Multiomics Analyses With Stool-Type Stratification in Patient Cohorts and Blautia Identification as a Potential Bacterial Modulator in Type 2 Diabetes Mellitus

Type 2 diabetes mellitus (T2DM) is defined solely by high blood glucose but is associated with a range of other clinical manifestations that demonstrate large interindividual variations, hampering precision treatment (1). Patient stratification would provide a solution to this problem. Several methods were proposed to subclassify T2DM, disentangling clinical variability (2). In particular, Ahlqvist et al. (2) classified Swedish individuals with nonautoimmune diabetes into four clusters using five easily measurable clinical parameters selected based on clinical experience and knowledge, including age at diabetes onset, glycosylated hemoglobin (HbA_1c), BMI, and measurements of insulin resistance (IR) and insulin secretion (IS). Others made efforts to disentangle heterogeneity by stratifying the population according to risk loci in the human genome, but this approach suffers from low reproducibility and utility (3). In the meantime, the gut microbiota, playing an important role in host metabolism and immunity and a promising source of novel therapies, has been observed to be vastly different between patients with T2DM and healthy control subjects (HCs), as well as within patients (4). Other omics data, such as metabolomics and transcriptomics, exhibit the same trends, showcasing the importance of these information levels for personalizing T2DM treatment options and, at the same time, adding to the complexity of patient stratification (1). Unfortunately, we are still a long way from the implementation of multiomics in diagnostic and therapeutic practice because of their costly measurement and complicated integration. As an alternative, we explore here the explanatory and predictive power on multiomics heterogeneities of some easily accessible parameters in order to use them for stratification of patients with T2DM in novel therapies involving gut microbiota. In addition to the diabetes-associated parameters commonly measured in clinical practice, including BMI, fasting blood glucose, IR, IS, and the existing subtypes, such as Ahlqvist subtypes, we assess other candidate parameters, especially those related to the vast variability in gut microbiota, as potential stratification parameters. One such parameter is stool consistency, which is firmly associated with enterotypes, richness, and bacterial growth rates of gut microbiota (5). A population-level analysis of gut microbial variation identified it as the parameter with the largest explanatory power (6).

Stool consistency, evaluated by the Bristol Stool Scale, is a proxy for colon transit time (CTT). Longer CTT causes more water absorption, rendering harder stools. In healthy individuals, varying CTT and stool consistency are associated not only with differences in community composition (5,6) but also in metabolic potential of gut microbiota (7,8). Both metagenomic and metabolomic analyses linked enhanced proteolytic activity to longer CTT because of a shift of microbial metabolism from saccharolytic to proteolytic fermentation forced by the exhaustion of fermentable substrates with longer CTT (7,8). Simultaneously, abnormal stool consistency has been observed in diseases; for example, constipation is proposed as a risk factor and a prodromal marker for Parkinson disease (9). However, there is a gap in interpreting diverse stool types under the pathophysiological context of diseases. Two stool-type extremes (dry or loose) are reportedly common and long lasting in T2DM according to population-based surveys (10), possibly related to underlying conditions, including obesity (11). Hence, we investigate here the stratification effects and practical implications of stool type, a feature of the microenvironment to which gut bacteria are exposed, in diagnosis and treatment.

Our analyses led to the identification of a potential bacterial modulator, namely Blautia. To advance microbial intervention in T2DM, many researchers have explored probiotics before. Bacteroides, Bifidobacterium, Akkermansia, Faecalibacterium, and Roseburia abundances have been shown to be reduced in T2DM in several studies (12). The favorable effects of these bacteria, including the inhibition of proinflammatory cytokines, the improvement of intestinal barrier functions, the maintenance of glucose homeostasis, etc., have been independently confirmed in T2DM mouse models (12). Bifidobacterium and Akkermansia have successfully been used in probiotic formulations for T2DM (12). Other bacteria, including Lactobacillus and Blautia, have presented inconsistent effects in different T2DM studies (12,13). Notably, Blautia, mainly consisting of several species previously belonging to the genus Clostridium or Ruminococcus, has been frequently noticed in microbial studies on metabolic diseases in the past decade (13). A recent study discovered the protective roles of Blautia in obesity and T2DM in a human cohort and animal models (14), but some earlier gut metagenomic studies reported the opposite results (12). Hitherto, cross-sectional studies of patients versus control subjects overlooked the individual differences in microbiota and host parameters in T2DM cohorts, which might account for the variable findings of previous microbial studies. In this study, we investigated effective parameters for patient stratification and potential bacterial modulators for precision treatment, leveraging both multiomics data of a cross-sectional human cohort and bacterial intervention experiments in mouse models.

We recruited 103 patients with new-onset, untreated T2DM and 25 HCs who met specific inclusion criteria (Supplementary Material). Patients, who were initially screened from 1,798 individuals in Beijing from July 2015 to December 2016, self-evaluated their stools with the initial guidance of clinicians and went through a 2-week screening performed by clinicians. To ensure that we recorded the representative and stable stool type for each patient over time, we finally enrolled patients with T2DM who self-reported consistently dry stool (DM-DS) or loose stool (DM-LS) over the past 3 months and maintained unchanged stool type at the beginning and end of the screening (DM-LS, n = 79; DM-DS, n = 24). HCs with normal stools were healthy volunteers who attended routine examinations in Guang’anmen Hospital from February to May 2016. The clinicians assessed the consistency of the collected stool samples using Bristol Stool Scale scores and confirmed the enrollment of all participants. The study followed the principles of the Declaration of Helsinki, was approved by the ethics committee of Guang’anmen Hospital (No. 2015EC060-02), and is registered in the Chinese Clinical Trial Registry (ChiCTR-IOR-15006626). The blood, stool, and urine samples of each participant were collected for the subsequent multiomics detection (Fig. 1A and Supplementary Material).

We measured HbA_1c, fasting plasma glucose (FPG), insulin, cholesterol, LDL, HDL, and triglyceride (TG). We used the HOMA2 method to quantify β-cell function (HOMA2-β), IS (HOMA2-IS), and IR (HOMA2-IR) based on FPG and fasting insulin levels. T2DM biomarkers, including C-peptide, ghrelin, gastric inhibitory peptide (GIP), glucagon-like peptide 1, glucagon, insulin, leptin, plasminogen activator inhibitor 1 (PAI-1), resistin, and visfatin, were also measured in serum samples.

To evaluate the host inflammation, we measured the concentrations of 27 cytokines in serum, namely eotaxin, fibroblast growth factor 2, granulocyte colony–stimulating factor, granulocyte-macrophage colony–stimulating factor (GM-CSF), interferon-γ, interleukin-10 (IL-10), IL-12, IL-13, IL-15, IL-17α, IL-1β, IL-1rα, IL-2, IL-4, IL-5, IL-6, IL-7, IL-8, IL-9, interferon γ–induced protein 10, monocyte chemoattractant protein 1, macrophage inflammatory protein 1α, macrophage inflammatory protein 1β, platelet-derived growth factor bb, RANTES (regulated upon activation, normal T-cell expressed and secreted), tumor necrosis factor α, and vascular endothelial growth factor (Supplementary Material).

