Integrating dietary patterns and metabolomics with machine learning for metabolic syndrome risk assessment in a rural population Authors: Ming Sun, Yixin Zhang, Jiajin Hu, Yang Yu, Yang Liu, Zhaoqing Sun, Liqiang Zheng, Yanan Ma, Deliang Wen Published: 2026 ABSTRACT This study aimed to identify dietary patterns and metabolomic signatures associated with metabolic syndrome (MetS) and develop machine learning models integrating both dietary and metabolomic data for risk assessment. In this cross-sectional study of 2810 adults, four dietary patterns derived using principal component analysis were associated with MetS. Serum metabolomic profiling was conducted in 400 participants using liquid chromatography-tandem mass spectrometry. Metabolic signatures of the meat and traditional Chinese patterns remained positively associated with MetS, whereas the rice and tuber pattern metabolic signature remained inversely associated. Metabolomic analysis revealed 71 differential metabolites associated with MetS, including elevated monoacylglycerols and decreased metabolites such as 4-Hydroxyisoleucine. A metabolomic risk score based on key metabolites showed a strong association with MetS (OR = 23.55, 95% CI: 11.77–47.12). Integrating dietary patterns with MetS-specific biomarkers significantly improved risk assessment models (AUC: 0.759–0.820) compared to using dietary patterns and pattern-associated metabolites (AUC: 0.640–0.709), with the Support Vector Machine (SVM) model performing optimally (AUC: 0.820). Specific dietary patterns and their associated metabolomic signatures are significantly associated with MetS. Integrating dietary and metabolomic data markedly improves the SVM model’s risk assessment accuracy for MetS risk, offering a robust approach for precision prevention in high-risk populations. Introduction Metabolic syndrome (MetS) is a clinical cluster of metabolic abnormalities. According to the Chinese Diabetes Society (CDS)1, MetS is characterized by the presence of at least three of the following conditions: abdominal obesity, hyperglycemia, hypertension, and dyslipidemia. This cluster poses a significant global public health threat due to its high and rising prevalence and its strong association with cardiovascular diseases and cancers2–4. In China, the burden of MetS is substantial, with reported prevalence rates as high as 33.9%, affecting an estimated 454 million adults and showing notable regional and gender disparities5. This escalating epidemic underscores the critical need to identify modifiable risk factors for early prevention and intervention. Among various lifestyle factors, dietary patterns play a pivotal role in the development of MetS. Compared with studies of single foods or nutrients, dietary patterns have been introduced to determine the broader impacts of diet through a comprehensive analysis of dietary intakes, providing a more comprehensive assessment of the association between diet and MetS6,7. Internationally, the Mediterranean and Dietary Approaches to Stop Hypertension (DASH) dietary patterns have been consistently associated with a lower risk of MetS8–10. However, given the significant regional variations in dietary habits in China11, the specific associations between indigenous dietary patterns and the risk of MetS, particularly in high-risk northern rural populations, require further elucidation12. Moreover, metabolomics captures a broad spectrum of both endogenous and exogenous metabolites, offering a powerful approach to investigate metabolic perturbations associated with MetS and dietary intake13,14. Metabolomics serves not only to identify potential biomarkers for MetS15 but also to discover objective biomarkers of dietary patterns, thereby overcoming the limitations of self-reported dietary assessment16. Therefore, metabolomics can provide deeper insights into metabolic responses to different