Exploring the gut microbiome in type 2 diabetes across different insulin resistance levels: a machine learning approach

TL;DR

This study explores how gut bacteria are linked to insulin resistance in type 2 diabetes and identifies specific microbes that could help improve metabolic health.

Contribution

The study introduces a machine learning approach to identify gut microbiome features associated with insulin resistance in type 2 diabetes.

Findings

XGBoost models using gut microbiome profiles effectively distinguished individuals with higher insulin resistance in T2DM.

Bacteroides and Faecalibacterium were key contributors to model performance and correlated with clinical measures of insulin resistance.

Several gut microbes were associated with IR indices, suggesting potential targets for metabolic interventions.

Abstract

Insulin resistance (IR) is central to type 2 diabetes mellitus (T2DM). Composite indices including the atherogenic index of plasma (AIP), metabolic score for insulin resistance (METS-IR), triglyceride–glucose index (TyG), and TyG-BMI, are widely used to quantify IR severity. The gut microbiome (GM) has been implicated in metabolic dysregulation, but its associations with IR remain incompletely defined.

We collected blood test results and stool samples from participants with T2DM and healthy controls. Stool samples underwent 16S rRNA gene sequencing. We trained XGBoost models to distinguish individuals with higher IR from healthy controls based on GM profiles and performed correlation analyses between GM features, clinical measures, and IR indices.

Triglycerides (TG), fasting blood glucose (FBG), and high-density lipoprotein cholesterol (HDL-C) differed significantly between the T2DM and control groups. IR indices (AIP, METS-IR, TyG, and TyG-BMI) were markedly higher in the T2DM group. XGBoost models based on GM profiles showed high discriminatory performance for identifying T2DM individuals with higher IR, with Bacteroides and Faecalibacterium contributing most to model performance. Correlation analyses further indicated that Lachnospiraceae_UCG-010, Bacteroides, Faecalibacterium, Lachnospira, Parasutterella, and Escherichia–Shigella were associated with clinical measures and IR indices.

Specific GM features are associated with IR-related clinical measures and composite indices in T2DM, supporting their potential as intervention targets to improve insulin resistance and restore carbohydrate and lipid metabolism.

Figures

Graphical abstract. Machine-learning-based identification of gut microbiome taxa associated with insulin resistance (IR) in type 2 diabetes. A cohort of healthy controls and T2DM participants provided stool samples for gut microbiome profiling. In parallel, clinical measurements including BMI, TG, FBG, and HDL-C were collected and used to derive IR indices including AIP, METS-IR, TyG, and TyG-BMI, to capture complementary dimensions of glucose-lipid dysregulation. Microbiome features were then selected using a machine learning framework, yielding a set of IR-related microbial features. Finally, correlation analyses integrated the selected taxa with clinical and IR indices to nominate microbial features potentially relevant to IR modulation.

Group comparisons of clinical measures between healthy controls and individuals with T2DM. Box-and-jitter plots show BMI, TG, FBG and HDL-C in the control group (blue) and the T2DM group (pink). Each dot represents a participant. Group differences are indicated above each panel: ns denotes no significant difference, indicates p < 0.0001.

Group comparisons of insulin resistance (IR) indices between healthy controls and individuals with T2DM. Box-and-jitter plots show AIP, METS-IR, TyG and TyG-BMI in the control group (blue) and the T2DM group (pink). Each dot represents a participant. Group differences are indicated above each panel: indicates p < 0.0001.

Discriminative performance of gut microbiome (GM)-based classifiers across four insulin resistance (IR) indices. Bar plots show the area under the receiver operating curve (AUC) for XGBoost classifiers constructed using GM features for AIP, METS-IR, TyG, and TyG-BMI. For each index, participants were stratified into higher and lower groups according to the index value, and classifier performance is reported for three pairwise comparisons: higher vs. control (pink), lower vs. control (light blue), and higher vs. lower (turquoise).