We performed 16S rRNA gene sequencing of stool samples from both human participants and mice, and got the taxonomic profile using Qiime2 (version 2020.8.0) (15) based on the Greengenes database (version 13.8.99) (16). PICRUSt2 was adopted to perform microbial functional profiling (17–19). We performed RNA sequencing of blood samples and used Trimmomatic (20), HISAT2 (21), and StringTie (22) for quality control, alignment, and quantification, respectively. We also measured microbial and host metabolites in stool, serum, and urine samples. Detailed methods for multiomics detection and processing can be found in the Supplementary Material.

Blautia producta (BP) (DSMZ 2950), a type strain of Blautia recognized as safe for use in food and medicine (23), was cultivated in a modified yeast casitone fatty acid (FA) medium, harvested in log phase, aliquoted, and stored in 10% glycerol at −80°C. Eighteen C57BL/6J male mice (18–22 g, 7 weeks old) were purchased from the Nanjing Biomedical Research Institute of Nanjing University. The experiment was approved by the ethics committee of Guang’anmen Hospital (No. IACUC-GAMH-2021-003). The mice were acclimatized in a specific pathogen-free facility, with controlled climate and light conditions (24 ± 2°C, 60–70% relative humidity, and 12-h light/dark cycles), for 1 week and were subsequently randomly divided into three groups (n = 6) as follows: 1) control group with a normal diet (10% fat) and sham gavage (1% glycerol in PBS), 2) diabetes group with an high-fat diet (HFD) (60% fat, D12492; Research Diets, New Brunswick, NJ) and sham gavage, and 3) BP group with an HFD and BP gavage (5 × 10¹¹ colony-forming units/kg body weight/day, 1% glycerol in PBS). The mice were subjected to sham or BP gavage fed orally once a day for 10 weeks.

Before the intervention, we collected the fecal pellets and measured the initial body weight and FPG for each mouse. During the intervention, the body weight and FPG were measured weekly and biweekly, respectively. After the intervention, we collected the feces and performed oral glucose tolerance tests (OGTTs) for each mouse. The mice were fasted for 12 h with free access to water and gavage fed with 3 g glucose/kg lean mass. Blood glucose was measured from tail vein blood before and 15, 30, 60, and 120 min after the glucose bolus. Then, we anesthetized mice with diethyl ether, collected ophthalmic artery blood, and centrifuged the serum within 15 min. The serum samples were used for the subsequent detection of cholesterol and LDL. The fecal samples collected before and after the intervention were used for 16S rRNA sequencing of the gut microbiome according to the detailed methods in the Supplementary Material (Fig. 1B).

For the analyses of the data from human participants, continuous variables, represented by mean ± SD, were tested by the two-tailed t test with Welch correction or Wilcoxon rank sum test, and discrete variables were compared using the χ² test or two-tailed Fisher exact test. P ≤ 0.05 or false discovery rate–adjusted P_adj  ≤ 0.05 represented statistically significant results. We applied the above statistical approaches to the microbial, metabolic, and clinical data. Differentially expressed genes (DEGs) in the transcriptome were additionally identified using R package ‘DESeq2’ (version 1.24.0), with P_adj  ≤  0.01 and abs(log₂-transformed fold change [FC])  ≥ 1.5 as thresholds. We performed enrichment analysis on DEGs against the Kyoto Encyclopedia of Genes and Genomes (KEGG) database using R package ‘clusterProfiler’ (version 3.16.1). All features were centered log-ratio (CLR) transformed and scaled to unit variance using clr in R package ‘compositions’ (version 2.0.1) in multivariable analysis. The variance of the CLR-transformed variable was used to describe the dispersion of each analyte. To further quantify the deviation in multiomics composition between groups, we performed orthogonal partial least squares discriminant analysis (OPLS-DA) on the CLR-transformed and scaled profile using opls in R package ‘ropls’ (version 1.22.0) and evaluated the performance using R²Y and Q² matrices.

The clinical covariates of multiomics, promising for T2DM stratification, were determined by linear correlation analysis (LCA) and permutational MANOVA (PERMANOVA), the two most commonly used methods (6). For the former, we calculated correlations between clinical variables and composition ordination of each omics type (principal component analysis based on Bray-Curtis dissimilarity) using envfit in R package ‘vegan’ (version 2.5.7) (10,000 permutations, P ≤ 0.05). For the latter, we used adonis2 in R package ‘vegan’ (version 2.5.7) to determine the proportion of explained variance and significance for each clinical covariate on Bray-Curtis dissimilarity matrices of each omics (10,000 permutations, P ≤ 0.05). Ahlqvist subtype was included as a discrete variable among other candidate clinical variables. The P values in the covariate analysis were also adjusted by false discovery rate correction. In the Ahlqvist subtyping of patients with T2DM (2), each patient was first characterized by five standardized (mean 0, SD 1) variables: age at diagnosis, BMI, HbA_1c, HOMA2-β, and HOMA2-IR. We partitioned the T2DM objects into several clusters using pam in R package ‘cluster’ (version 2.1.1). The optimal cluster number (k = 4) was determined by pamk in R package ‘fpc’ (version 2.2.9).

Binary stool type was the most significant covariate identified in our analyses, outperforming other clinical parameters. Therefore, we evaluated the discrimination of stool-type–stratified T2DM subgroups (DM-DS and DM-LS) on multiomics data using machine learning (ML) algorithms. After the comparison of different ML algorithms, including logistic regression, ridge regression, random forest, and LightGBM (24), we opted for LightGBM classifiers to distinguish DM-DS, DM-LS, and HC, incorporating all discriminatory microbial and/or host features. The multiomics features were first filtered by statistical approaches as mentioned above, then selected and ranked by feature_importance in the ‘lightgbm’ Python package. The comparison of different ML algorithms, splitting of training and independent test sets, fivefold cross validation, and feature selection process are described in detail in Supplementary Material.