dietary patterns and, among others, their association with metabolic disease17. For instance, a U.S. cohort study identified a Mediterranean diet metabolomic signature comprising 67 metabolites and found that both the self-reported diet and its metabolic signature were inversely associated with cardiovascular disease risk18. Similarly, a study in the Netherlands identified that specific plasma metabolites, including apolipoprotein A1 and fatty acid ratios, mediate the association between healthy dietary patterns and glucose metabolism status19. A study in southern China has successfully characterized dietary patterns using metabolomics in a community-based cohort and associated these metabolomic patterns with metabolic diseases, such as hypertension and hyperglycemia20. Despite this critical finding, the application of untargeted metabolomics to systematically dissect the diet-MetS interplay, particularly in high-risk rural northern Chinese populations, remains largely unexplored. Advancements in machine learning (ML) provide robust analytical frameworks for integrating complex, multi-dimensional data to build predictive models21,22. However, the performance of these models fundamentally depends on the quality and biological relevance of the input data. In metabolomics, this relevance is determined by the extent to which the analytical approach captures metabolic pathways associated with the outcome of interest. If key pathways are not adequately represented, even the most sophisticated ML algorithms may fail to achieve meaningful predictive performance. Despite the promise of such integrative approaches, their application remains limited. A Korean study developed ML models for MetS prediction by incorporating demographic, clinical, and laboratory factors; however, its dietary assessment was limited to a single food item (seaweed), potentially restricting its ability to capture overall dietary patterns23. Similarly, a Chinese study constructed MetS prediction models using health examination data while omitting dietary factors altogether24. The integration of metabolomic data remains particularly limited. A UK cohort study demonstrated that incorporating metabolomic biomarkers into traditional risk factors significantly improved diabetes prediction accuracy, particularly for waist circumference, blood pressure, and dyslipidemia25. However, a separate Chinese study developed MetS risk models using only conventional metabolic indicators, including inflammatory factors, erythrocyte parameters, and lipid metabolism markers26. The synergistic combination of these multimodal data could substantially enhance model accuracy and refine individual risk stratification, thereby advancing personalized nutrition and precision prevention for MetS. Therefore, this study aims to develop and validate a comprehensive risk assessment model for MetS by integrating dietary patterns and metabolomic biomarkers using ML algorithms. The research is designed to first identify significant dietary patterns associated with MetS in a rural northern Chinese population. Subsequently, an untargeted metabolomics analysis will be conducted to identify key metabolites and metabolic signatures associated with both the identified dietary patterns and MetS status. Finally, the study will construct and compare multiple ML algorithms to develop risk assessment models that synergistically integrate both dietary and metabolomic data, thereby optimizing assessment performance. Results Study population characteristics A total of 3377 adults were initially enrolled in the study. After applying the exclusion criteria, 567 participants were removed due to incomplete dietary questionnaires (n = 265), missing anthropometric or biochemical data required for MetS definition (n = 245), or extreme energy intake (<500 or ≥4000 kcal/day) (n = 57). As a result, 2810 participants were included in the final analysis. 