Key gut microbiome (GM) taxa contributing to classification of the higher index T2DM subgroup versus healthy controls across four insulin resistance (IR) indices. Bar plots show the feature-importance scores for the higher vs. control comparison for each index (AIP, METS-IR, TyG, and TyG-BMI). Within each panel, genera are ranked by importance (highest to lowest). The pie charts summarize the mean relative abundances of representative discriminatory genera in controls and T2DM, including Escherichia-Shigella, Bacteroides, Faecalibacterium, Lachnospira, Parasutterella, Lachnospiraceae_UCG-010, and the remaining genera are grouped as others.

Spearman correlation heatmap between selected GM features and clinical/IR indices. Rows indicate GM features selected by the machine learning models, and columns indicate clinical/IR measures. Each cell shows the Spearman correlation coefficient. Colors represent correlation direction and magnitude, with purple indicating negative correlations and yellow indicating positive correlations.

Tables

Table 1. Demographic summary of the control and T2DM groups.

Characteristics	Control	T2DM	p-value
Male	20	55	0.09
Female	18	23
Age (years)	51.71 ± 6.66	52.90 ± 11.13	0.48

Full text

Introduction

1

Type 2 diabetes mellitus (T2DM) is a chronic metabolic condition characterized by dysregulation of carbohydrate metabolism, imposing an increasing burden to the global population (1). According to the International Diabetes Federation (IDF), approximately 536.6 million individuals aged 20–79 years were living with diabetes worldwide in 2021, and this number is expected to rise to 783.2 million by 2045 (2). Among the various types of diabetes, T2DM is the predominant form, with insulin resistance (IR) as its core pathophysiological feature.

IR refers to a condition in which insulin-sensitive tissues, especially the liver, muscle, and adipose tissue, are unable to effectively lower blood glucose levels in response to normal insulin concentrations (3). Although the level of IR cannot be measured directly in clinical practice, several indices have been developed to estimate IR status from different physiological perspectives. The atherogenic index of plasma (AIP), calculated from triglyceride (TG) and high-density lipoprotein cholesterol (HDL-C), reflects the balance between adverse and protective lipoproteins (4). Notably, a non-linear relationship between AIP and IR has been reported, with a positive correlation below an inflection point of 0.45 (5). The metabolic score for insulin resistance (METS-IR) is another index for quantifying IR, positively associated with visceral fat accumulation and fasting insulin levels, and has demonstrated superior performance in evaluating insulin sensitivity (6). The triglyceride-glucose index (TyG), derived from fasting blood glucose and triglycerides, has shown high sensitivity and specificity for diagnosing IR when benchmarked against euglycemic-hyperinsulinemic clamp, the gold standard for IR assessment (6). An extended version, TyG-BMI, incorporates body mass index (BMI) into the TyG index and exhibits a strong correlation with IR (7). Collectively, these indices enable accessible and cost-effective evaluation of IR, which is central to the development and progression of T2DM as well as its chronic complications. Consequently, strategies that ameliorate IR are critical for improving glycemic control and preventing T2DM-related morbidity.

Growing evidence indicates that the gut microbiome (GM) plays a significant role in IR. Individuals with T2DM exhibit distinct alterations in microbial composition compared with healthy controls (8). The GM contributes to host metabolism by extracting energy from dietary carbohydrates and generating bioactive metabolites (9). Experimental studies have shown that the GM dysbiosis can alter the production of microbial metabolites thereby influencing insulin sensitivity in the liver, adipose tissue, and muscle (10). Human observational studies further support associations between specific taxa, such as Firmicutes and Bacteroidetes, and IR (11). These findings highlight the potential of targeting specific microbial taxa and their metabolites as a therapeutic strategy for improving IR.

However, GM data are inherently high-dimensional and highly sparse, posing challenges for traditional statistical methods. Machine learning (ML) methods, by contrast, excel at handling such data structures, enabling the detection of complex, non-linear patterns within large microbiome datasets (12). ML also facilitates feature selection, offering a powerful framework for GM-based disease classification (13).

Therefore, we aimed to characterize the GM profiles of T2DM individuals across different IR levels, thereby refining the association between GM and IR. Specifically, we leveraged multiple IR indices to capture complementary dimensions of dysregulated glucose and lipid metabolism, and implemented a XGBoost machine learning framework to select GM features. By further correlating selected GM features with clinical and IR measures, we strengthened the biological understanding of IR-GM relationships and identified GM features with potential relevance for IR modulation. The detailed graphical abstract is shown in Figure 1.