Single- and multiomics associations were assessed through Spearman and sparse correlations for compositional data (SparCC) correlations and association networks. Pairwise Spearman rank correlation between any two CLR-transformed variables was calculated using corr.test in R package ‘psych’ (version 2.0.12). Absolute Spearman correlation (R) ≥0.4 with P ≤ 0.05 was considered significant and meaningful. To determine more reliable co-occurrences among bacteria, we additionally calculated SparCC correlation using the Python-based SparCC tool, where the absolute SparCC correlation coefficient (R) ≥0.4 with P ≤ 0.05 was used as the significant cutoff (25). The bacteria-bacteria co-occurrences detected by only one approach were removed. After that, a large-scale cross-measurement association network was integrated by the correlations between any two variables in six measurements, including genera, gut microbial metabolites (short-chain fatty acids [SCFAs]) and secondary bile acids [BAs]), host metabolites, cytokines, T2DM-associated biomarkers, and clinical phenotypes. We used Cytoscape (version 3.8.2) to visualize the association networks (26). The overlap of the top three nodes based on maximal clique centrality, maximum neighborhood component, and density of maximum neighborhood component algorithms were identified as hub objects using the cytoHubba plugin (version 0.1) in Cytoscape. To disclose more bacterial regulation, we identified the shortest pathways between microbes or their metabolites and T2DM-related clinical parameters in the cross-measurement network using get.all.shortest.paths in R package ‘igraph’ (version 1.2.6).

For the analyses of the data from the mouse intervention, variables, depicted by mean ± SE, were compared among groups using one-way ANOVA followed by Tukey multiple comparison test. After the taxonomic profiling of microbiota using QIIME2 (version 2020.8.0) and the functional profiling using PICRUSt2, we performed the Wilcoxon signed rank test on the paired microbial data to detect taxa/pathways significantly altered after a certain intervention and performed Wilcoxon rank sum test on the changes of the microbial data to detect the taxa/pathway change differentially in different interventions. The microbial compositional deviations after intervention were assessed by the analysis of similarity approach using anosim in R package ‘vegan’ (version 2.5.7) after principal component analysis based on Bray-Curtis dissimilarity. P ≤ 0.05 was used as the significant cutoff.

All data were deposited to the National Center for Biotechnology Information Sequence Read Archive under BioProject number PRJNA795858.

This study enrolled 103 patients with T2DM and 25 HCs and measured and analyzed multiomics data for each participant (Fig. 1A and Supplementary Material). We performed 42 clinical laboratory tests, including T2DM-related phenotypes and biomarkers (Supplementary Table 1). Regarding multiomics measurements, we identified 96 genera from 16S rRNA sequence data of stool samples, 58,395 genes from whole-blood RNA sequences, 27 cytokines in serum, 5 kinds of SCFAs and their isomers, 19 kinds of secondary BAs in serum and stool, and 294 host metabolite features, including 226 in serum (amino acids [AAs], arachidonic acids [ARAs], primary BAs, choline, phospholipids, medium-chain FAs [MCFAs], and total FAs), 11 in stool (primary BAs, aromatic amino acids [AAAs] and their derivatives), and 23 in urine (AAAs and their derivatives) (Supplementary Table 2). To explore the regulation and therapeutic potential of microbial modulators inferred from multiomics analyses, we performed a microbial challenge in a T2DM mouse model (Fig. 1B).

Assessing variation on multiomics levels, we observed low variability in most clinical phenotypes (Fig. 2A), revealing several common clinical manifestations in our patients with recent-onset T2DM. The patients generally exhibited impaired β-cell function, severe IR, or both. Compared with HCs, most patients with T2DM reached a higher level of IR (HOMA2-IR) for every value of the β-cell function (HOMA2-β) and a lower level of β-cell function (HOMA2-β) for every value of IR (HOMA2-IR) (Fig. 2B). Additionally, the T2DM group showed significantly higher levels of HbA_1c, FPG, cholesterol, TG, hip circumference, BMI, waist circumference, weight, fasting insulin, PAI-1, and resistin and lower levels of ghrelin (Supplementary Table 1). More varied signatures were observed in other omics data of T2DM, especially in microbial taxa and metabolites, and host transcripts (Fig. 2A). These profiles thus hold considerably more information with stratification potential, and we set out to assess the power of accessible clinical variables and existing subtyping methods to explain these heterogeneities.

To examine the explanatory power of the candidate parameters on variation in multiomics, we performed LCA and PERMANOVA, the two most commonly used methods (6) (Fig. 2C–E). We first assessed the power of the standard subtyping method, Ahlqvist subtyping. This was also the only assessable and applicable one for our data. Applying the Ahlqvist subtyping, determining T2DM subtypes using five commonly available parameters, our patients with T2DM were divided into four expected subtypes recapitulating the subtype-specific characteristics defined by Ahlqvist subtyping (Supplementary Fig. 1). We found that Ahlqvist subtype captured the variation of gut microbiota (P < 0.05) and serum choline (P < 0.05, P_adj < 0.1), with an effect size of 1.93% and 2.85% using the PERMANOVA method (Fig. 2C and E). The subtype was also identified as the covariate of serum choline (P < 0.01, P_adj < 0.1) and AAs (P < 0.001, P_adj < 0.01), with an effect size of 21.25% and 25.97% in LCA (Fig. 2E).

We further identified the clinical covariates with the greatest effect sizes for each omics layer to determine the optimal parameter for patient stratification. Our LCA revealed that among the parameters considered, the binary stool type showed the most robust explanatory power across multiple omics layers, encompassing both the gut microbiome (P < 0.01, P_adj < 0.05) and host transcriptome (P < 0.001, P_adj < 0.01) (Fig. 2C and D). For the metabolic layer, stool type was the most important covariate of serum FAs (P < 0.001, P_adj < 0.1) (Fig. 2E), while Ahlqvist subtype showed the largest effect sizes covarying with serum choline (P < 0.01, P_adj < 0.1) and AAs (P < 0.001, P_adj < 0.01) (Fig. 2E). Other parameters, including GIP, BMI, waist circumference, ghrelin, C-peptide, TG, leptin, height, and diastolic pressure, explained the most variation for stool BAs, AAAs, and SCFAs and serum ARAs, MCFAs, phospholipids, SCFAs, and BAs, respectively (P < 0.05) (Fig. 2E). The PERMANOVA method also proved stool type as the strongest covariate of the microbiota (P < 0.05) and transcriptome (P < 0.001, P_adj < 0.01) (Fig. 2C and D) and identified some additional parameters as the most important covariates for metabolic profiles, including systolic pressure for stool BAs and serum choline, and HbA_1c, fasting insulin, hip circumference, TG, and PAI-1 for serum AAs, ARAs, BAs, and FAs, respectively (P < 0.05) (Fig. 2E).