786 (27.97%) participants were diagnosed with MetS, while the remaining 2024 (72.03%) were classified as non-MetS (Table 1). MetS prevalence was slightly higher in females (29.72%) than in males (24.67%, P = 0.004). Age distribution showed no significant difference between the two groups (P = 0.164), with 29.02% MetS in participants aged >60 years. Moreover, the MetS group had significantly higher BMI (26.70 [3.82] vs. 23.88 [4.19] kg/m²), WC (90.00 [9.00] vs. 82.00 [11.00] cm), SBP (143.33 [25.67] vs. 130.67 [27.00] mmHg), DBP (88.33 [14.00] vs. 83.00 [14.67] mmHg), FBG (6.25 [2.47] vs. 5.26 [0.76] mmol/L), and TG (2.05 [1.23] vs. 1.06 [0.60] mmol/L), while HDL (1.14 [0.36] vs. 1.47 [0.43] mmol/L) was significantly lower (all P < 0.001).Table 1Characteristics of the study participants by MetSCharacteristicOverallMetSPNoYesN (%)28102024 (72.03)786 (27.97)Sex0.004 Male973 (34.63)733 (75.33)240 (24.67) Female1837 (65.37)1291 (70.28)546 (29.72)Age (years)0.164 ≤601235 (43.95)906 (73.36)329 (26.64) >601575 (56.05)1118 (70.98)457 (29.02)Ethnicity0.295 Han1883 (67.01)1268 (72.65)515 (27.35) Mongolian927 (32.99)656 (70.77)271 (29.23)Marital status0.273 Married2456 (87.68)1778 (72.39)678 (27.61) Others345 (12.32)240 (69.58)105 (30.43)Education status0.336 Middle school or below1110 (39.50)788 (70.99)322 (29.01) Middle school above1697 (60.39)1233 (72.66)464 (59.03)Annual income (yuan)0.588 <10,0001473 (52.42)1806 (72.15)697 (27.85) ≥10,0001330 (47.33)212 (70.67)88 (29.33)Smoking status0.033 Yes948 (33.83)707 (74.58)241 (25.42) No1855 (66.17)1312 (70.77)542 (29.23)Physical activity<0.001 Light or below768 (27.33)492 (64.32)274 (35.31) Moderate or vigorous1994 (70.96)1492 (74.82)502 (25.18)Family history of diabetes0.003 Yes105 (3.74)62 (59.05)43 (40.95) No2705 (96.26)1962 (72.53)743 (27.47)Family history of hypertension<0.001 Yes525 (18.68)333 (63.43)192 (36.57) No2285 (81.32)1691 (74.00)594 (26.00)Energy intake (kcal/d)1571.23 (931.35)1555.60 (898.32)1622.77 (1000.74)0.083BMI (kg/m2)24.67 (4.55)23.88 (4.19)26.70 (3.82)<0.001WC (cm)85.00 (10.00)82.00 (11.00)90.00 (9.00)<0.001SBP (mmHg)134.67 (28.33)130.67 (27.00)143.33 (25.67)<0.001DBP (mmHg)85.00 (15.00)83.00 (14.67)88.33 (14.00)<0.001FBG (mmol/L)5.40 (1.05)5.26 (0.76)6.25 (2.47)<0.001TG (mmol/L)1.23 (0.89)1.06 (0.60)2.05 (1.23)<0.001HDL (mmol/L)1.38 (0.47)1.47 (0.43)1.14 (0.36)<0.001Values were median (interquartile range) or n (%). Bolded values indicated P < 0.05.Values of polytomous variables may not sum to 100% due to rounding.MetS metabolic syndrome, BMI body mass index, WC waist circumference, SBP systolic blood pressure, DBP diastolic blood pressure, FBG fasting blood glucose, TG triglyceride, HDL high-density lipoprotein, SD standard deviation. Table 1Characteristics of the study participants by MetSCharacteristicOverallMetSPNoYesN (%)28102024 (72.03)786 (27.97)Sex0.004 Male973 (34.63)733 (75.33)240 (24.67) Female1837 (65.37)1291 (70.28)546 (29.72)Age (years)0.164 ≤601235 (43.95)906 (73.36)329 (26.64) >601575 (56.05)1118 (70.98)457 (29.02)Ethnicity0.295 Han1883 (67.01)1268 (72.65)515 (27.35) Mongolian927 (32.99)656 (70.77)271 (29.23)Marital status0.273 Married2456 (87.68)1778 (72.39)678 (27.61) Others345 (12.32)240 (69.58)105 (30.43)Education status0.336 Middle school or below1110 (39.50)788 (70.99)322 (29.01) Middle school above1697 (60.39)1233 (72.66)464 (59.03)Annual income (yuan)0.588 <10,0001473 (52.42)1806 (72.15)697 (27.85) ≥10,0001330 (47.33)212 (70.67)88 (29.33)Smoking status0.033 Yes948 (33.83)707 (74.58)241 (25.42) No1855 (66.17)1312 (70.77)542 (29.23)Physical activity<0.001 Light or below768 (27.33)492 (64.32)274 (35.31) Moderate or vigorous1994 (70.96)1492 (74.82)502 (25.18)Family history of diabetes0.003 Yes105 (3.74)62 (59.05)43 (40.95) No2705 (96.26)1962 (72.53)743 (27.47)Family history of hypertension<0.001 Yes525 (18.68)333 (63.43)192 (36.57) No2285 (81.32)1691 (74.00)594 (26.00)Energy intake (kcal/d)1571.23 (931.35)1555.60 (898.32)1622.77 (1000.74)0.083BMI (kg/m2)24.67 (4.55)23.88 (4.19)26.70 (3.82)<0.001WC (cm)85.00 (10.00)82.00 (11.00)90.00 (9.00)<0.001SBP (mmHg)134.67 (28.33)130.67 (27.00)143.33 (25.67)<0.001DBP (mmHg)85.00 (15.00)83.00 (14.67)88.33 (14.00)<0.001FBG (mmol/L)5.40 (1.05)5.26 (0.76)6.25 (2.47)<0.001TG (mmol/L)1.23 (0.89)1.06 (0.60)2.05 (1.23)<0.001HDL (mmol/L)1.38 (0.47)1.47 (0.43)1.14 (0.36)<0.001Values were median (interquartile range) or n (%). Bolded values indicated P < 0.05.Values of polytomous variables may not sum to 100% due to rounding.MetS metabolic syndrome, BMI body mass index, WC waist circumference, SBP systolic blood pressure, DBP diastolic blood pressure, FBG fasting blood glucose, TG triglyceride, HDL high-density lipoprotein, SD standard deviation. Characteristics of the study participants by MetS Values were median (interquartile range) or n (%). Bolded values indicated P < 0.05. Values of polytomous variables may not sum to 100% due to rounding. MetS metabolic syndrome, BMI body mass index, WC waist circumference, SBP systolic blood pressure, DBP diastolic blood pressure, FBG fasting blood glucose, TG triglyceride, HDL high-density lipoprotein, SD standard deviation. Associations between dietary patterns and metabolic syndrome Four distinct dietary patterns (meat, protein–sweets, traditional Chinese, and rice and tuber) were identified using principal component analysis (PCA) based on FFQ data, with their factor loadings detailed in Supplementary Table 2. Associations between the four dietary patterns and MetS risk were analyzed using logistic regression models with progressive adjustment for confounders (Table 2). After adjusting for age, sex, ethnicity, marital status, education, income, smoking, physical activity, family history of diabetes/hypertension, energy intake, BMI, and other dietary patterns (Model 4), highest adherence (Tertile 3 [T3]) to the meat pattern remained significantly associated with increased MetS risk (OR = 1.34, 95% CI: 1.04–1.74; P for trend = 0.026). Conversely, the protein-sweets pattern showed a inverse association at the highest adherence level (OR = 0.71, 95% CI: 0.56–0.91; P for trend = 0.009). The traditional Chinese pattern demonstrated the strongest positive association with MetS risk (T3: OR = 1.77, 95% CI: 1.39–2.26; P for trend < 0.001; per-SD: OR = 1.29, 95% CI: 1.06–1.57), while the rice and tuber pattern exhibited a significant inverse association (T3: OR = 0.66, 95% CI: 0.52–0.83; P for trend = 0.001; per-SD: OR = 0.80, 95% CI: 0.70–0.90). Furthermore, subgroup analyses confirmed the robustness of these associations across different population strata (Supplementary Tables 3–6).Table 2Associations between the different dietary patterns and MetS in different modelsDietary patternsNo. of CasesModel 1Model 2Model 3Model 4OR (95% CI)OR (95% CI)OR (95% CI)OR (95% CI)Meat pattern Tertile 1233/9371.001.001.001.00 Tertile 2274/9371.25 (1.02, 1.53)1.32 (1.07, 1.63)1.34 (1.07, 1.69)1.38 (1.09, 1.75) Tertile 3279/9361.28 (1.05, 1.57)1.45 (1.17, 1.80)1.36 (1.07, 1.72)1.34 (1.04, 1.74) P for trend0.0170.0010.0120.026 Per-SD1.08 (0.995, 1.17)1.12 (1.04, 1.22)1.07 (0.96, 1.19)1.09 (0.98, 1.22)Protein-sweets pattern Tertile 1283/9371.001.001.001.00 Tertile 2266/9360.92 (0.75, 1.12)0.93 (0.76, 1.14)0.97 (0.78, 1.21)0.90 (0.71, 1.12) Tertile 3237/9370.78 (0.64, 0.96)0.80 (0.66, 0.98)0.85 (0.68, 1.07)0.71 (0.56, 0.91) P for trend0.0180.0350.1680.009 Per-SD0.94 (0.86, 1.02)0.95 (0.87, 1.03)0.83 (0.73, 0.95)0.90 (0.77, 1.05)Traditional Chinese pattern Tertile 1217/9371.001.001.001.00 Tertile 2254/9361.24 (1.002, 1.52)1.30 (1.05, 1.60)1.36 (1.09, 1.72)1.33 (1.05, 1.69) Tertile 3315/9371.68 (1.37, 2.06)1.84 (1.49, 2.27)1.75 (1.39, 2.21)1.77 (1.39, 2.26) P for trend<0.001<0.001<0.001<0.001 