The human gut microbiota data for this study were obtained from the GMrepo database,¹ including individuals aged 18 years and older classified as either healthy or overweight. Overweight was defined as BMI ≥25, encompassing individuals with BMI ≥30 (commonly classified as obese). To ensure homogeneity, samples from individuals with other underlying diseases were excluded, and participants were categorized into two groups: overweight (BMI ≥25) and control (BMI <25). Stratified sampling was applied to the overweight group to minimize biases related to age and gender distribution, resulting in 3,606 samples analyzed (2,043 overweight and 1,563 control). The inclusion of individuals with obesity under the overweight category is discussed as a potential limitation. Data were imported into Python using pandas, enabling operations like filtering, alignment, and transformation. Missing values for bacterial relative abundance and gender were replaced with zeros to ensure data integrity. Z-score normalization was applied within cross-validation to avoid data leakage by calculating normalization parameters only on the training data and applying them to the test data. Samples with low sequencing coverage (<1%) were excluded, and potential contaminants were identified and removed through background sample comparisons, ensuring a robust and high-quality dataset.

We calculated the Shannon index and Chao1 index to assess the diversity of gut microbiota among the study groups. The Shannon index provides insights into the richness and evenness of microbial communities, while the Chao1 index estimates species richness based on the number of observed species and their frequency.

To visualize the microbial diversity and community structure, we generated heatmaps and boxplots. Heatmaps illustrate the relative abundance of microbial taxa across samples, highlighting patterns of similarity and differences between groups. Boxplots provide a comparative overview of diversity indices between the overweight and control groups, facilitating the identification of significant differences.

In this study, we applied z-score standardization to all input features to ensure that the features were on the same scale for comparison. The standardization was performed before model training to eliminate differences in scale between features, preventing certain features from disproportionately influencing the model results due to their different units. All numerical data (such as microbial community relative abundance and sample information) were standardized before being input into the machine learning models. This step helped improve the model’s stability and ensured that all features had equal weight during model training.

To investigate the potential of gut microbiota dysbiosis as biomarkers for overweight, we developed machine learning models using XGBoost, support vector machine (SVM), logistic regression, and decision tree algorithms. The optimization target for model performance was AUC (area under the ROC curve), though additional metrics such as accuracy, precision, recall, and F₁-score were also assessed. For each algorithm, we used the following settings: In XGBoost, hyperparameter optimization was performed using GridSearchCV, exploring the ranges of colsample_bytree: [0.8, 0.9, 1.0], gamma: [0, 0.1, 0.2], learning_rate (eta): [0.01, 0.1, 0.2], max_depth: [3, 4, 5], n_estimators: [50, 100, 150], and subsample: [0.8, 0.9, 1.0]. A 5-fold cross-validation strategy was employed to prevent overfitting and ensure robust model performance. For SVM (support vector machine), the radial basis function (RBF) kernel was used, and hyperparameters such as C (regularization parameter) and gamma (kernel coefficient) were optimized using a grid search. The values tested for C ranged from 1 to 100, and gamma ranged from 0.001 to 0.1. In Logistic Regression, default parameters were used, with the main hyperparameter being the regularization parameter C, which was optimized using GridSearchCV with values from 0.01 to 10. For the Decision Tree model, the default max_depth parameter was used, and different values for min_samples_split (ranging from 2 to 10) and min_samples_leaf (ranging from 1 to 4) were explored to control tree growth and avoid overfitting. After comparing performance metrics, including AUC, we determined that XGBoost was the most effective model for predicting overweight status based on gut microbiota data.

To interpret the XGBoost model’s predictions, we employed SHAP (Shapley Additive Explanations), an approach grounded in game theory (Li et al., 2023). SHAP quantifies the contribution of each feature to the model’s output by assigning a Shapley value that indicates the direction and magnitude of each feature’s impact. Positive SHAP values (>0) suggest a positive influence on the prediction, while negative SHAP values (<0) indicate a negative influence (Song et al., 2023). SHAP provides both global and local insights: Global Interpretation calculates the average SHAP value of each feature across all samples, highlighting overall feature importance, while local interpretation assesses each feature’s impact on individual predictions. The SHAP framework enabled us to generate “SHAP Summary Plots,” visualizing feature importance and contributing to a deeper understanding of the influence of gut microbiota features on overweight classification.

