coefficient for first-stage fitted values in the second-stage model is the causal estimate32,33. F-statistics from the first-stage regression were examined to avoid weak instrument bias34. The Durbin-Wu-Hausman (DWH) test for endogeneity35 was used to evaluate whether there is any evidence that the causal estimate differs from the ordinary least square estimate of exposure and outcome. For two-sample MR, the average F-statistic was used to avoid weak instrument bias. The inverse-variance-weighted (IVW) method, Cochran’s Q test, and MR-PRESSO global test were used to confirm the heterogeneity assumption, and I2 was used for the no measurement error (NOME) assumption. To enhance the validity of MR analysis, we considered the extensive range of existing MR methods, including IVW, MR-egger, MR-egger with SIMEX correction, median-weighted method, and MR-PRESSO, and selected the recommended MR method based on the violations of MR assumptions36.


Longitudinal changes in the urine microbial composition over four years

The alpha-diversity of the urine microbiome decreased during the follow-up period, which may have been an effect of aging (Supplementary Fig. 2). A nonmetric multidimensional scaling plot based on beta diversity also revealed a gradual change in microbiota composition with age (Supplementary Fig. 3). The overall microbiome composition at the phylum and genus levels is presented in Fig. 2 and Supplementary Fig. 4, respectively. Verrucomicrobia, Bacteroidetes, and Firmicutes were the predominant phyla, whereas Akkermansia and Bacteroides were the predominant genera.

Fig. 2: Taxonomic composition at the phylum level for phases 1, 2, and 3.
figure 2

Profile plots for phases 1 (a), 2 (b), and 3 c showing the relative abundances of bacterial taxa at the phylum level for each individual are presented.

T2D and other clinical traits explained by microbial variance

We investigated the associations between various clinical phenotypes and microbial compositions using PERMANOVA (Supplementary Fig. 5). HbA1c, WBC, hematocrit, binary_T2D, and age in phase 1 significantly explained changes in microbial composition during the follow-up period (p = 0.0061, 0.0107, 0.0110, 0.0409, and 0.0290, respectively; FDR-adjusted p = 0.1027, 0.1027, 0.1027, 0.2290, and 0.2030, respectively). HbA1c and binary_T2D partially explained the variance in microbial changes over the 4 years, indicating that the longitudinal change in microbiome composition may be more closely associated with T2D-related phenotypes than with other clinical traits.

Taxa and functional profiles associated with T2D and diabetes risk indicators

In an association analysis of 70 genera with binary_T2D and T2D-at-risk/T2D phenotypes, GU174097_g, an unclassified Lachnospiraceae, was found to exhibit a significant association with these phenotypes and was more abundant in healthy subjects than in diabetic or prediabetic patients (Table 1). We divided the samples into four groups. The Healthy in Phases 1-3 group included subjects who were healthy in phases 1, 2, and 3. The T2D in Phases 1-3 group consisted of subjects who had T2D in phases 1, 2, and 3. The Healthy to T2D-at-risk/T2D group included subjects who were healthy in phase 1 and became T2D patients or T2D-at-risk in phase 3. The T2D-at-risk/T2D to Healthy group included subjects who were T2D-at-risk/T2D in phase 1 and healthy in phase 3. The relative abundance of GU174097_g in subjects who were healthy at baseline but changed to the T2D-at-risk/T2D group at phase 2 or 3 decreased with the development of T2D (p = 0.0001). Conversely, its relative abundance in the T2D-at-risk/T2D to Healthy group exhibited no tendency to decrease (p = 0.19) (Fig. 3). Supplementary Fig. 6 shows the profiles of GU174097_g for randomly selected subjects. The relative abundance of GU174097_g in subjects who were healthy at baseline but changed to T2D-at-risk/T2D at phase 2 or 3 tended to decrease. Most T2D patients had small relative abundances of GU174097_g at baseline. In summary, GU174097_g was clearly associated with the progression of diabetes over time, and this association was not simply based on diabetic or nondiabetic status.