Our covariate analyses demonstrated that stool type could achieve better stratification performance than the state-of-the-art subtyping method, Ahlqvist subtyping. Particularly, LCA revealed stool type to be the sole covariate of the gut microbiome, with an effect size of 13.93% (P < 0.01, P_adj < 0.05) (Fig. 2C), and the PERMANOVA method showed that stool type explained more gut microbiome variation than Ahlqvist subtyping (effect size 2.26% vs. 1.93%) (Fig. 2C). Additionally, while the Ahlqvist subtype failed to capture the variation in the host transcriptome data, stool type significantly covaried with the transcriptome data, reaching remarkable effect sizes of 40.2% and 6.01% in LCA and PERMANOVA, respectively (P < 0.001, P_adj < 0.01) (Fig. 2D). Moreover, although the Ahlqvist subtype explained more variation in serum choline (P < 0.01, P_adj < 0.1 in LCA; P < 0.05, P_adj < 0.1 in PERMANOVA) and AAs (P < 0.001, P_adj < 0.01 in LCA), stool type covaried with serum FAs (P < 0.001, P_adj < 0.01 in LCA; P < 0.05 in PERMANOVA) and phospholipids (P < 0.05 in LCA), two metabolite categories containing the largest number of metabolite species in our data (P < 0.05) (Fig. 2E). Overall, stool type had a stronger explanatory power than the Ahlqvist subtype on extensive omics profiles, despite the exceptions in serum choline and AA. Stool type, therefore, is an important variable to stratify T2DM cohorts, especially in gut metagenome-wide association studies.

Stratification based on stool type revealed differences in clinical laboratory tests that are biologically relevant. Specifically, we found that two T2DM biomarkers, glucagon and visfatin, were enriched in DM-LS compared with DM-DS (HC, n = 25; DM-DS, n = 24; DM-LS, n = 79; P < 0.05, Wilcoxon rank sum test) (Fig. 2F). Higher glucagon secretion is generally thought to be detrimental in T2DM, contributing to the progression of hyperglycemia and IR (27). A meta-analysis demonstrated a positive correlation between circulating visfatin and IR (28). We also observed a significant elevation of IR in DM-LS compared with HC, with slightly higher levels in DM-DS. Together, these alterations suggest a more serious IR of DM-LS than DM-DS. Furthermore, DM-LS was identified as having lower levels of HOMA2-IS and higher levels of cholesterol, BMI, body weight, and waist circumference than HC, indicating lower insulin sensitivity and aggravated obesity in DM-LS (P < 0.05, Wilcoxon rank-sum test) (Supplementary Table 1). No other significant differences were found in demographic characteristics and other clinical parameters between the two stool-type–divided subgroups.

To further investigate the discrimination of stool type across multiomics, we used ML algorithms, based on heterogeneous fecal microbiota taxonomic and/or host features, to discriminate DM-DS, DM-LS, and HC from one another. LightGBM (24) outperformed other commonly used ML algorithms, including logistic regression, ridge regression, and random forest, in this study (Supplementary Fig. 2). Fivefold cross validation, in which LightGBM models stably performed well (Supplementary Fig. 3), was used to select key discriminatory features. This method identified 3 genera, 41 metabolites, and 6 genes to construct DM-DS/HC classifiers, with an area under the curve (AUC) of 0.900, 0.911, and 0.813, respectively (Fig. 2G and Supplementary Table 2). The integration of microbial and host features improved AUC to 0.980. Comparatively, predicting DM-LS and HC proved less challenging. The classifiers, based on 8 genera, 59 metabolites, 196 genes, or the combination, distinguished DM-LS from HC with higher AUCs of 0.952, 0.989, 0.986, and 1.000, respectively (Fig. 2H and Supplementary Table 2). Finally, the DM-DS/DM-LS classifiers, composed of 10 discriminatory genera, 37 metabolites, 171 genes, or the integrated features, separately reached satisfactory AUCs of 0.935, 0.909, 0.935, and 0.986 (Fig. 2I and Supplementary Table 2). Subsequent analyses showed that DM-DS and DM-LS were clearly discrepant in multiomics characteristics. Several of the identified discriminating factors have been described previously in T2DM- or health-associated mechanisms. Blautia, a bacterium potentially inversely linked to metabolic syndrome (13,14,23), was identified as the genus with the most discriminatory power using LightGBM models (13) (Fig. 2J). Additionally, several serum FAs, phospholipids, and cysteine, which influence the homeostasis of glucose and lipid (29–31), were identified as the most critical host metabolites (Supplementary Table 2).

We analyzed the gut microbiota as characterized by amplicon sequence variant–level features (HC, n = 25; DM-DS, n = 24; DM-LS, n = 79). A reduced species richness (Chao1 index) was observed in DM-LS compared with both DM-DS and HC (Fig. 3A). Lower species richness is often associated with disease and is generally thought to cause decreased resilience (32). Next, we assessed co-occurrences among gut microbiota through both Spearman and SparCC methods, and found these relationships to be the weakest in DM-LS (Supplementary Table 3). Our previous microbiome analyses linked the co-occurrence network to gut health, revealing a reduction in bacterial co-occurrences in disease samples (18,19). Furthermore, OPLS-DA was performed to evaluate the variation in community composition among the HC, DM-DS, and DM-LS groups. The overall microbial taxonomic structure of DM-DS and DM-LS significantly deviated from HC as well as from each other (Fig. 3B and Supplementary Fig. 4A and B). Compared with HC, three genera and eight genera were differentially abundant in DM-DS and DM-LS, respectively (Supplementary Table 2). Among them, Bacteroides decreased concordantly in stool-type–stratified T2DM subgroups. Ten genera were distinguished between DM-DS and DM-LS (Supplementary Table 2). For these 10 genera, 7 increased in DM-DS, including Blautia, [Ruminococcus] (names in square brackets are contested names according to Greengenes), Dorea, Ruminococcus, Faecalibacterium, Bifidobacterium, and Collinsella, while the remaining 3 were more abundant in DM-LS, including Bacteroides, Parabacteroides, and Phasolarctobacterium (Fig. 3C). The increased Ruminococcus and decreased Bacteroides in DM-DS are consistent with previous reports on the gut microbial composition of Belgian individuals with dry stools (5,6). Among the genera increased in DM-DS versus DM-LS, most are known to be associated with health benefits (12,13). Particularly, Bifidobacterium and Faecalibacterium are the genera most consistently reported to be negatively associated with T2DM in microbiome studies of human cohorts (12). Bifidobacterium, which showed positive effects in some animal experiments and limited, but valuable randomized crossover trials, has been introduced as a T2DM probiotic (12). Blautia abundances have been mostly negatively linked to obesity-related diseases, such as T2DM and nonalcoholic fatty liver disease, despite the reverse findings in some studies (13). The common advantageous mechanisms of these bacteria include the synthesis or upregulation of the favorable metabolites (12,13) SCFAs and secondary BAs, which have anti-inflammatory properties and protective effects on glucose-lipid metabolism and host energy homeostasis against metabolic diseases, such as T2DM (33).