We cultured B. pseudocatenulatum strains (JXL-01, JXL-02, JXL-03, and JXL-05) anaerobically in MRS medium (Shanghai Macklin Biochemical Technology Co., Ltd.) at 37°C for 24 h. Once the bacterial concentration reached 10^9 CFU/mL, the culture was centrifuged at 7,000 rpm for 15 min. The supernatant was discarded, and the bacterial pellet was washed several times with sterile PBS. The bacteria were then resuspended in complete DMEM medium and incubated anaerobically at 37°C for an additional 2 h. The supernatant was collected for subsequent cell experiments.

L4-stage C. elegans were collected in a centrifuge tube, and 1 mL of M9 buffer was added to wash away excess E. coli OP50. After allowing the worms to settle, the supernatant was removed, and this washing step was repeated three times. Next, 1 mL of lysis solution was added and thoroughly mixed. After discarding the supernatant, the worms were washed three times with M9 buffer via centrifugation (3,000 rpm, 1 min), retaining the eggs. The eggs were then transferred to new NGM plates and incubated at 20°C in a biochemical incubator. After approximately 16–18 h, the eggs developed into L1-stage larvae.

B. pseudocatenulatum strains (JXL-01, JXL-02, JXL-03, and JXL-05) and E. coli OP50 were cultured in liquid medium until reaching the logarithmic phase. After centrifugation at 8,000 rpm for 10 min at 4°C, the bacterial pellets of B. pseudocatenulatum were collected and resuspended in M9 buffer to an OD of 0.5. A total of 100 μL of the bacterial suspension was then applied to 60 mm NGM solid plates containing the worm eggs and incubated at 20°C. After approximately 3 to 4 days, the worms developed into L4-stage larvae, with E. coli OP50 serving as a control group. Following an additional 3 days of continued culture with the bacterial suspension, 20 randomly selected worms were subjected to Oil Red O staining to measure fat accumulation. The worms were rinsed three times with M9 buffer, anesthetized with 25 mM levamisole hydrochloride, and treated with 200 μL of 4% paraformaldehyde solution for 20 min. The solution was removed, and the worms underwent three freeze–thaw cycles using liquid nitrogen. Subsequently, the worms were stained with 60% isopropanol and Oil Red O solution for 5 h. Excess dye was rinsed off with M9 buffer, and the staining was observed under an optical microscope to analyze fat content.

Mouse mesenchymal stem cells (C3H10 cells) were generously provided by Professor En-Dong Zhu’s research group at the Endocrinology Institute of Tianjin Medical University. The cells were cultured in DMEM medium (Gibco, United States) at 37°C, supplemented with 10% fetal bovine serum (FBS) (Hyclone, United States) and 100 U/mL penicillin/streptomycin (Beijing ZSbio). They were divided into two groups: a control group and a B. pseudocatenulatum metabolite intervention group (B.p), with the intervention lasting for 24 h. Following this, the cells were treated with specific differentiation media to induce differentiation into adipocytes. Differentiation Induction Agent A, consisting of DMEM high-glucose complete medium supplemented with 0.2 μg/mL dexamethasone, 5 μg/mL insulin, 100 μM indomethacin, and 250 μM IBMX, was applied for 3 days. This was followed by Differentiation Induction Agent B, which included DMEM high-glucose complete medium with 5 μg/mL insulin, administered for an additional 3 days. This differentiation process ensured that C3H10 cells functionally mimic adipose tissue, making them suitable for subsequent fat accumulation studies. For co-culture experiments, we used 10% of the metabolites from the bacterial cultures, ensuring that the conditions for cell differentiation were consistent across all experiments.

Cell viability was assessed using the CCK-8 Cell Counting Kit (Shanghai Taosu Bio Technology Co., Ltd.) according to the manufacturer’s instructions. C3H10 cells were seeded in a 96-well plate with 100 μL of cell suspension per well and incubated at 37°C with various concentrations of B. pseudocatenulatum for 24 h. Following treatment, 10 μL of CCK-8 reagent was added to each well and incubated for an additional 30 min. Absorbance at 450 nm was then measured using a multifunctional microplate reader (Thermo Fisher Scientific), and cell viability percentages were calculated for each concentration.