Per-SD1.24 (1.14, 1.34)1.29 (1.19, 1.40)1.43 (1.21, 1.69)1.29 (1.06, 1.57)Rice and tuber pattern Tertile 1301/9371.001.001.001.00 Tertile 2261/9370.82 (0.67, 0.994)0.83 (0.68, 1.01)0.80 (0.64, 0.991)0.78 (0.63, 0.98) Tertile 3224/9360.67 (0.54, 0.82)0.70 (0.57, 0.87)0.68 (0.54, 0.85)0.66 (0.52, 0.83)P for trend<0.0010.0010.0010.001 Per-SD0.85 (0.78, 0.93)0.87 (0.80, 0.96)0.77 (0.69, 0.87)0.80 (0.70, 0.90)Model 1: Adjust for none. Bolded values indicated P < 0.05.Model 2: Age (<60, ≥60); Sex (Female, Male).Model 3: Adjust for Age (<60, ≥60); Sex (Female, Male); Ehnicity (Han, Mongolian); Marital status (Married, others); Education status (Middle school or below, Middle school above); Annual income (yuan) (<10,000, ≥10,000); Smoking status (Yes, No); Moderate physical inactivity (Light or below, Moderate or vigorous); Family history of diabetes (Yes, No); Family history of hypertension (Yes, No), Energy intake (kcal/d), and BMI (kg/m2).Model 4: Adjust for Age (<60, ≥60); Sex (Female, Male); Ehnicity (Han, Mongolian); Marital status (Married, others); Education status (Middle school or below, Middle school above); Annual income (yuan) (<10,000, ≥10,000); Smoking status (Yes, No); Moderate physical inactivity (Light or below, Moderate or vigorous); Family history of diabetes (Yes, No); Family history of hypertension (Yes, No), Energy intake (kcal/d), BMI (kg/m2), and other dietary patterns. Table 2Associations between the different dietary patterns and MetS in different modelsDietary patternsNo. of CasesModel 1Model 2Model 3Model 4OR (95% CI)OR (95% CI)OR (95% CI)OR (95% CI)Meat pattern Tertile 1233/9371.001.001.001.00 Tertile 2274/9371.25 (1.02, 1.53)1.32 (1.07, 1.63)1.34 (1.07, 1.69)1.38 (1.09, 1.75) Tertile 3279/9361.28 (1.05, 1.57)1.45 (1.17, 1.80)1.36 (1.07, 1.72)1.34 (1.04, 1.74) P for trend0.0170.0010.0120.026 Per-SD1.08 (0.995, 1.17)1.12 (1.04, 1.22)1.07 (0.96, 1.19)1.09 (0.98, 1.22)Protein-sweets pattern Tertile 1283/9371.001.001.001.00 Tertile 2266/9360.92 (0.75, 1.12)0.93 (0.76, 1.14)0.97 (0.78, 1.21)0.90 (0.71, 1.12) Tertile 3237/9370.78 (0.64, 0.96)0.80 (0.66, 0.98)0.85 (0.68, 1.07)0.71 (0.56, 0.91) P for trend0.0180.0350.1680.009 Per-SD0.94 (0.86, 1.02)0.95 (0.87, 1.03)0.83 (0.73, 0.95)0.90 (0.77, 1.05)Traditional Chinese pattern Tertile 1217/9371.001.001.001.00 Tertile 2254/9361.24 (1.002, 1.52)1.30 (1.05, 1.60)1.36 (1.09, 1.72)1.33 (1.05, 1.69) Tertile 3315/9371.68 (1.37, 2.06)1.84 (1.49, 2.27)1.75 (1.39, 2.21)1.77 (1.39, 2.26) P for trend<0.001<0.001<0.001<0.001 Per-SD1.24 (1.14, 1.34)1.29 (1.19, 1.40)1.43 (1.21, 1.69)1.29 (1.06, 1.57)Rice and tuber pattern Tertile 1301/9371.001.001.001.00 Tertile 2261/9370.82 (0.67, 0.994)0.83 (0.68, 1.01)0.80 (0.64, 0.991)0.78 (0.63, 0.98) Tertile 3224/9360.67 (0.54, 0.82)0.70 (0.57, 0.87)0.68 (0.54, 0.85)0.66 (0.52, 0.83)P for trend<0.0010.0010.0010.001 Per-SD0.85 (0.78, 0.93)0.87 (0.80, 0.96)0.77 (0.69, 0.87)0.80 (0.70, 0.90)Model 1: Adjust for none. Bolded values indicated P < 0.05.Model 2: Age (<60, ≥60); Sex (Female, Male).Model 3: Adjust for Age (<60, ≥60); Sex (Female, Male); Ehnicity (Han, Mongolian); Marital status (Married, others); Education status (Middle school or below, Middle school above); Annual income (yuan) (<10,000, ≥10,000); Smoking status (Yes, No); Moderate physical inactivity (Light or below, Moderate or vigorous); Family history of diabetes (Yes, No); Family history of hypertension (Yes, No), Energy intake (kcal/d), and BMI (kg/m2).Model 4: Adjust for Age (<60, ≥60); Sex (Female, Male); Ehnicity (Han, Mongolian); Marital status (Married, others); Education status (Middle school or below, Middle school above); Annual income (yuan) (<10,000, ≥10,000); Smoking status (Yes, No); Moderate physical inactivity (Light or below, Moderate or vigorous); Family history of diabetes (Yes, No); Family history of hypertension (Yes, No), Energy intake (kcal/d), BMI (kg/m2), and