We further assessed alterations in functional profiles predicted by PICRUSt2 (17,18). Using the OPLS-DA method, we found that the predicted microbial gene composition of DM-LS deviated more from HC compared with DM-DS (Fig. 3D). The functional gene profiles were preserved between the HC and DM-DS groups but largely altered in the DM-LS group. Using the predicted cellular processes and metabolic pathways, we further found different microbial functions between DM-DS and DM-LS samples. Compared with DM-DS, 10 processes or pathways were upregulated and 15 were downregulated in DM-LS (Fig. 3E). Notably, we observed that the cell motility process was significantly elevated in DM-LS (Fig. 3E), while the cell growth process was upregulated in DM-DS (Supplementary Table 4). It is reasonable to infer that DM-LS had higher levels of motile bacteria, while DM-DS seemed to have a higher investment in cell growth. These might be related to alternative microbial life strategies selected by the intestinal environment controlled by host physiology. In addition, several different pathways, including the glycolysis/gluconeogenesis, sulfur metabolism, butanoate metabolism, and secondary BA biosynthesis pathways, participating in the metabolism of important gut microbial metabolites, SCFAs, and secondary BAs were inconsistently altered between the two stool groups (Fig. 3E and Supplementary Fig. 5).

To investigate the metabolism and host use of SCFAs and secondary BAs further, we also measured their circulation concentration in serum and excretion concentration in stool samples. We detected five kinds of SCFAs and their isomers and 19 kinds of secondary BAs (Supplementary Table 2). Compared with DM-LS or HC, DM-DS exhibited lower fecal excretion but higher serum circulation with regard to total SCFAs, acetic acid, propionic acid, and butyric acid (Fig. 3F). Reversely, DM-LS displayed the concentrations of SCFAs as higher in stool but lower in serum. Intriguingly, a similar observation could be made for the total secondary BAs (Fig. 3F). Especially, serum glycoursodeoxycholic acid, ursodeoxycholic acid, and tauroursodeoxycholic acid were more abundant in DM-DS than DM-LS (Supplementary Table 2). In conclusion, the integrated analysis of stool and serum samples revealed a higher circulation load of these generally recognized as favorable microbial metabolites in DM-DS and a more significant excretion in DM-LS.

Metabolic disorder and chronic low-level inflammation are two aspects that affect the onset and progression of T2DM (34). We investigated host cytokines, metabolites, and transcripts.

We evaluated host inflammation by measuring 27 cytokines in serum (HC, n = 25; DM-DS, n = 24; DM-LS, n = 79) (Supplementary Table 2). IL-1b, IL-13, and GM-CSF were consistently elevated in DM-DS and DM-LS compared with HC. The increased cytokines IL-1b and GM-CSF are regarded as proinflammatory (35,36). Samples of the DM-DS and DM-LS groups did not differ significantly with regard to cytokine profiles, but the majority showed slightly higher concentrations in the DM-LS group (Supplementary Table 2).

We performed both OPLS-DA and differential abundance analyses through the Wilcoxon rank sum test to evaluate host metabolism on the metabolic profiles of the stool, serum, and urine samples (HC, n = 25; DM-DS, n = 24; DM-LS, n = 38). OPLS-DA revealed that the overall metabolic composition was different among the HC, DM-DS, and DM-LS groups, but DM-LS deviated more from HC than DM-DS (DM-LS vs. HC: R²Y = 0.72, Q² = 0.511; DM-DS vs. HC: R²Y = 0.617, Q² = 0.336; DM-LS vs. DM-DS: R²Y = 0.724, Q² = 0.231) (Fig. 4A). The variations among the three groups were most apparent for AAs, FAs, and phospholipids in serum (Supplementary Fig. 6A–C), with more upregulated phospholipids and AAs in the DM-LS group and more elevated FAs in the DM-DS group. The patient samples differed from HC regarding serum MCFAs (Supplementary Fig. 6D). For the metabolite profiles of serum ARAs, serum choline, stool AAAs, and urine AAAs, the three groups were indistinguishable (Supplementary Fig. 6E–H). The separate analysis results of serum, stool, and urine samples are described in the Supplementary Material.

Compared with HC, 81 metabolites were differently abundant in DM-DS (55 upregulated and 26 downregulated), and 103 metabolites were different in DM-LS (84 upregulated and 19 downregulated) (P < 0.05, two-tailed Wilcoxon rank sum test) (Fig. 4B and Supplementary Table 2). Among them, 51 metabolites were concordantly altered in DM-DS and DM-LS (42 upregulated and 9 downregulated). Significantly, some deleterious metabolites in serum, including trimethylamine-N-oxide (TMAO) and several AAs (alanine, aspartic acid, glutamic acid, isoleucine, leucine, phenylalanine, tyrosine, and valine) and phospholipids, especially phosphatidylethanolamine subclass, were increased in our patients with T2DM. TMAO is a gut microbial product of choline and can induce inflammation (37). A previous metabolic analysis reported negative associations between IS and the above-mentioned serum AAs (31). Some plasma lipid profiling studies in large population-based cohorts linked phosphatidylethanolamine to T2DM and obesity (38). The concordantly decreased metabolites contained MCFAs (including C7–11) and Cis-11-eicosenoic-acid in serum. MCFAs not only can promote the insulin sensitization of the liver, adipose tissue, and skeletal muscle but also can increase glucose-stimulated IS (39). Cis-11-eicosenoic-acid is generally thought to be anti-inflammatory (40). Altogether, these results show that disadvantageous metabolic disorders were common in our patients with T2DM.

Despite the consistent changes, the DM-LS group exhibited more group-specific metabolic alterations compared with the HC group, especially in metabolites disadvantageous to T2DM, than the DM-DS group (Fig. 4B and Supplementary Table 2). Of the 47 metabolites specifically changed in DM-LS compared with HC, serum phospholipids accounted for the majority. We found a significant phospholipid overload in the DM-LS group, where the phospholipid subclasses, including phosphatidylcholine, phosphatidylethanolamine, phosphatidylglycerol, and phosphatidylserine, were increased, even though the small numbers of sphingomyelin species were reduced. Notably, phosphatidylethanolamine and phosphatidylglycerol were proven to be associated with T2DM and obesity in previous population studies (39). Additionally, it is generally thought that a high circulating level of phospholipids is an important driver of lipotoxicity that impairs the tissues, including the liver and pancreatic islets (29). Furthermore, the insulin-sensitizing MCFA C12 was decreased in serum samples of the DM-LS group compared with the HC group. Of the 19 metabolites uniquely altered in DM-DS compared with HC, most were serum FAs and phospholipids. The increased FAs included C14:00, C16:00, C18:01, C18:3n-6, and C20:3n-6, and the reduced FAs included C20:01 and C24:00. Previous studies reported the associations between FA species and T2DM risk: long even-chain saturated FAs (SFAs) were associated with increased risk and very-long even-chain saturated FAs and unsaturated FAs with reduced risk (30). Summarizing the effects of both increased and decreased FAs, we found that FA alterations in the DM-DS group were more likely to increase the T2DM risk compared with the HC group. Regarding phospholipids, in addition to the species increasing concordantly with the DM-LS group, the lysophosphatide subclass was specifically reduced in the DM-DS group. Serum thromboxane B2, a potential proinflammatory ARA (41), was downregulated in DM-DS compared with HC. In summary, the comparisons of the two stool-type–stratified T2DM subgroups with the HC group revealed a more beneficial metabolic profile in individuals with DM-DS.