Graphical abstract. Machine-learning-based identification of gut microbiome taxa associated with insulin resistance (IR) in type 2 diabetes. A cohort of healthy controls and T2DM participants provided stool samples for gut microbiome profiling. In parallel, clinical measurements including BMI, TG, FBG, and HDL-C were collected and used to derive IR indices including AIP, METS-IR, TyG, and TyG-BMI, to capture complementary dimensions of glucose-lipid dysregulation. Microbiome features were then selected using a machine learning framework, yielding a set of IR-related microbial features. Finally, correlation analyses integrated the selected taxa with clinical and IR indices to nominate microbial features potentially relevant to IR modulation.

Methods

2

Healthy controls were individuals undergoing routine laboratory tests at the Health Examination Center of Hospital of Chengdu University of Traditional Chinese Medicine. Patients with T2DM were recruited from the Inpatient Department of Endocrinology at the same hospital. The study received approval from the Medical Ethics Committee of Hospital of Chengdu University of Traditional Chinese Medicine (approval number 2020KL-060)1 and was conducted in strict adherence to the principles outlined in the Declaration of Helsinki. Informed consent was duly acquired from all participants.

Inclusion and exclusion criteria

2.1

Participants eligible for this study were individuals aged between 18 and 75 years. Enrollment included healthy subjects free from any ongoing diseases, as well as patients with a confirmed diagnosis of T2DM. These participants were allocated into respective groups based on their health status. The exclusion criteria were delineated as follows: (1) Individuals who have experienced acute metabolic disturbances, such as ketoacidosis, within the preceding month. (2) Presence of severe infectious diseases. (3) Concurrent liver pathologies including infections by hepatitis B or C virus, autoimmune hepatitis, primary biliary cholangitis, primary sclerosing cholangitis, drug-induced steatosis or liver injury, etc. (4) Diagnoses of gastrointestinal or hepatic malignancies. (5) History of alcoholism or substance misuse. (6) Use of antibiotics or ursodeoxycholic acid at the point of study entry. (7) Use of prebiotics or probiotics at the point of study entry. (8) Periods of pregnancy and lactation.

Demographic statistics

2.2

A cohort of 116 participants from Chengdu city, China, was recruited and categorized into two groups based on their health conditions: healthy controls (n = 38) and patients diagnosed with T2DM (n = 78). A Pearson chi-square test was employed to assess differences in gender distribution between the groups. Additionally, t-tests were utilized to evaluate disparities in age between the groups. These analyses revealed no significant differences in gender distribution or age between the groups (Table 1).

Gut microbiome-16S rRNA gene sequencing

2.3

Stool specimens were collected from healthy controls and T2DM patients to analyze gut microbiome profiles, which were elucidated through 16S rRNA gene sequencing. Initially, DNA was isolated from fecal samples using DNeasy PowerSoil kit, followed the assessment of amplicon quality via gel electrophoresis. The generation of the amplicon sequence variant (ASV) abundance table was performed using DADA2, employing the default parameters of set forth in QIIME2. Subsequently, all representative sequences were annotated and subjected to BLAST analysis against the Silva database (Version 138) using the q2-feature-classifier (14). The sequencing data have been deposited in the NCBI Sequence Read Archive (SRA)2 under the BioProject accession PRJNA1397912.

Blood test outcomes

2.4

After obtaining informed consent, body mass index (BMI) along with clinical measures of triglycerides (TG), high-density lipoprotein cholesterol (HDL-C), and fasting blood glucose (FBG) were extracted from the medical records of each participant. We used independent two-sample t-tests to compare these variables between groups. Based on these measurements, the AIP, METS-IR, TyG, and TyG-BMI were calculated for each participant using established formulas (4, 6, 7).

