This study was approved by the Ethics Committee of Qilu Hospital of Shandong University (Approval number: KYLL-202212-010), and written informed consent was obtained from all subjects. A total of 49 UC patients, 20 CD patients, and 26 control subjects were recruited from hospitalized patients in the Qilu Hospital of Shandong University between April 2023 and June 2024. Eligible patients were between 18 and 75 years old, with a diagnosis of ulcerative colitis (UC) or Crohn’s disease (CD) confirmed according to the European Crohn’s and Colitis Organization criteria (Magro et al., 2017; Gomollón et al., 2017). Clinical activity was scored using the Mayo score for UC (Paine, 2014) and the Crohn’s disease activity index (CDAI) for CD (Best et al., 1976). The inclusion criterion for control group was healthy adults between 18 and 75 years of age.

To minimize inter-individual variation due to hydration status and recent dietary intake, all urine samples were collected as first-morning voids following an overnight fast of at least 8 h. Participants were instructed to avoid food and fluid intake after midnight and to collect their first urination immediately upon waking. This standardized collection helps ensure consistency in urinary metabolite concentrations by reducing the influence of short-term fluctuations in fluid balance and nutritional status.

To minimize possible confounding effects in the results, exclusion criteria were as follows: indeterminate colitis; structural abnormalities of the gastrointestinal tract; urinary dysfunction or infection; and combined with diabetes mellitus, inborn errors of metabolism, renal or hepatic disease, severe infection, or evidence of malignancy that could affect the results of this study. All participants adhered to the same exclusion criteria.

Fresh midstream urine samples were collected from subjects in the morning. Immediately after collection, the urine was centrifuged at 1,000 × g for 10 min to collect the supernatant. After centrifugation at 12,000 × g for 10 min at 4°C, the middle layer of the supernatant was aspirated into a 1.5 mL centrifuge tube and stored at −80°C until use.

A total of 49 central carbon metabolism-related standards were accurately weighed and individually dissolved in 50% methanol-water to prepare single-compound stock solutions. Appropriate volumes of each stock solution were combined and diluted with 50% methanol-water to obtain a mixed working standard solution at suitable concentrations.

Isotopically labeled internal standards—including succinic acid-D4, L-carnitine-D3, cholic acid-D4, and salicylic acid-D4—were also weighed and dissolved individually in 50% methanol-water to prepare their respective stock solutions. Equal volumes of these solutions were then combined and diluted to prepare a mixed internal standard solution with final concentrations of 5 μg/mL (succinic acid-D4), 5 μg/mL (L-carnitine-D3), 20 μg/mL (cholic acid-D4), and 30 μg/mL (salicylic acid-D4).

To construct the calibration curves, 50 μL of the working standard solution was mixed with 10 μL of the isotope internal standard mixture and 140 μL of acetonitrile. The mixture was vortexed for 1 min and centrifuged at 14,000 rcf for 20 min at 4°C. Then, 100 μL of the resulting supernatant was transferred to a 1.5 mL centrifuge tube, followed by the addition of 25 μL of 200 mM 3-nitrophenylhydrazine hydrochloride (3NPHHCl) and 25 μL of 120 mM EDCHCl solution containing 6% pyridine. The mixture was vortexed for 30 s, briefly centrifuged for 5 s, and incubated at 60°C for 40 min using a thermostatic shaker. This reaction forms hydrazone derivatives with carbonyl-containing metabolites, which is known as derivatization. After the reaction, the sample was vortexed again for 30 s, centrifuged at 14,000 rcf for 20 min at 4°C, and the supernatant was transferred into an autosampler vial for LC-MS/MS analysis. Calibration curves for all 49 central carbon metabolites were constructed using 12 concentration points, each of which was injected in triplicate. The resulting peak area ratios (analyte/internal standard) were used to generate calibration curves through weighted linear regression. Detailed infromation were listed in Supplementary Table S5.

A 20 μL aliquot of urine was mixed with 10 μL of isotope internal standard, 30 μL of 50% methanol, and 140 μL of acetonitrile, and vortexed for 1 min. After centrifugation at 14,000 rcf for 20 min at 4°C, 100 μL of the supernatant was transferred to 1.5 mL centrifuge tubes containing 25 μL of 200 mM 3NPH.HCL and 25 μL of 120 mM EDC. HCL (containing 6% pyridine) solution. The mixture was shaken in a vortex for 30 s, centrifuged for 5 s, and then reacted at 60°C for 40 min in a thermostatic oscillator. After the reaction, the mixture was vortexed for 30 s and centrifuged at 14,000 rcf for 20 min at 4°C. Finally, the supernatant was transferred to an injection vial for subsequent analysis.

Targeted metabolomic analysis of the urine samples was performed using liquid chromatography–tandem mass spectrometry (LC-MS/MS) on an ExionLC AD system coupled with a QTRAP® 6500+ mass spectrometer (Sciex, United States) at Majorbio Bio-Pharm Technology Co. Ltd. (Shanghai, China).