When comparing DM-DS and DM-LS, we found significant differences in 58 metabolites (P < 0.05, two-tailed Wilcoxon rank sum test) (Fig. 4B and Supplementary Table 2). DM-LS showed an increase in most of the differential metabolites (n = 46), especially in serum phospholipid subclasses phosphatidylcholine and phosphatidylglycerol. The higher level of circulating phospholipid in DM-LS compared with DM-DS and HC might be indicative of lipotoxicity in patients with T2DM (29). DM-LS also had several enriched AAs, such as cysteine, known to inhibit IS (31,33). In contrast, DM-DS had 12 upregulated metabolites, mainly comprising serum unsaturated FAs, which are beneficial to T2DM (30), indicating the relatively advantageous FA profiles in DM-DS compared with DM-LS, despite some disadvantageous alterations away from the HC group. The serum level of sphingomyelins was also higher in DM-DS than DM-LS, but close to HC, revealing the healthier state of sphingomyelin composition in the DM-DS group. DM-DS additionally demonstrated increased levels of insulin-sensitizing MCFAs (C8 and C12), indicating favorable metabolic alterations. Together these results indicated that the DM-LS group presented with more severe metabolic dysbiosis, which is not only a cause but also a consequence of IR (42), compared with the DM-DS and HC groups.

Host whole-blood transcriptome assessment further confirmed the more severe metabolic disorder of DM-LS (HC, n = 25; DM-DS, n = 24; DM-LS, n = 79). OPLS-DA and between-group DEG numbers showed that DM-DS was less discriminable from HC, while DM-LS was significantly distinctive from the other groups (Fig. 4C and D). Notably, most DEGs in DM-LS were upregulated (DM-LS vs. HC, n = 2,498 [97.1%]; DM-DS vs. DM-LS, n = 2,533 [98.0%]). We performed KEGG pathway enrichment for DEGs to understand their related functions and metabolic pathways. Compared with both HC and DM-DS, 20 pathways were upregulated in DM-LS (Fig. 4E). The raised protein digestion and absorption and glycerolipid metabolism were consistent with the overflow of excess levels of circulating AAs and phospholipids mentioned above and might provide insights into aspects of metabolic dysbiosis in DM-LS. We also found excess IS in DM-LS: an upregulated insulin secretion pathway and several pathways participating in insulin signaling, including cAMP signaling pathway and calcium signaling pathway. Additional evidence for increased IS lay in the moderately higher circulating levels of fasting insulin in DM-LS compared with both HC and DM-DS (Supplementary Table 1).

The host metabolomics and transcriptomics data collectively suggest more metabolic dysbiosis and excessive IS in DM-LS compared with both HC and DM-DS. Considering the vicious cycle between these disadvantageous alterations and IR (42,43), as well as the more adverse clinical manifestations of the DM-LS group, including the significantly higher levels of glucagon and visfatin, the slightly higher levels of HOMA2-IR and obesity, and the slightly lower level of HOMA2-IS (Supplementary Table 1), we have ample reason to assume that patients with loose stools generally present with an unhealthier state than those with dry stools. In general, our analyses show profound discrepancies in metabolism, and less so in inflammation parameters between stool-type–stratified patient samples, with less favorable profiles associated with loose stools.

To characterize and compare the host-microbe associations with T2DM in DM-DS and DM-LS, we constructed a large-scale cross-association network incorporating gut microbes and their metabolites (including SCFAs and secondary BAs), host metabolites, cytokines, T2DM-associated biomarkers, and clinical parameters. All the correlations in the integrated network are included in our online database DMA (Diabetes Mellitus & Microbiota Associations [https://cqb.pku.edu.cn/zhulab/info/1006/1208.htm]).

We focused on host-microbe associations involving microbes and microbial metabolites identified as differentiating between the two stool-type–stratified groups to unveil the microbiological mechanisms underlying the stratification. DM-DS had a higher number of significant correlations between microbe and host characteristics. Interestingly, the microbial parameters were often associated with T2DM-associated phenotypes and biomarkers (Fig. 5A and B). Blautia, the most important genus in distinguishing DM-DS and DM-LS (Fig. 2G), with higher abundance in DM-DS, was positively related to HOMA2-β and fasting insulin and negatively associated with glucagon in DM-DS (R = 0.551, 0.440, 0.486, and −0.536, respectively; P < 0.05) (Fig. 5A), suggesting a role for Blautia in promoting IS and lowering blood glucose. Two other bacteria, Dorea and Faecalibacterium, also engaged in the Blautia-centered association network involving HOMA2-β, HOMA2-IS, fasting insulin, and glucagon, where Blautia was recognized as the hub through the hub identification software Cytohubba (P < 0.05) (Fig. 5A and Supplementary Table 5). Dorea and Faecalibacterium were also enriched in DM-DS (Fig. 3C). These bacteria are generally known for their production of SCFA (12,13,44). In contrast, in DM-LS, gut microbes were mainly linked to serum phospholipids, and no clear associations with T2DM were found (P < 0.05) (Fig. 5B). Together, these findings suggest that different microbial community compositions are associated with DM-DS versus DM-LS, and these discordant gut microbial communities might contribute to the observed differences in host metabolic disorders, though of course, tests of causality have yet to be performed.

We identified Blautia as the most critical microbe indicator in patients with T2DM (Figs. 2J and 3 C), with potential modulating effects given its co-occurrences with other beneficial bacteria and significant correlations with T2DM phenotypes in DM-DS (Fig. 5A). To verify its regulatory effect on T2DM, we conducted a Blautia intervention experiment in an HFD-induced obesity T2DM mouse model.

We divided 18 male mice into three groups: a control group (n = 6) fed a normal diet, a diabetes group (n = 6) fed an HFD, and a BP group (n = 6) fed an HFD and BP (DSMZ 2950) gavage (Fig. 1B). Blautia was significantly increased after the intervention in the BP group, as evidenced by fecal 16S rDNA amplicon analysis, suggesting its successful colonization and growth over time (Fig. 5C). During the 10-week intervention, the weight of obese T2DM mice in the diabetes group was significantly higher than the control group, in contrast to the BP group, which only showed significantly higher weight from the 4th week onward (P < 0.05, one-way ANOVA test) (Fig. 5D). The BP group manifested significantly lower body weight than the diabetes group since the 2nd week (P < 0.01, one-way ANOVA test) (Fig. 5D). Blautia supplementation, therefore, seems to have a protective effect on weight gain induced by an HFD. In addition, serum cholesterol and LDL also remarkably decreased in the BP group (P < 0.001, one-way ANOVA test) Fig. 5E and F) compared with the diabetes group, which seemed to indicate an inhibitory effect of Blautia on lipid accumulation on an HFD. To further assess the ability of glycemic control, we measured FPG and performed OGTT. In both cases, we found a sustained reduction in the glucose level of the BP group compared with the diabetes group (P < 0.01, one-way ANOVA test) (Fig. 5G–I).