Oil Red O staining was performed, and lipid droplets were analyzed using cell imaging. The procedure followed the Oil Red O kit instructions (Beijing Solarbio Science & Technology Co., Ltd.). First, Solution A and Solution B were mixed in a 3:2 ratio to prepare the staining solution. Cells were fixed with ORO fixative for 25 min, followed by staining with freshly prepared Oil Red O solution for the recommended time. After staining, cells were washed to remove excess dye, and the cell nuclei were counterstained with Mayer’s Hematoxylin, followed by washing with distilled water. Finally, the cells were treated with ORO buffer for 1 min and observed under a microscope to analyze the lipid droplets.

Total RNA was purified from cells using the Total RNA Extraction Kit (Tiangen Biotech Co., Ltd.) according to the manufacturer’s instructions. After quantifying the RNA concentration and purity, 1 μg of RNA was reverse transcribed using the M-MuLV First Strand cDNA Synthesis Kit (Tiangen Biotech). The resulting cDNA was then amplified on a PCR cycler (Eppendorf, Germany) using the SYBR Green Real-Time PCR Kit (Tiangen Biotech). Relative mRNA levels were determined using the 2 − Δ Δ C T method, with GAPDH mRNA serving as the reference.

The primer sequences used in the study are shown in the Supplementary Table S1.

Statistical analysis in this study was conducted using several methods. ML analysis was performed using Python (version 3.10.2), while the Mann–Whitney U test was applied for non-parametric data, with p-values adjusted using the Bonferroni method and a significance level set at p < 0.01. Experimental data were analyzed using SPSS (version 25.0). Data normality was assessed, and for normally distributed data, independent samples t-tests were used to compare two groups. Results were expressed as mean ± standard deviation (SD), with a p-value of p < 0.05 considered statistically significant. Data processing for image analysis was carried out using Fiji ImageJ. These methods ensured rigorous and accurate data analysis, providing reliable and meaningful results for the study.

From the GMrepo database, a total of 14,028 samples containing human gut microbiome information were downloaded. After sample screening and processing, samples that did not meet the criteria were excluded, leaving 5,259 samples. Using BMI ≥25 as the threshold for overweight classification, the samples were divided into an overweight group and a control group. Further stratified screening of the samples resulted in the final inclusion of 3,606 samples, with 1,563 in the normal group and 2,043 in the overweight group, as illustrated in Figure 1 and detailed in Supplementary Table S1.

The study included a total of 4,419 bacterial species. Analysis of the gut microbiota in the two sample groups revealed that the relative abundance and frequency of 705 bacterial species differed significantly between the groups (p < 0.01). After applying Bonferroni correction, 283 bacterial species remained significantly different (adjusted p-value <0.01). These bacterial species were then ranked by frequency of occurrence. Considering model complexity and computational power, the top 40 bacterial species were selected as features for constructing the ML model (Table 1). Additionally, to enhance prediction accuracy, gender and age were included as feature variables, given their potential influence on the gut microbiome composition (Hales et al., 2018; Ge et al., 2024).

The analysis revealed significant differences in gut microbiota diversity between the overweight and control groups. Figure 2A presents a stacked bar chart illustrating the changes in microbial composition, emphasizing the substantial impact of overweight status on gut microbiota diversity and composition. Figure 2B displays boxplots comparing the Shannon index and Chao1 index, showing that the control group exhibited a wider range and higher median values for both indices. This further reinforces the observation of reduced microbial diversity in individuals with overweight status. Specifically, the Shannon index, which reflects the richness and evenness of microbial communities, was significantly lower in the overweight group compared to the control group. Similarly, the Chao1 index, an indicator of species richness, also demonstrated a significant decrease in the overweight group, indicating a shift in microbial diversity associated with overweight status. Figure 2C presents a heatmap generated from the relative abundance of microbial taxa, clearly illustrating the clustering of samples, with distinct groupings observed between the overweight and control groups. Specific taxa showed significant differences in abundance, supporting the findings from the diversity indices.

Using BMI classification as the response variable and incorporating the 40 bacterial species listed in Table 1, along with gender and age as feature variables, we constructed prediction models using XGBoost, logistic regression, decision tree, and support vector machine (SVM). The results represent the average performance metrics (accuracy, precision, recall, F₁-score, and AUC) obtained from 5-fold cross-validation on the validation data. Among these models, the XGBoost model achieved the highest overall performance (Figure 3A), demonstrating its suitability for this task.

Subsequently, we used GridSearchCV to fine-tune the parameters of the XGBoost model. The performance metrics and ROC curve of the optimized model are shown in Figures 3B,C. The fine-tuned model achieved an accuracy of 0.8144, a precision of 0.8191, a recall of 0.8481, and an F₁-score of 0.8333. Additionally, the AUC was 0.90, indicating that the model performs well and is suitable for further research.