To examine differences across IR levels, we stratified the T2DM cohort into high and low subgroups for each index using the median as the cut-off. For AIP, the median in the T2DM group was 0.32. Participants with AIP >0.32 were assigned to the higher AIP subgroup, and those with AIP ≤0.32 to the lower AIP subgroup. The same approach was applied to other indices, with medians of 54.46 for METS-IR, 7.76 for TyG, and 199.59 for TyG-BMI.

Machine learning

2.5

We used subject-level leave-one-out cross-validation (LOOCV) to maximize the number of samples available for model training. In each outer LOOCV iteration, one subject was held out for testing, and the remaining subjects formed the training set. All preprocessing was performed in a leakage-free manner. Missing values in the GM matrix were imputed using a k-nearest neighbor (KNN) imputer fitted on the training set.

Classification was performed using an extreme gradient boosting model (XGBoost), selected for its ability to handle high-dimensional, sparse feature spaces and nonlinear feature interactions, and for its built-in L1/L2 regularization that helps reduce overfitting in small-sample settings (15). Hyperparameters were optimized within the training set using an inner cross-validation loop, and the tuned model was subsequently evaluated on the held-out subject in the outer loop. Overall model performance was computed by aggregating predictions across all LOOCV iterations.

To quantify the contribution of each feature, permutation feature importance was estimated in a cross-validation fashion by permuting one feature at a time in the held-out data and measuring the associated decrease in the chosen performance metric relative to the unpermuted baseline.

For each IR index, we constructed three separate classification tasks on the GM data: higher-index subgroup vs. controls, lower-index subgroup vs. controls, and higher vs. lower subgroups, training an XGBoost model for each comparison.

Correlation analysis

2.6

Correlation analyses were performed to evaluate associations among clinical measurements, IR indices and GM features selected by machine learning models. Shapiro–Wilk tests were employed to assess normality. Pearson’s correlation was used for approximately normally distributed variables with linear relationships. Otherwise, Spearman’s rank correlation was applied. All tests were two-tailed, and statistical significance was evaluated at a 95% confidence level. To account for multiple comparisons, p-values were adjusted using the false discovery rate (FDR) controlled by the two-stage step-up method of Benjamini, Krieger and Yekutieli, with adjusted q value <0.01 considered significant.

Results

3

Comparison of BMI, TG, FBG, and HDL-C between groups

3.1

As shown in Figure 2, there was no significant difference in BMI between the T2DM and control groups (mean BMI: 25.48 vs. 24.84 kg/m^2^, p > 0.05). In contrast, TG, FBG, and HDL-C differed markedly between groups. Compared with controls, patients with T2DM exhibited significantly higher TG (2.60 vs. 1.49 mmol/L, p < 0.0001) and FBG (8.32 vs. 5.10 mmol/L, p < 0.0001), while HDL-C levels were significantly lower (0.91 vs. 1.27 mmol/L, p < 0.0001). These findings indicate that dyslipidemia and hyperglycemia, but not elevated BMI, are key distinguishing features between the groups.

Group comparisons of clinical measures between healthy controls and individuals with T2DM. Box-and-jitter plots show BMI, TG, FBG and HDL-C in the control group (blue) and the T2DM group (pink). Each dot represents a participant. Group differences are indicated above each panel: ns denotes no significant difference, **** indicates p < 0.0001.

Comparison of AIP, METS-IR, TyG, and TyG-BMI between groups

3.2

Indices reflecting IR were consistently higher in participants with T2DM compared with healthy controls (Figure 3). The AIP was markedly elevated in the T2DM group (mean AIP 0.35 vs. 0.05 in controls), and the difference was highly significant (p < 0.0001). Similar patterns were seen for the METS-IR (53.89 vs. 43.04), the TyG (7.80 vs. 7.05), and the TyG-BMI (198.77 vs. 175.4), all group comparisons reached p < 0.0001. These findings reflect more pronounced metabolic dysregulation in T2DM group compared with healthy individuals.

Group comparisons of insulin resistance (IR) indices between healthy controls and individuals with T2DM. Box-and-jitter plots show AIP, METS-IR, TyG and TyG-BMI in the control group (blue) and the T2DM group (pink). Each dot represents a participant. Group differences are indicated above each panel: **** indicates p < 0.0001.