other dietary patterns. Associations between the different dietary patterns and MetS in different models Model 1: Adjust for none. Bolded values indicated P < 0.05. Model 2: Age (<60, ≥60); Sex (Female, Male). Model 3: Adjust for Age (<60, ≥60); Sex (Female, Male); Ehnicity (Han, Mongolian); Marital status (Married, others); Education status (Middle school or below, Middle school above); Annual income (yuan) (<10,000, ≥10,000); Smoking status (Yes, No); Moderate physical inactivity (Light or below, Moderate or vigorous); Family history of diabetes (Yes, No); Family history of hypertension (Yes, No), Energy intake (kcal/d), and BMI (kg/m2). Model 4: Adjust for Age (<60, ≥60); Sex (Female, Male); Ehnicity (Han, Mongolian); Marital status (Married, others); Education status (Middle school or below, Middle school above); Annual income (yuan) (<10,000, ≥10,000); Smoking status (Yes, No); Moderate physical inactivity (Light or below, Moderate or vigorous); Family history of diabetes (Yes, No); Family history of hypertension (Yes, No), Energy intake (kcal/d), BMI (kg/m2), and other dietary patterns. Differential metabolites associated with dietary patterns After propensity score matching (PSM), 200 patients with MetS and 200 matched controls were included in the metabolomics analysis, with no significant differences in most characteristics (Supplementary Table 7). We conducted an untargeted metabolomic analysis to identify serum metabolites associated with four distinct dietary patterns by comparing subjects with the highest adherence (T3) versus those with the lowest adherence (T1). Multivariate analysis revealed distinct clustering patterns between the T3 and T1 groups. In the unsupervised PCA, a discernible separation trend was observed between the high and low-adherence groups, indicating inherent metabolic differences (Supplementary Fig. 3A, C, E, G). This separation further enhanced in the supervised partial least squares-discriminant analysis (PLS-DA) models, which clearly discriminated between the T3 and T1 groups across all four dietary patterns (Supplementary Fig. 3B, D, F, H). The supervised PLS-DA models showed discrimination among dietary patterns, with R2 values ranging from 0.45 to 0.55. Although the Q2 values remained low (ranging from −0.16 to −0.03), the models were rigorously validated. As shown in the 200-iteration permutation tests (Supplementary Fig. 5), all permuted R2 and Q2 values were lower than the original values (R2 > Q2), and all Q2 regression lines intercepted the Y-axis below zero. The meat pattern exhibited a near-equal distribution of altered metabolites, with 46 significantly up-regulated and 26 down-regulated metabolites (Fig. 1A and Supplementary Table 8). Sensitivity analysis using a stricter quality control (QC) threshold (coefficient of variation [CV] < 25%) identified 61 differential metabolites in the meat pattern (Supplementary Table 19). Among these, 20 key discriminators had Variable importance in projection (VIP) scores ≥2 (Fig. 1I), which comprised 18 up-regulated metabolites, including Phe-Pro (VIP = 3.29), N-Acetyl-L-carnosine (VIP = 3.22), and steroid hormones such as Testosterone. Only two lysophosphatidic acids (LPA 16:1 and LPA 18:3) were downregulated. Notably, Pantethine stood out as a leading biomarker, demonstrating both the highest fold change (Fold change [FC] = 3.76) and a strong VIP score (2.22) (Fig. 1E). Sensitivity analyses using false discovery rate (FDR) correction (FDR < 0.2) yielded results that were highly consistent with the primary analyses across all comparisons, supporting the robustness of our findings. Sensitivity analysis using a stricter FDR < 0.2 threshold identified 32 metabolites for the meat pattern (Supplementary Table 8).Fig. 1Differential metabolomic analysis from dietary patterns.Volcano plots of differential metabolites in A meat pattern, B