We further investigated the modulation effects of Blautia gavage on the composition and metabolic function of the mouse gut microbiota in addition to the beneficial phenotypic changes. In the BP group, Blautia increases strongly correlated with Butyricicoccus, a well-known butyric acid–producing bacterium (R = 0.975, P = 0.005, Spearman). The co-occurrences between Blautia and other bacteria were also detected in our cross-association network of the DM-DS subgroup (Fig. 5A). These relationships hinted at the collaboration between Blautia and other microbes in modifying the gut environment, which was supported by a cross-feeding assay in a recent study (14). Additionally, despite the similar microbial composition initially, the control, diabetes, and BP groups exhibited obvious compositional deviation in both longitudinal intragroup and cross-sectional intergroup comparisons (Supplementary Fig. 7A and B). Meanwhile, 39 genera from 21 families were altered differently among the three groups (P < 0.05, two-tailed Wilcoxon rank sum test) (Supplementary Fig. 7C). We noticed that several genera significantly increased specifically following the intervention in the BP group, including Blautia and Tuzzerella in the Lachnospiraceae family, Megamonas in the Selenomonadaceae family, and Candidatus Saccharimonas in the Saccharimonadaceae family. These bacteria are known for their SCFA production capabilities. Furthermore, among the predicted metabolism pathways differently altered between groups, the glucose metabolism–related pathways were abnormally altered in the diabetes group but not in the BP group, compared with the control group (P < 0.05, two-tailed Wilcoxon rank sum test) (Supplementary Fig. 7D), revealing the inhibition effect of Blautia gavage on glucose metabolism disorder. All the alterations above suggest that Blautia supplementation modulated the microbial composition, resulting in altered fermentation and microbial glucose metabolism in the stool environment. These experimental results confirmed the presumed beneficial effect of Blautia apparent from the stool-type–stratified analysis of the patients with T2DM.

In this study, we performed a cross-sectional investigation of patients with T2DM with long-lasting abnormal stools and HCs, integrating gut microbiota, gut microbiota–derived metabolites, host metabolome, host transcriptome, cytokines, T2DM biomarkers, and clinical parameters. Significant heterogeneities were observed in both host and microbial profiles, hampering the precision treatment. While the prevalent Ahlqvist subtypes, based on common clinical variables, poorly explained the heterogeneities in gut microbiota and host transcriptome data, stool-type–based stratification (DM-DS and DM-LS) better captured the variations. In-depth analyses using stool-type–based stratification revealed a more beneficial profile in DM-DS compared with DM-LS, as well as a potential bacterial modulator of T2DM, namely Blautia. Our mouse intervention experiment confirmed the presumed positive effects of this bacterium. Together, our results indicate that stool-type–based stratification has great potential in T2DM analyses and clinical practice and provide a proof of concept of a microbial precision intervention for T2DM.

Several host signatures distinguished DM-LS from DM-DS. Generally, DM-LS exhibited graver IR and more impaired IS than DM-DS. HOMA2-IR (and HOMA2-IS) values of DM-LS were significantly higher (or lower, respectively) than HC and slightly higher (or lower, respectively) than DM-DS. Similar findings were obtained for other parameters positively associated with T2DM risk or IR, including obesity, cholesterol, and fasting insulin. The circulating levels of glucagon and visfatin, positively linked to IR in previous studies (27,28), were also remarkably elevated in DM-LS compared with DM-DS, underscoring the more serious IR found in patients with DM-LS. Additionally, the upregulated IS pathway in the transcriptome of DM-LS compared with other groups, together with the phenotype of a relatively higher fasting insulin level, reflected more excessive IS in DM-LS. Furthermore, DM-LS showed more metabolic dysbiosis, namely in the raised gene expression related to the metabolism of AAs and glycerolipids in the transcriptome and the significant overload of circulating AAs and phospholipids in the metabolome, compared with HC and DM-DS. These disadvantageous metabolic alterations might cause and be aggravated by IR in a vicious cycle (42) and subsequently augment the pathological changes of patients with T2DM (43). Regarding host inflammation parameters, the recent-onset T2DM groups did not differ significantly from each other, but most cytokines were slightly elevated in the DM-LS group. As the metabolic and immune systems are commonly thought to be highly integrated and interdependent (34), future deviations in immune response and metabolic regulation between DM-DS and DM-LS are worth noting in follow-up research of disease trajectories.

DM-LS also exhibited more gut dysbiosis than DM-DS, despite some common altered characteristics in the two groups, including an overall shifted composition of the microbial community as well as individual taxa abundances, especially the predominant genus Bacteroides. DM-LS uniquely showed a significant reduction in species richness and bacterial co-occurrences, and a noticeable shift in functional composition deviated from HC. We further identified 10 differentially abundant individual taxa between the stool-type–stratified T2DM subgroups, with more abundant Blautia, Ruminococcus, Faecalibacterium, and Bifidobacterium in DM-DS. Faecalibacterium and Ruminococcus produce SCFAs known for their protective effects on glucose-lipid homeostasis and IS in both human cohorts and animal models (13,33). Bifidobacterium improves glucose tolerance and has been successfully used in probiotic formulations for T2DM (12). Notably, the commonly reported decrease of Faecalibacterium and Bifidobacterium in patients with T2DM versus healthy people (12) was only observed in DM-LS, while the inverse alterations were found in DM-DS, revealing discordant behaviors of bacteria in different stool microenvironments and emphasizing the crucial importance of stool-type assessment, especially in gut metagenome-wide association studies investigating diseases.

DM-DS and DM-LS were also discrepant with regard to the concentrations of two important types of microbial metabolites: SCFAs and secondary BAs. Functional gene analysis revealed that the microbiota contributed differently to the related metabolic pathways in the two groups. Also, DM-DS and DM-LS had largely different concentrations of these metabolites. More importantly, we reached opposite between-group comparison results using stool and serum samples. We found higher fecal excretion density in DM-LS but advantageous serum circulation concentrations in DM-DS. This apparent contrast can be explained by different absorption and carbohydrate availability linked to CTT (45). Previous studies on CTT and fermentation products revealed that long CTT, which is more common in individuals with dry stools, lowers the availability of fermentable polysaccharides and increases the absorption rate (45). SCFAs and secondary BAs are produced in the cecum and the proximal colon but used and metabolized by various tissues, with only 5–10% excreted in the feces (46,47). Therefore, CTT and its proxy stool consistency are confounding factors that will lead to contradictory results in determining the discriminative power of SCFAs and secondary BAs as measured in stool and serum as biomarkers in both health and disease.