Classification performance on GM data

3.3

The XGBoost classifiers trained on GM data distinguished participants with higher IR index values from healthy controls much more effectively than other comparisons (Figure 4). For AIP, METS-IR, TyG, and TyG-BMI, the area under the ROC curve (AUC) for the higher-index vs. control models were 0.69, 0.84, 0.77 and 0.81, respectively. METS-IR achieved the strongest discrimination and AIP the weakest among these. When discriminating between higher-index and lower-index subgroups, or between lower-index and control groups, performance declined substantially. AUCs for higher vs. lower comparisons ranged from 0.54 to 0.71, and for lower vs. control comparisons from 0.43 to 0.61. Overall, these results indicate that GM signatures robustly differentiate individuals with pronounced IR from healthy controls than they separate those with less severe metabolic profiles.

Discriminative performance of gut microbiome (GM)-based classifiers across four insulin resistance (IR) indices. Bar plots show the area under the receiver operating curve (AUC) for XGBoost classifiers constructed using GM features for AIP, METS-IR, TyG, and TyG-BMI. For each index, participants were stratified into higher and lower groups according to the index value, and classifier performance is reported for three pairwise comparisons: higher vs. control (pink), lower vs. control (light blue), and higher vs. lower (turquoise).

GM feature selection based on relative contribution

3.4

Figure 5 depicts the taxa that contributed most to XGBoost models distinguishing individuals with T2DM and high IR indices from controls. Each index was driven by a distinct set of genera. In the AIP model, Terrisporobacter showed the greatest discriminative power, with other notable contributions from Lachnospiraceae_UCG-010, Anaerococcus, and Romboutsia, suggesting a broad involvement of Firmicutes. The METS-IR was dominated by Lachnospiraceae_UCG-010, followed by Oscillospira and Butyricimonas. For TyG, Bacteroides and Faecalibacterium were the most influential genera. Finally, the TyG-BMI classifier was characterized by a leading contribution from Lachnospira, accompanied by pathobionts including Escherichia-Shigella and Parasutterella. Consistent with these feature-importance patterns, Escherichia-Shigella was markedly enriched in the T2DM group compared with controls (8.48% vs. 1.97%). In contrast, the relative abundances of Bacteroides, Faecalibacterium, Lachnospira, Parasutterella, Lachnospiraceae_UCG-010 were substantially reduced in T2DM group (25.33% vs. 9.39, 12.61% vs. 5.51, 2.49% vs. 0.19, 1.28% vs. 0.66, 0.18% vs. 0.05%, respectively).

Key gut microbiome (GM) taxa contributing to classification of the higher index T2DM subgroup versus healthy controls across four insulin resistance (IR) indices. Bar plots show the feature-importance scores for the higher vs. control comparison for each index (AIP, METS-IR, TyG, and TyG-BMI). Within each panel, genera are ranked by importance (highest to lowest). The pie charts summarize the mean relative abundances of representative discriminatory genera in controls and T2DM, including Escherichia-Shigella, Bacteroides, Faecalibacterium, Lachnospira, Parasutterella, Lachnospiraceae_UCG-010, and the remaining genera are grouped as others.

Correlations among clinical measures, IR indices and GM features

3.5

Normality testing indicated that BMI, METS-IR, and TyG-BMI were approximately normally distributed, whereas all other GM features, IR indices, and clinical measures deviated from normality (Supplementary Table 1). Given the non-normal distribution of GM features, we therefore used Spearman’s rank correlation to examine associations between the selected GM features and clinical/IR measures, with FDR correction applied for multiple comparisons (Supplementary Table 2).