protein-sweets pattern, C traditional Chinese pattern, and D rice and tuber pattern (Red: Up-regulated; Blue: Down-regulated). Stick plot of top 20 differential metabolites ranked by FC in E meat pattern, F protein-sweets pattern, G traditional Chinese pattern, and H rice and tuber pattern. VIP plot of the metabolites with VIP scores ≥2, highlighting those with the highest separation power between the high and low adherence groups of the I meat pattern, J protein-sweets pattern, K traditional Chinese pattern, and L rice and tuber pattern in the PLS-DA model. M–P Network diagram integrating differential metabolites and their impacted KEGG pathways. Differential metabolomic analysis from dietary patterns. Volcano plots of differential metabolites in A meat pattern, B protein-sweets pattern, C traditional Chinese pattern, and D rice and tuber pattern (Red: Up-regulated; Blue: Down-regulated). Stick plot of top 20 differential metabolites ranked by FC in E meat pattern, F protein-sweets pattern, G traditional Chinese pattern, and H rice and tuber pattern. VIP plot of the metabolites with VIP scores ≥2, highlighting those with the highest separation power between the high and low adherence groups of the I meat pattern, J protein-sweets pattern, K traditional Chinese pattern, and L rice and tuber pattern in the PLS-DA model. M–P Network diagram integrating differential metabolites and their impacted KEGG pathways. In contrast, the protein-sweets pattern showed an inverse association, with only 15 up-regulated and 36 down-regulated metabolites (Fig. 1B and Supplementary Table 9). Sensitivity analysis using a stricter QC threshold (CV < 25%) identified 46 differential metabolites in the protein-sweets pattern (Supplementary Table 20). VIP analysis identified 11 key discriminatory metabolites with VIP scores ≥2 (Fig. 1J), which segregated into two groups: up- and down-regulated. The up-regulated metabolites comprised 6 out of 11, with the most prominent being N-Acetylvaline (VIP = 2.68, FC = 2.12), LPC 17:0 (VIP = 2.67, FC = 1.27), and Pentadecanoic Acid (VIP = 2.39, FC = 1.23). The down-regulated metabolites were represented by 5 compounds, including Sebacic acid (VIP = 2.23, FC = 0.82), N2-Acetyl-L-ornithine (VIP = 2.21, FC = 0.72), and 6-(4-methoxyphenyl) pyrimidine-2,4-diamine (VIP = 2.20, FC = 0.83). The overall magnitude of these changes for the top metabolites is visually summarized in the stick plot (Fig. 1F). A sensitivity analysis with FDR < 0.2 identified 13 metabolites for the protein-sweets pattern (Supplementary Table 9). The traditional Chinese pattern exhibited the most pronounced down-regulatory trend, with only 6 up-regulated metabolites compared to 24 down-regulated ones (Fig. 1C and Supplementary Table 10). Sensitivity analysis using a stricter QC threshold (CV < 25%) identified 27 differential metabolites (Supplementary Table 21) in the traditional Chinese pattern. VIP analysis identified 14 key discriminatory metabolites with VIP scores ≥2 (Fig. 1K), which were predominantly downregulated (10 out of 14). The down-regulated metabolites included 4-Methylphenol (VIP = 3.00, FC = 0.70), 2-(14,15-Epoxyeicosatrienoyl) glycerol (VIP = 2.67, FC = 0.62), and Xanthurenic acid (VIP = 2.61, FC = 0.79). The up-regulated metabolites comprised only 4 compounds, notably ethyl 2-cyano-3-(tetrahydro-3-thiophenylamino) acrylate (VIP = 2.58, FC = 1.95), Isobutyryl carnitine (VIP = 2.25, FC = 1.30), and branched fatty acid esters (FAHFAs). The overall magnitude of these changes is visually summarized in the stick plot (Fig. 1G). A sensitivity analysis with FDR < 0.2 identified 11 metabolites for the traditional Chinese pattern (Supplementary Table 10). Notably, the rice and tuber pattern revealed the strongest up-regulatory signature with 55 up-regulated and 20 down-regulated metabolites (Fig. 1D and Supplementary Table 11). Sensitivity