With ML methods, we were able to evaluate the significance of stool type in T2DM and learn the host and microbial patterns and biomarkers across our stool-type–stratified patients. These data-based models place high demands on the quality of multiomics data, and our high-quality paired multiomics data thus provide an opportunity to investigate the complex host-microbe interactions from different angles, yet more validation is needed before results can be generalized to a larger population. We observed divergent host-microbe interactions in DM-DS and DM-LS in the cross-measurement association network. In DM-DS, Blautia, the most crucial microbe indicator of stool type, exhibited significant associations with HOMA2-β, fasting insulin level, and glucagon. Dorea and Faecalibacterium also linked some T2DM-related parameters and cooperated with Blautia, hinting at the combined effects of the three bacteria. By contrast, no clear associations with T2DM were observed in DM-LS. In addition to the interpretation of the benefits of Faecalibacterium above, Dorea is potentially inversely linked to T2DM via its positive regulation of secondary BAs reported in previous studies (44). Encouragingly, our microbial intervention experiments in the obese T2DM mouse model pointed out the inhibition capacity of Blautia on weight gain, lipid accumulation, and hyperglycemia and the modulation of gut microbial composition, confirming the general picture arising from the host-microbe associations inferred from multiomics data. Although different metagenome studies reported inconsistent alterations of Blautia in T2DM, our mouse experiments confirmed its probiotic potential. Our results are in accordance with those of Hosomi et al. (14), the only other animal study investigating the effects of Blautia. Therefore, we hypothesize that the enriched beneficial bacteria, through cooperation between microbes and interactions with the host, might contribute to the observed differences in metabolic disorders between T2DM subtypes. We expect the inhibitory effects of Blautia on lipid accumulation, weight gain, and blood glucose elevation in T2DM to be largely similar for both patient groups. However, because of the different stool microenvironments, Blautia will face different competitive forces, possibly influencing its colonization and functioning within the gut. Using more advanced experimental tools, further follow-up research investigating the potential of this bacterium should include stool type, alongside others, in order to reveal whether Blautia supplementation is equally beneficial for all patients or stratification for this, or other, parameters might be advantageous.

Next to supplementation with Blautia, we identified potential alternative microbial interventions favorable for T2DM based on cross-measurement correlation analysis within the HC group (Supplementary Fig. 8). We found clues to a preventive effect of secondary BAs and Collinsella on T2DM because of their notable associations with fasting insulin level or HOMA2 indexes in the HC group. Secondary BAs regulated human immune response and improved insulin sensitivity in earlier reports (33). In addition, a previous microbiome study revealed a positive link between Collinsella and circulating insulin (48). Based on cumulating evidence, both are thus promising candidates for follow-up research.

The host-microbe associations identified in this study stress the importance of the bacterial component in T2DM. However, these direct one-to-one relationships are far from accurately depicting the gut environment, which is shaped as an interdependent and multilevel network involving cascading links between many bacteria and host parameters (49). To disclose more bacterial regulations, we detected the shortest pathways from microbes and their metabolites to T2DM-related clinical parameters, implying the possible ways by which microbes can modulate the disease. For example, in the HC group, we identified phospholipids as the intermediate node connecting Blautia with HOMA2-β. This indirect host-microbe association between Blautia and HOMA2-β, in line with their direct link recognized in the T2DM population, reveals the benefits of Blautia in both HCs and patients and suggests that phospholipids are likely to be involved in the beneficial regulation. All the direct and indirect host-microbe associations identified here are collected in our online database DMA to stimulate future research and increase our insights into the complex system of host-microbiota interactions in T2DM.

Our study represents the first endeavor into microbial precision intervention for T2DM. It showcases the great potential of stool-type–based stratification, improving the identification of individuals differing in the severity and presentation of the disorder. We collected multiomics data from patients with new-onset T2DM who were screened from a large population using stringent inclusion criteria. The rich data set allowed us to identify potential bacterial modulators of T2DM, which we further verified by a mouse intervention experiment. However, several limitations have to be acknowledged as well. First, the confined region of the enrolled patients, who are likely to have similar dietary structures and life habits, might cause sample bias in DM-DS and DM-LS and restrict the generalization of the results. Future studies of multiple cohorts from different ethnic populations are necessary to verify the higher prevalence of loose stools found in both our cohorts and an Australian population (10) and to confirm our research findings. Second, although our cross-sectional design regarding the omics layers revealed many associations, it is not possible to disentangle cause and effect. Third, we recognize that our study design could not assess the longitudinal stability of the associations. However, we did observe the stability of microbe-microbe associations in longitudinal samples from patients with prediabetes in the Human Microbiome Project (Supplementary Fig. 9). Despite these preliminary analyses, more time series cross-sectional data and in vitro experiments are warranted to dissect the more defined mechanisms underlying the microbial effects associated with different stool types. Furthermore, we used stool type as a single parameter to stratify patients with T2DM. In doing so, we provide a thorough description of its importance and effect, yet our approach does not provide a complete stratification system. This would require an assessment of the stratification power of stool type in combination with other parameters. While our data set was sufficiently powered for the former, the latter sort of analysis requires a larger sample size with sufficient variation in all investigated parameters.

To advance our understanding of the clinical implications of stool-type–based stratification, future prospective studies on long-term complications and outcomes should be performed to reveal the progression trajectories of DM-DS and DM-LS. The efficacy of the traditional intervention, including drug administration and dietary regimes, and its combination with probiotics therapy also deserve to be looked at through the lens of stool-type–based subclassification. Additionally, the consistency-based binary stool types (dry/loose stools) probably need to be refined further to a continuous variable (e.g., Bristol Stool Scale, moisture content) and/or by including other stool characteristics (e.g., pH). Moreover, further follow-up research on bacterial intervention in T2DM should consider the stool type, alongside other parameters, to investigate the most effective therapy strategies for individuals with T2DM. With future replication, verification, and improvement, the bacteria and stratification parameters as proposed here offer a potential framework to explore probiotic precision interventions in T2DM.

Acknowledgments. The multiomics data process and analysis were performed on the High Performance Computing Platform of the Center for Life Sciences of Peking University.

Funding. This work was supported by the National Key Research and Development Program of China (2021YFC2300300), the Innovation Team and Talents Cultivation Program of National Administration of Traditional Chinese Medicine (ZYYCXTD-D-202001), and the National Natural Science Foundation of China (81430097, 81973837, 32070667, 31671366, and 32300078).

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. Q.G., Z.G., H.W., and D.V. codesigned the study based on the sequenced data. Q.G., Z.G., L.Z., D.V., and H.Z. wrote and revised the manuscript. Q.G., M.L., J.H., and X.J analyzed the data and made the plots and tables of the results. Z.G., L.Z., H.W., Z.L., L.H., S.D., and Y.L. codesigned the experiments and sampling. L.Z., H.Z., and X.T. cosupervised the study. All authors proofread and improved the manuscript. H.Z. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.