After FDR adjustment, Lachnospiraceae_UCG-010 was not significantly correlated with BMI, but showed negative correlations with TG, FBG, AIP, METS-IR, TyG and TyG-BMI, and a positive correlation with HDL-C. bacteroides was not significantly correlated with BMI or TG after FDR adjustment, but was negatively correlated with FBG, AIP, METS-IR, TyG, and TyG-BMI, and positively correlated with HDL-C. Faecalibacterium and Lachnospira were not significantly correlated with BMI after FDR adjustment, while showing negative correlations with TG, FBG, AIP, METS-IR, TyG and TyG-BMI, and positive correlations with HDL-C. For Escherichia-Shigella, correlations with BMI, TG, AIP, METS-IR, TyG and TyG-BMI were not significant after FDR adjustment. It showed a negative correlation with HDL-C and a positive correlation with FBG. For Parasutterella, correlations of BMI, TG, and TyG-BMI were not significant after FDR adjustment, whereas it showed negative correlations with FBG, AIP, METS-IR, and TyG, and a positive correlation with HDL-C (see Figure 6).

Spearman correlation heatmap between selected GM features and clinical/IR indices. Rows indicate GM features selected by the machine learning models, and columns indicate clinical/IR measures. Each cell shows the Spearman correlation coefficient. Colors represent correlation direction and magnitude, with purple indicating negative correlations and yellow indicating positive correlations.

Discussion

4

In this study, TG, FBG, and HDL-C levels differed significantly between participants with T2DM and healthy controls. Indices of IR were also markedly higher in the T2DM group. Machine learning analysis of GM profiles demonstrated high performance in distinguishing T2DM individuals with high IR indices from healthy controls, with GM features including Bacteroides and Faecalibacterium contributing most to the XGBoost models. Correlation analyses further revealed links among clinical measures, IR indices and GM features.

Beyond elevated FBG, our results also showed that participants with T2DM exhibited abnormal lipid profiles, underscoring that the disease reflects dysregulation of both carbohydrate and lipid metabolism. IR is the core mechanism bridging these disturbances by reshaping how the liver, skeletal muscle, and adipose tissue process glucose and lipids.

In the insulin-resistant liver, gluconeogenesis remains inappropriately active and glycogen synthesis is relatively suppressed, resulting in an inadequate suppression of hepatic glucose output. Notably, this defect can coexist with preserved insulin signaling along anabolic branches. In particular, insulin upregulates SREBP1 in an mTORC1-dependent manner, and mTORC1 further enhances SREBP1 function through post-translational regulation, collectively augmenting hepatic cholesterol and fatty-acid biosynthesis and thereby promoting steatosis and hyperglyceridemia (16). Moreover, selective activation of the Akt*-mTORC1-*SREBP-1c axis favors intrahepatic accumulation of TG and diacylglycerols (DAGs). DAG-driven protein kinase C activation, in turn, impairs insulin receptor signaling and exacerbating hepatic IR (17).

In skeletal muscle, reduced insulin-stimulated glucose uptake limits peripheral glucose disposal, thereby increasing substrate delivery to the liver and favoring its conversion into fatty acids via de novo lipogenesis (18, 19). In adipose tissue, loss of insulin’s antilipolytic effect enhances adipocyte lipolysis, increasing circulating free fatty acids (FFAs) that deposit ectopically (20). Excessive lipolysis also promotes secretion of pro-inflammatory cytokines, such as IL-6 and TNF-α, which further impair insulin signaling and aggravate systemic insulin resistance (18).

Across these tissues, chronic nutrient excess can maintain elevated mTORC1 level, which not only promotes anabolism but also suppresses catabolic quality-control programs. Mechanistically, mTORC1 phosphorylates transcription factor EB (TFEB), retaining it in the cytoplasm and downregulating lysosomal biogenesis and autophagy gene programs. Conversely, mTORC1 inhibition activates TFEB and enhances lysosomal-autophagic processes, support proteostasis and mitochondrial clearance (16). Consistently, dietary restriction associated signaling includes mTOR inhibition along with AMPK, SIRT1, and PGC-1α, facilitating mitophagy and reducing inflammatory cytokines output, which in turn restores mitochondrial quality and improves metabolic resilience (21). However, it is important to note that overly reduced mTORC1 activity in pancreatic β cells may compromise insulin production and β-cell maintenance (22).

Collectively, these mechanisms highlight IR as the driver of both hyperglycemia and dyslipidemia in T2DM.