Table 1 Analysis of the associations between type 2 diabetes (T2D) and bacterial genera.
Fig. 3: Relative proportions of GU174097_g in different type 2 diabetes (T2D) groups.
figure 3

The mean relative proportions of GU174097_g are provided for the Healthy in Phases 1-3, T2D in Phases 1-3, Healthy to T2D-at-risk/T2D, and T2D-at-risk/T2D to Healthy groups and are compared according to the T2D status. The Healthy in Phases 1-3 group included subjects who were healthy in phases 1, 2, and 3. The T2D in Phases 1-3 group included subjects who had T2D in phases 1, 2, and 3. The Healthy to T2D-at-risk/T2D group included subjects who were healthy in phase 1 and became T2D patients or T2D-at-risk in phase 3. The T2D-at-risk/T2D to Healthy group included subjects who were T2D-at-risk/T2D in phase 1 and healthy in phase 3. The p values were calculated based on simple linear regression using the log relative proportion of GU174097_g as a response variable and phase variable. The relative proportion of GU174097_g was log-transformed after adding one to avoid zero values. The phase variable is coded by 1, 2, and 3 for phases 1, 2, and 3, respectively.

To investigate the T2D-associated microbial functional profiles, 238 functional profiles were evaluated. The significant associations at an FDR-adjusted significance of 0.1 are presented in Supplementary Table 4. The T2D-at-risk/T2D phenotype was related to the cationic antimicrobial peptide. Furthermore, the biosynthesis of fatty acids, coenzyme A (CoA), and secondary metabolites as well as oxidative phosphorylation were significantly associated with the Binary_T2D phenotype at an FDR-adjusted significance of 0.1.

Next, we investigated the associations between the log-transformed diabetes risk indicators and genera, and significant associations at an FDR-adjusted significance of 0.1 were identified. Twelve, four, and 20 genera were significantly associated with HbA1c, glucose, and insulin levels, respectively. In particular, Hafnia was associated with HbA1c and 60- and 120-min insulin levels, whereas AB185816_g and Akkermansia were associated with HbA1c, fasting glucose, and 60-min insulin levels (Supplementary Table 5).

Associations between T2D-related unclassified Lachnospiraceae and diabetes risk indicators and ketone bodies

To confirm the association between GU174097_g and T2D, we performed extensive validation analysis using clinical and metabolite data. We analyzed the association between GU174097_g and diabetes risk indicators (Table 2). Among all glucose- and insulin-related variables, GU174097_g was significantly and positively associated with the 60-min insulin level.

Table 2 Analysis of the associations between diabetes risk indicators and GU174097_g.

Thereafter, we analyzed the potential associations between ketone bodies and the T2D-related taxon, since ketone bodies have been suggested as markers of disrupted glucose metabolism in prediabetic patients37. The ketone bodies 3-hydroxybutyrate and acetoacetate exhibited significant negative correlations with GU174097_g (r = –0.0829 and –0.0938, respectively; p = 0.0069 and 0.0022, respectively) (Table 3). Supplementary Fig. 7 shows the tendency of high acetoacetate and 3-hydroxybutyrate concentrations coinciding with the low abundance of GU174097_g. The ion abundances of acetoacetate and 3-hydroxybutyrate did not rise beyond 2000 and 10000, respectively, when the relative abundance of GU174097_g was high.

Table 3 Analysis of the associations between GU174097_g and ketone bodies.

Finally, we established an association network for diabetes risk indicators and ketone bodies, as the same observed correlations can imply completely different biological processes. For example, if high levels of glucose or HbA1c tend to appear in parallel to high levels of insulin, insulin resistance may be present. However, if high levels of glucose or HbA1c are observed in parallel to low levels of insulin, insulin secretion may have suppressed glucose or HbA1c levels. Network analysis indicated strong associations among the diabetes risk indicators (Supplementary Fig. 8). In particular, the 60-min insulin level exhibited a strong negative correlation with HbA1c levels, suggesting that the former can decrease the latter. Ketone bodies exhibited negative correlations with fasting insulin and 60-min insulin levels and positive correlations with 60- and 120-min glucose levels.