Chromatographic separation was performed on an ExionLC™ AD system equipped with a Waters HSS T3 chromatography column (2.1 × 150 mm, 1.8 μm). Mobile phase A was 0.03% formic acid in water, and mobile phase B was 0.03% formic acid in methanol. The gradient was as follows: 0.0–2.0 min, hold at 1% B; 2.0–8.0 min, from 1% to 22% B; 8.0–12.0 min, hold at 22% B; 12.0–13.0 min, from 22% to 40% B; 13.0–17.0 min, from 40% to 65% B; 17.0–19.0 min, hold at 65% B; 19.0–20.0 min, from 65% to 100% B; 20.0–21.0 min, hold at 100% B; 21.0–21.01 min, from 100% to 1% B; 21.01–22.0 min, hold at 1% B. The injection volume was 2 μL, and the column temperature was 40°C.

Mass spectrometric analyses were performed on a QTRAP® 6500+ mass spectrometer (Sciex, United States) equipped with an electrospray ionization (ESI) source operating in negative and positive modes. The parameters were set as follows: source temperature (TEM) at 550°C; curtain gas (CUR) at 35 psi; collision gas (CAD) at medium; both Ion Source Gas1 and Gas2 at 55 psi; IonSpray Voltage (IS) at +4500/−4500 V.

Data acquisition was conducted in multiple reaction monitoring (MRM) mode, a highly sensitive and specific mass spectrometric technique that enables the selective quantification of predefined metabolites. In MRM mode, the mass spectrometer first selects precursor ions (parent ions) of interest in the first quadrupole (Q1), induces fragmentation through collision-induced dissociation (CID) in the second quadrupole (Q2), and then monitors specific product ions in the third quadrupole (Q3). This approach ensures high analytical specificity, sensitivity, and reproducibility for the targeted detection and quantification of known metabolites. Metabolite identification was carried out in a targeted manner using MRM based on authentic reference standards. Both precursor and corresponding product ions were monitored for each metabolite, enabling high-confidence identification. The complete list of ion pairs and retention time data is provided in Supplementary Table S3.

The quality control (QC) sample was a mixed standard solution at a moderate concentration level, primarily used to evaluate the stability of the analytical system. In this study, the QC sample was prepared using the C10 concentration level (i.e., the 10th calibration level, see Supplementary Table S5). During the LC-MS/MS analysis, one QC sample was injected after every 5 to 10 sample injections to monitor instrument stability and repeatability. The stability of QC signals across the analytical batch further validated the reproducibility of the instrument and analytical conditions. The relative standard deviations (RSDs) of all target analytes were below 15%, indicating that the method and analytical system were stable and reliable.

To evaluate the robustness of the LC-MS/MS platform and exclude the potential influence of instrumental variation, we conducted method validation for 49 central carbon metabolites. QC samples at low, medium, and high concentrations were analyzed in six replicates within a single day and across 3 days. Intra-day and inter-day precision were calculated as RSD%, and recovery rates were calculated based on spiked and measured concentrations. Acceptable thresholds were defined as RSD% ≤15% and recovery rates within 80%–120%, following common metabolomics validation guidelines. All metabolites met these criteria, confirming the stability of the measurement system.

The raw data were processed by Sciex software OS by using the default parameters and assisting manual inspection. A linear regression standard curve was created with the ratio of the mass spectral peak area of the analyte to the internal standard peak area as the vertical coordinate and the concentration of the analyte as the horizontal coordinate. The ratio of the mass spectral peak area of the analyte to the internal standard peak area was substituted into the linear equation to calculate the sample’s concentration (Supplementary Tables S1, S2).

SPSS version 27.0.1 and R software (version 4.4.2) were used to process and analyze clinical and metabolomic data. Normally distributed data were expressed as mean ± standard deviation and compared using a t-test. Non-normally distributed data were expressed as median and interquartile ranges and analyzed using the Mann-Whitney U-test. Qualitative data were presented using frequencies and percentages and compared with the Chi-square test. Multivariate statistical analysis was performed using principal component analysis (PCA), partial least squares discriminant analysis (PLS-DA), and orthogonal partial least squares discriminant analysis (OPLS-DA) to identify differential metabolites and visualized by the ggplot2 package. Seven-fold cross-validation and permutation tests were used to evaluate the quality of the model. Enrichment of Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways was also performed. Multivariable logistic regression, adjusting for age, sex, smoking, weight, and height, was used to evaluate the association of each metabolite with disease status. Metabolites with a VIP >1 in OPLS-DA analysis and Padj values <0.05 were defined as significantly differential metabolites.