Consistently, IR indices that integrate glucose- and lipid-related measures, including AIP, METS-IR, TyG, and TyG-BMI, were significantly higher in the T2DM group. Prior studies show that higher METS-IR values often accompanied by elevated TG and reduced HDL-C, reflecting increased risk of obesity, fatty liver disease, and atherosclerosis (23–25). AIP, which captures TG/HDL-C imbalance, serves as an indicator of aggravated lipid abnormalities and heightened cardiovascular risk (26). The TyG index is a widely validated surrogate of IR, with higher values denoting impaired insulin sensitivity and disruption of glucose-lipid homeostasis (27). TyG-BMI, which incorporates adiposity, often outperforms TyG alone in detecting IR, particularly in overweight or obese populations, thereby improving early risk stratification and metabolic phenotyping in T2DM (28). Taken together, interventions targeting IR are essential not only for controlling glycemia but also for mitigating dyslipidemia and preventing cardiometabolic complications in T2DM.

Compared with medications such as metformin and GLP-1 agonists, which are effective in improving IR, GM interventions offer distinct and complementary advantages. By directly modulating host–microbe interactions, GM-targeted strategies can restore microbial balance, increase the abundance of short-chain fatty acid (SCFA)-producing bacteria, and thereby enhance endogenous GLP-1 secretion (29). These microbes not only mimic incretin-based pharmacotherapy but also reinforce the intestinal barrier and attenuate chronic low-grade inflammation (30). Clinical studies further indicate that combining probiotics with metformin reduces gastrointestinal side effects and yields greater reductions in HbA1c compared with metformin alone (31). Moreover, unlike conventional drugs that mainly target specific metabolic pathways, GM interventions adopt a multifaceted approach, simultaneously improving glucose homeostasis, lipid metabolism, and systemic inflammatory status.

In this study, we identified GM features including Lachnospiraceae_UCG-010, Bacteroides, Faecalibacterium, Lachnospira, Parasutterella, and Escherichia-Shigella that distinguished T2DM participants with high IR indices from healthy controls and were associated with clinical measures and IR indices.

Lachnospiraceae_UCG-010, a member of the Lachnospiraceae family, is a major short-chain-fatty-acid (SCFA) producer. Butyrate, in particular, modulates inflammation and stimulates glucagon-like peptide-1 (GLP-1) secretion, thereby improving glucose uptake and insulin sensitivity. Reports also show that Lachnospiraceae_UCG-010 is correlated with oxidative stress and lipid-metabolism indices in metabolic disorders, suggesting that its abundance reflects lipid dysregulation (32). Overall, targeting Lachnospiraceae_UCG-010 likely to support carbohydrate and lipid homeostasis through SCFA production in T2DM individuals.

Faecalibacterium, particularly F. prausnitzii, is another butyrate producer. Animal studies demonstrate that supplementation with F. prausnitzii significantly lowers FFA, TG, total cholesterol and the atherogenic index, while also reducing hepatic TG and cholesterol content (33). Mechanistically, this bacterium suppresses hepatic lipogenic enzymes, such as fatty-acid synthase and β-oxidation activities, and decreases pro-inflammatory cytokines including TNF-α, IL-6 and IFN-γ. F. prausnitzii is considered as an indicator of gut dysbiosis as its abundance is diminished in T2DM individuals (33). Restoring F. prausnitzii may improve insulin sensitivity and lipid metabolism through butyrate-mediated anti-inflammatory effects.

Bacteroides species produce diverse metabolites with opposing metabolic effects. On one hand, aromatic-amino-acid metabolites such as phenylacetic acid promote hepatic TG accumulation and impair insulin signaling by reducing AKT phosphorylation in hepatocytes (10). On the other, Bacteroides are major SCFA producers, generating acetate, propionate and butyrate (10). These SCFAs suppress hepatic lipogenesis by inhibiting lipogenic gene expression, enhance lipid oxidation via AMPK activation, increase energy expenditure, and promote adipose browning. In diabetic mice, supplementation with Bacteroides uniformis improved carbohydrate and lipid metabolism, lowered TG and LDL-C, down-regulated hepatic gluconeogenic and lipogenic genes, and restored bile acid signaling through the TGR5/AMPK pathway (34).