Causal relationship between the T2D-related taxon and ketone bodies and the diabetes risk indicators

One-sample MR did not reveal any significant causal relationship between GU174097_g and ketone bodies and vice versa (Supplementary Table 6). To verify whether a causal relationship existed between the abundance of GU174097_g or the levels of ketone bodies and diabetes risk indicators, two-sample MR analysis was performed. Extensive assumption checks were conducted to enhance the validity of the two-sample MR analysis (Supplementary Table 7). No weak instrument bias was observed (F-statistic >10). However, NOME assumptions were violated for all tests because GU174097_g, 3-hydroxybutyrate, and acetoacetate had seven, eight, and five SNPs as their instrument variables, respectively, and these values were not sufficiently large for I2 > 90. In this case, if heterogeneity exists, MR–Egger (SIMEX) is recommended; otherwise, IVW is recommended. As the InSIDE assumption cannot be statistically tested38, the weighted median method—a robust approach used in cases of InSIDE assumption violation—has to be considered with each recommended method36. Therefore, MR–Egger (SIMEX) was used to estimate the causal effect of 3-hydroxybutyrate on 60-min insulin as well as that of acetoacetate on HbA1c levels. The IVW method was used to estimate all other causal effects. To determine the causal effect of 3-hydroxybutyrate on 60-min insulin, rs2259835 was detected as an outlier via MR-PRESSO at a significance level of 0.05 (Supplementary Table 8). Thus, rs2259835 had to be removed to prevent potential horizontal pleiotropy. The result of MR-PRESSO is shown in Supplementary Table 9 and shows the estimates without outliers. The effect of acetoacetate on the HbA1c level was the only significant effect at an FDR-adjusted significance of 0.05, indicating that acetoacetate increases HbA1c levels (Supplementary Table 9). The results obtained using the weighted median method corroborated this significant association (p = 0.0475).


Recent microbiome studies have shown that T2D is associated with gut dysbiosis39,40,41 that can result in altered intestinal barrier function and a dysregulation of host metabolic and signaling pathways42. Intestinal bacteria can promote insulin resistance by triggering inflammation via polysaccharides, which are components of the gram-negative bacterial cell wall43. Furthermore, microbiota-derived EVs are expected to affect insulin resistance and provide a more in-depth understanding of T2D pathogenesis11. Various bacterial metabolites, such as short-chain fatty acids (SCFAs), can modulate the function of various signaling pathways implicated in human health and can protect against insulin resistance43,44.

The human microbiota is highly variable, and this variability is determined by various external factors, such as diet, exercise, mobility, medication, and microbial cooccurrence patterns. Many of these external factors also determine the risk of metabolic disease and are age-related45; that is, the intestinal microbiota and host phenotype are substantially altered with aging. Furthermore, the effect of the intestinal microbiota on the host phenotype is also dependent on the age of the host. The estimation of within-subject covariate effects represents a robust approach against between-subject confounders, and longitudinally measured microbiome data enable characterization of the effects of the microbiota on host disease risk. As most existing studies have been cross-sectional in nature, the validity and interpretation of their results are limited. In turn, longitudinal studies are needed to comprehensively investigate the association between the human microbiome and diseases, including T2D.

Our longitudinal study revealed that a low abundance of GU174097_g is a risk factor for T2D development. GU174097_g has not been cultured to date. Multiomics data, including host genomic data, T2D-related metabolites, clinical information, and predicted functional metagenomic profiles, were utilized to extensively validate our results via causality analysis. GU174097_g is a member of the family Lachnospiraceae, and an association between Lachnospiraceae and T2D risk has been reported in several previous studies46,47. SCFA pathways, including the propanediol and acrylate signaling pathways, play important roles in mediating the effects of Lachnospiraceae on T2D45. Additionally, SCFA-producing bacteria affect epigenetic regulation in T2D patients and reduce the risk of developing T2D44,48. We found that GU174097_g is positively correlated with the 60-min insulin level, and in turn, it is negatively correlated with HbA1c levels. This indicates that GU174097_g reduces HbA1c levels and, thus, the risk of developing T2D by stimulating insulin secretion.

