Integrated Metabolic and Inflammatory Clustering Reveals Distinct Risk Profiles for Digestive Diseases

Abstract

Emerging research highlights the complex relationship between metabolic dysfunction and chronic low-grade inflammation, which disrupts gut homeostasis and drives disease progression. However, most current studies evaluate metabolic and inflammatory markers separately, relying on basic indicators such as body mass index (BMI) or individual biomarkers. In this study, a scalable clustering framework is developed to integrate six clinical parameters in 398 432 participants from the UK Biobank, identifying four distinct metabolic-inflammatory subtypes. Cox proportional hazards models demonstrate significant associations between these subtypes and digestive disease risk. Using 251 plasma metabolites and elastic net regression, cluster-associated metabolite signatures are identified. Mediation analyses indicate that metabolic signatures mediate the association between clusters and digestive disease risk. Machine learning algorithms are applied to construct disease-specific metabolic risk scores, achieving C-indices above 0.70 for ten digestive disease endpoints. Explainable machine learning approaches further identify both shared and disease-specific predictors, with glycoprotein acetyls, valine, tyrosine, and fatty acids emerging as key risk indicators. This integrative approach provides a comprehensive framework for digestive disease risk assessment and offers novel insights into the metabolic mechanisms underlying disease susceptibility.

Keywords: cluster analysis; digestive diseases; machine learning; metabolomics.

Figure 1

The proportion and features of different clusters in the cohort study. A) Radar plots summarize scaled variables for each cluster. Values closer to the outer ring are higher than the cohort average for each of the key variables. B) Elbow plot for determining the optimal number of clusters. C)The violin plot of the feature variables of different clusters. Abbreviations: BMI, body mass index; CRP, C‐reactive protein; NLR, Neutrophil‐to‐Lymphocyte Ratio; TyGBMI, TyG index multiplied by BMI; WHtR, Waist‐to‐Height Ratio; IS, inflammatory status; OHS, overweight with high strength; HSLS, healthy status with low strength; OIR, obesity with insulin resistance.

Figure 2

Cox analysis of the associations of anthropometric clusters with risks of digestive diseases, compared to Healthy Status with Low Strength, respectively. Model adjusted for age, age squared, sex, ethnicity, Townsend deprivation index, smoking status, alcohol drinking, education level, physical activity, medicinal intake (aspirin, non‐aspirin NSAIDs, and lipid‐lowering drugs use), and comorbidities (lipidaemia, hypertension, diabetes).

Figure 3

Volcano diagrams showing the differential metabolites across clusters. The red dot on the figure represents the upregulated metabolite, and the blue dot represents the downregulated metabolite. The y‐axis corresponds to log2 (fold change).

Figure 4

Associations between selected metabolites constituting the metabolic signature and four clusters. Each panel presented the associations between metabolites in the metabolic signature and relevant cluster. Colors indicate the direction of association, with red representing positive associations and blue indicating inverse associations. The darkness of the color corresponds to the magnitude of the association. Asterisks denote the significance level of associations (^* p < 0.05 and Bonferroni‐corrected p < 0.05). Abbreviations: C, cholesterol; CE, cholesteryl ester; FC, free cholesterol; M, medium; PL, phospholipid; TG, triglyceride; S, small; L, large; VLDL, very LDL; XL, very large; XS, very small; XXL, especially large.

Figure 5

Observed event frequency for 25 incident disease categories plotted against best model percentiles. Age is represented by varying sizes of dots, while sex is distinguished by different colors of the dots.

Figure 6

Stacked bar chart of standardized SHAP values from machine learning models across 25 disease categories. We highlighted 12 metabolites that exhibited the most important discriminatory value in three or more disease categories. Abbreviations: SHAP SHapley Additive exPlanations.

Figure 7

Associations between metabolites shown in Figure 6 and 25 outcomes. In each square, the first row presents the HR values from the Cox analysis, while the second row indicates the corresponding 95% confidence intervals. The color of squares indicates the effect size (HR). Model adjusted for age, age squared, sex, ethnicity, Townsend deprivation index, smoking status, alcohol drinking, education level, physical activity, medicinal intake (aspirin, non‐aspirin NSAIDs, and lipid‐lowering drugs use), and comorbidities (lipidaemia, hypertension, diabetes). Abbreviations: HR, hazard ratio.