Machine learning operations were performed using the “tidymodels” package. The SMOTE algorithm was utilized to address unbalanced class distribution issues in the models. Several optimized machine learning models were used to discriminate between IBD and control group: decision tree (DT), random forest (RF), ridge regression (Ridge), support vector machine (SVM), light gradient boosting machine (LightGBM), and neural networks (NN). The hyperparameters were obtained using grid search with 5-fold cross-validation. The performance of these machine learning methods was evaluated by the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA). We also utilized Shapley additive explanation (SHAP) values to improve the interpretability of the final model.

A total of 95 participants were enrolled in this study, including 49 patients with ulcerative colitis (UC), 20 patients with Crohn’s disease (CD), and 26 control individuals. The clinical characteristics of the three groups are summarized in Table 1. Statistically significant differences were observed in age and weight between the IBD groups and the control group (p < 0.01), as well as in hematological parameters such as hemoglobin, hematocrit, and platelet count. These differences highlight the clinical and physiological alterations associated with IBD, which may be reflected in the metabolomic profiles.

Targeted metabolomic profiling identified 49 urinary metabolites associated with central carbon metabolism across all subjects. Principal component analysis (PCA) showed limited separation among the three groups, indicating the necessity for supervised methods. Orthogonal partial least squares discriminant analysis (OPLS-DA) revealed clear group separation between UC and controls (R ²X = 0.906, R ²Y = 0.598, Q ² = 0.439), and between CD and controls (R ²X = 0.543, R ²Y = 0.733, Q ² = 0.594), suggesting that metabolite patterns differ significantly between disease and non-disease states (Figure 2, Supplementary Figure S1).

Hierarchical clustering of metabolite intensities showed distinct metabolic signatures among UC, CD, and control groups (Figure 3A). Metabolites such as glucosamine-6-phosphate were elevated in UC, whereas cis-aconitic acid, trehalose-6-phosphate, nicotinic acid, and glucaric acid were more abundant in CD compared to both UC and control groups. KEGG pathway enrichment analysis further supported that metabolic pathways including the TCA cycle, glucagon signaling, PI3K-Akt, and mTOR signaling were perturbed in IBD, suggesting potential disease-specific metabolic reprogramming (Figures 3B,C).

Logistic regression models adjusting for age, sex, smoking status, weight, and height identified several significantly differential metabolites. Six metabolites—xylose, isocitric acid, fructose, L-fucose, GlcNAc, and glycolic acid—were significantly altered in UC patients compared to controls. In CD, three metabolites—xylose, L-fucose, and citric acid—were significantly changed. Notably, L-fucose and xylose were elevated in both UC and CD, highlighting their potential as shared biomarkers. The differences in metabolite levels were more pronounced in patients with higher disease activity, as shown in Figure 4.

To assess the potential impact of instrumental variation, we performed method validation based on intra- and inter-day precision as well as recovery rate analyses. A total of 49 central carbon metabolites were evaluated using quality control (QC) samples prepared at three concentration levels (low, medium, and high). Each level was analyzed in six replicates within a day (intra-day precision) and across 3 days (inter-day precision). The RSD of intra-day precision ranged from 1.20% to 11.92%, and that of inter-day precision ranged from 2.33% to 12.66%. Recovery rates ranged from 85.33% to 113.71%. These results confirm that the analytical platform is highly stable and reproducible under the current workflow. Therefore, the observed metabolite differences are unlikely to be attributed to instrumental drift. The detailed information was listed in Supplementary Table S4.

Six machine learning algorithms were trained to distinguish IBD subtypes from controls. Among them, the random forest (RF) model achieved the best performance for UC, with a mean cross-validated AUC of 0.84 (SE = 0.036), and the support vector machine (SVM) model performed best for CD, with a mean AUC of 0.93 (SE = 0.035). Calibration curves showed good model fit, and decision curve analysis (DCA) demonstrated that both models offered higher net clinical benefit compared to baseline strategies (Figures 5A–F).

Stratified analysis by disease activity showed that the diagnostic accuracy remained high in both mild/remission and moderate/severe stages. The AUC for distinguishing UC from controls was 0.864 in remission/mild patients and 0.904 in moderate/severe cases. Similarly, CD patients showed AUCs of 0.963 in remission and 0.988 in active disease, confirming the robustness of the models across disease stages (Figures 6A–D).

Longitudinal data from a subset of patients with samples collected at ≥6-month intervals showed consistent metabolite profiles during periods of equivalent disease activity, indicating good temporal stability of the biomarkers (Figure 7).

To interpret the contributions of individual metabolites to model predictions, SHAP (Shapley Additive exPlanations) analysis was performed (Lundberg and Lee, 2017). In the RF model for UC, L-fucose, xylose, and GlcNAc were identified as the top predictors, with positive SHAP values indicating higher disease risk (Figure 8A). For CD, citric acid and xylose were most influential in the SVM model, with decreased citric acid and elevated xylose contributing to increased CD probability (Figure 8B). These findings confirm the biological relevance and diagnostic utility of the selected metabolites.