Next, we aimed to elucidate how GU174097_g affects T2D through the regulation of 60-min insulin and HbA1c. Multiple mechanisms may underlie these associations, including the effects of various microbiota-derived metabolites, including SCFAs, as previously suggested. In addition, ketone bodies have been reported not only as indicators of diabetic hyperglycemia but also as markers of disturbed glucose metabolism in the prediabetic state37,48. Furthermore, fatty acid metabolism, CoA synthesis, and oxidative phosphorylation, all of which are involved in ketogenesis or ketolysis, have been associated with T2D49. In our study, the ketone bodies 3-hydroxybutyrate and acetoacetate were negatively correlated with GU174097_g but positively correlated with the 60- and 120-min glucose levels. MR analysis was employed to investigate the effects of GU174097_g and ketone bodies on diabetes risk indicators. Although no causal relationship was observed between GU174097_g and ketone bodies or other clinical variables, acetoacetate was found to be causally related to an increased HbA1c level. HbA1c level is a major biomarker of T2D and explains the microbial beta-diversity. Furthermore, GU174097_g was negatively correlated with acetoacetate. Therefore, our study not only confirmed the importance of ketone bodies in T2D pathogenesis but also suggests an underlying mechanism for the association between GU174097_g and T2D development.

Previous studies have reported that gut microbe–derived EVs can infiltrate the circulatory system through the gut barrier11,50. Furthermore, microbe-derived EVs in urine can reflect the lung and gut microbiota of children with asthma51. Interestingly, T2D increases the co-occurrence of the same OTUs within the gut microbiome and microbe-derived EVs in urine samples52, which indicates that these EVs may reflect the gut microbiota composition. Coprococcus, a member of the Lachnospiraceae family, is one of the major butyrate-producing bacteria. It is known to utilize metabolic intermediates essential for the synthesis of ketone bodies, such as acetoacetyl-CoA, 3-hydroxybutyryl-CoA, and crotonyl-CoA26, as energy sources to produce the SCFA butyrate. SCFAs are considered beneficial for health and are considered to protect against T2D53. Thus, we hypothesize that GU174097_g consumes acetoacetate to produce SCFAs. These SCFAS can promote insulin secretion and decrease HbA1c levels, leading to a decreased risk of T2D.

Our study had several limitations. First, as it was based on the metagenomic profiles of EVs, the microbial compositions observed can differ from, and need to be further compared with, those of the intestinal microbiota. Second, as the genus-level taxonomy of GU174097_g is unknown, ecological and biological information on this species is limited. Third, published summary statistics of microbial GWAS are limited, and the sample size in the current microbial GWAS was small. Therefore, the number of SNPs used as instrumental variables in our MR analysis was also suboptimal. Future studies should include a large sample size to identify more associated SNPs and increase the power of MR analysis. Therefore, the mechanisms underlying T2D pathogenesis could be further identified and characterized. Fourth, even though extensive methods were used to validate assumptions in our MR analysis and enhance the validity of causal analysis, the MR results were not easy to interpret. Ketone bodies and diabetes risk indicators were highly correlated and interacted with each other. Additional in vivo and in vitro experiments may clarify the associations identified herein.

Our study revealed that GU174097_g, an unclassified Lachnospiraceae, is associated with T2D and ketone bodies. Furthermore, we found a potential causal relationship between ketone body acetoacetate and HbA1c levels. Our findings indicate that GU174097_g may lower the risk of developing T2D via the reduction in ketone body levels. Although the mechanisms by which GU174097_g and ketone bodies affect T2D development have not been elucidated, further large-scale longitudinal studies as well as in vivo and in vitro experiments could contribute to unraveling these mechanisms.