The importance matrix of the XGBoost model, shown in Figure 4, highlights the top 10 most significant variables contributing to the model. These variables are host age, Akkermansia muciniphila, Alistipes putredinis, sex, Alistipes finegoldii, Blautia obeum, Barnesiella viscericola, Alistipes onderdonkii, B. pseudocatenulatum, and Anaerotruncus colihominis.

By calculating the Shapley values for each feature, we gain a clear understanding of how each feature influences the model’s predictions. This is crucial for revealing feature importance and the predictive mechanisms of the model (Ning et al., 2022). To identify the features with the most significant impact on the predictive model, we plotted the SHAP summary plot for the XGBoost model (Figure 5). Higher SHAP values for a feature indicate a greater likelihood of being overweight.

Our analysis revealed that features such as Akkermansia muciniphila, Barnesiella viscericola, Alistipes onderdonkii, B. pseudocatenulatum, and Anaerotruncus colihominis have negative SHAP values, suggesting that their abundance is inversely related to BMI. In contrast, Odoribacter splanchnicus exhibits a positive impact on BMI. While features like age and sex significantly contribute to the model’s predictions, their biological implications are less clear in terms of explanatory power.

SHAP values illustrate the contribution of each feature to the final prediction, providing a clear explanation of the model’s predictions for individual patients. As shown in Figure 6, we selected an overweight sample to demonstrate the model’s interpretability. In this case, age had the highest positive contribution to BMI, with a SHAP value of 0.62. Conversely, B. pseudocatenulatum, Roseburia hominis, and Akkermansia muciniphila had negative contributions to BMI, with SHAP values of −0.58, −6.1, and −0.92, respectively.

This study employed machine learning (ML) to predict several bacterial strains associated with BMI. To validate the potential functions of these strains, we examined the effect of B. pseudocatenulatum on fat reduction using Caenorhabditis elegans as a model, with Escherichia coli OP50 as the blank control. Four strains of B. pseudocatenulatum (JXL-01, JXL-02, JXL-03, and JXL-05) were selected based on their availability and prior verification of their fat-reducing properties, which were identified through our ML analysis of bacterial strains associated with BMI. Additionally, these strains have been implicated in obesity-related studies (Cano et al., 2013), providing further justification for their selection. These initial findings allowed us to identify the most effective strains for further cell experiments, and lipid accumulation in C. elegans was compared after feeding with each of these four B. pseudocatenulatum strains.

Results showed that, compared to the E. coli OP50 group, fat content in nematodes fed with the four strains of B. pseudocatenulatum was significantly reduced (Figure 7A). Further analysis of red fluorescence intensity, which indicates fat content, revealed that the strains JXL-01, JXL-02, JXL-03, and JXL-05 resulted in significant reductions in fat fluorescence intensity by 39.25, 24.95, 38.6, and 19.29%, respectively (Figure 7B). These findings suggest that B. pseudocatenulatum can significantly reduce fat deposition in nematodes, thereby influencing fat metabolism.

Based on the results from the C. elegans experiment, we selected the B. pseudocatenulatum JXL-01 strain for further investigation into its effect on adipogenic differentiation using mouse mesenchymal stem cells (C3H10 cells). The cells were divided into two groups: a control group and a B. pseudocatenulatum intervention group (B.p). To assess the impact of B. pseudocatenulatum metabolites on cell viability, a CCK8 assay was first performed. We chose to co-culture C3H10 cells with 10% of the bacterial metabolites (Figure 8A). To evaluate the effects on adipogenic differentiation, we used Oil Red O staining to visualize lipid accumulation. The results showed that lipid droplets were significantly reduced in the B.p group compared to the control group (Figure 8B). Quantitative analysis of the lipid content confirmed the reduction in lipid droplets in the B.p group (Figure 8C). Furthermore, we assessed the expression of key adipogenic transcription factors, including PPARγ and FABP4, which are upregulated during adipocyte differentiation. The mRNA levels of PPARγ and FABP4 were significantly inhibited in the B.p group (Figure 8D–E), suggesting that B. pseudocatenulatum can significantly reduce adipogenic differentiation. These findings demonstrate that B. pseudocatenulatum JXL-01 significantly influences fat deposition and adipogenic differentiation in C3H10 cells, promoting a reduction in fat accumulation.