Lachnospira spp. produce SCFAs and contribute to flavonoid biosynthesis. Their abundance was positively correlated with pathways for flavone/flavanol biosynthesis and negatively associated with insulin and HOMA-IR (35). Studies have suggested that lower Lachnospira may exacerbate IR and inflammation, suggesting a protective metabolic role. Lower Lachnospira abundance has been associated with increased IR and inflammation, implying that this genus supports insulin sensitivity by providing butyrate and flavonoid metabolites.

In contrast, Parasutterella has been linked to adverse metabolic profiles. High abundance Parasutterella correlates with obesity and T2DM, greater carbohydrate intake, reduced serum L-cysteine (an amino acid essential for glucose homeostasis), and activation of host fatty-acid biosynthesis pathways (36). It has also been associated with hypothalamic inflammation, potentially contributing to weight gain and IR through. Importantly, weight-loss interventions have been shown to reduce Parasutterella abundance, suggesting its role as a modifiable marker of metabolic dysfunction.

Similarly, Escherichia-Shigella, a Gram-negative Enterobacteriaceae group, is consistently enriched in diabetic models. Its abundance has been positively correlated with FBG, HOMA-IR and circulating lipopolysaccharide (LPS) levels (37). Elevated Escherichia-Shigella exacerbates metabolic endotoxemia by increasing LPS release, which triggers systemic inflammation and worsens IR. In db/db mice, quercetin supplementation reduced Escherichia-Shigella abundance, restored intestinal barrier integrity, decreased circulating LPS and improved insulin sensitivity. Intermittent fasting has similar effects, reducing Escherichia-Shigella, improving glucose tolerance, and enhancing insulin sensitivity (38).

Overall, supplementation with SCFA-producing genera such as Lachnospiraceae_UCG-010, B. uniformis, F. prausnitzii, Lachnospira may improve IR and metabolic homeostasis by enhancing insulin signaling, promoting lipid oxidation, and reducing inflammation. Conversely, interventions such as quercetin supplementation and lifestyle modifications (e.g., weight loss, intermittent fasting) may reduce harmful genera such as Parasutterella and Escherichia-Shigella, thereby alleviating hepatic TG accumulation and systemic endotoxin-mediated inflammation.

The primary limitation of this work is the modest sample size, which may constrain generalizability and increase the risk of overfitting in ML models. In addition, the cross-sectional design captures GM composition at a single time point and therefore cannot resolve temporal trajectories, directionality, or causality in the GM-IR relationship during T2DM onset and progression. Third, residual confounding cannot be fully excluded, factors such as adiposity distribution, smoking status, dietary patterns, and medication exposure may influence both IR and GM. Fourth, the use of 16S rRNA sequencing provides limited taxonomic resolution and does not directly measure microbial functional capacity, restricting pathway-level mechanistic inference.

Future work should therefore prioritize larger, multi-center cohorts with independent external validation to improve the robustness and transportability of the identified signatures. Longitudinal designs with repeated GM sampling and comprehensive metabolic phenotyping are also needed to assess within-person dynamics and strengthen causal inference. Moreover, future studies should incorporate richer covariate collection to better control confounding and evaluate potential effect modification. Finally, mechanistic insights would be further strengthened by integrating complementary omics layers, such as shotgun metagenomics, together with targeted experimental validation.

Conclusion

5

In summary, this study demonstrated that the gut microbiota profiles of T2DM participants with high IR levels were significantly different from those of healthy controls. Key genera, including Lachnospiraceae_UCG-010, Bacteroides, Faecalibacterium, Lachnospira, Parasutterella, and Escherichia-Shigella, emerged as potential targets for interventions aimed at improving insulin resistance and restoring carbohydrate and lipid metabolism. Importantly, incorporating microbiota-informed strategies could translate these findings into timelier and more precise T2DM management. For instance, routine microbiome profiling may help clinicians stratify patients by dysbiosis and IR risk, anticipate inter-individual variability in response to orally administered therapies, and enable earlier, personalized adjunct interventions with longitudinal monitoring to track therapeutic response and metabolic trajectory (39).