analysis using a stricter QC threshold (CV < 25%) identified 69 differential metabolites in the rice and tuber pattern (Supplementary Table 22). VIP analysis identified 14 key discriminatory metabolites with VIP scores ≥2 (Fig. 1L), which were overwhelmingly up-regulated (12 out of 14). The up-regulated metabolites included Phe-Pro (VIP = 2.88, FC = 1.27), N-Acetyl-L-carnosine (VIP = 2.79, FC = 1.34), Testosterone (VIP = 2.63, FC = 1.38), and trans-4-Hydroxy-L-proline (VIP = 2.45, FC = 1.38), representing a consistent elevation of dipeptides, amino acid derivatives, and steroid hormones. The down-regulated metabolites were represented by only 2 compounds: 3-(5,7-dimethoxy-4-oxo-4H-chromen-2-yl) propanoic acid (VIP = 2.28, FC = 0.70) and LPC 17:0 (VIP = 2.16, FC = 0.79). Sensitivity analysis using FDR < 0.2 identified 40 metabolites for the rice and tuber pattern (Supplementary Table 11). The overall magnitude of these changes is visually summarized in the stick plot (Fig. 1H). Correlation analysis among the top 20 differential metabolites revealed complex co-regulation networks associated with dietary patterns (Supplementary Fig. 7). KEGG pathway enrichment analysis was performed for all four dietary patterns (Fig. 1M–P). Distinct metabolic pathways were identified across patterns. In the meat pattern and traditional Chinese pattern, the most significantly enriched pathway was Oxidative phosphorylation. For the protein-sweets pattern, the differential metabolites were primarily enriched in Sulfur metabolism. In the rice and tuber pattern, Caffeine metabolism showed the highest degree of enrichment. Differential metabolites associated with MetS Multivariate analysis revealed distinct metabolic separation between MetS and non-MetS groups. Both PCA and PLS-DA models demonstrated clear discrimination between the two groups (Supplementary Fig. 4). The PLS-DA model exhibited an R2 of 0.52 and a Q2 of 0.16. Permutation testing (200 iterations) confirmed the model’s validity with R2 > Q2 and a Q2 intercept of -0.30, indicating no over-fitting (Supplementary Fig. 6). Metabolomic analysis revealed extensive metabolic alterations associated with MetS, identifying 36 significantly upregulated and 35 significantly downregulated metabolites in the MetS group compared to the controls (Fig. 2A and Supplementary Table 12). For MetS group comparison, 70 metabolites were identified as differentially expressed under the CV < 25% threshold, of which 70 (98.6%) overlapped with the 71 metabolites identified under the CV < 30% threshold (Supplementary Table 18). VIP analysis identified 20 key discriminatory metabolites with VIP scores ≥2 (Fig. 2B), which segregated into distinct up- and down-regulated groups. The up-regulated metabolites included Monoolein (VIP = 3.62, FC = 1.73), MAG (18:3) (VIP = 3.45, FC = 1.67), 6-Sialyllactose (VIP = 3.05, FC = 1.27), and N-lactoyl-phenylalanine (VIP = 3.05, FC = 1.24), representing elevations in lipid compounds and carbohydrate derivatives. The down-regulated metabolites comprised 6-Deoxy-D-glucose (VIP = 2.98, FC = 0.74), 5-(2,5-dihydroxyhexyl)oxolan-2-one (VIP = 2.93, FC = 0.76), 4-Hydroxyisoleucine (VIP = 2.76, FC = 0.71), and Pipecolic acid (VIP = 2.69, FC = 0.71), indicating suppression of amino acid metabolism and carbohydrate derivatives. Notably, several metabolites demonstrated powerful effects, with Pantethine showing the highest up-regulation (FC = 4.74, VIP = 1.57) and L-threo-3-Phenylserine exhibiting both substantial up-regulation (FC = 4.16) and discriminatory power. The overall magnitude of these metabolic changes is visually summarized in the corresponding stick plot (Fig. 2C). A sensitivity analysis with FDR < 0.2 identified 49 metabolites for MetS (Supplementary Table 12). Correlation analysis among the top 20 differential metabolites revealed complex co-regulation networks associated with MetS (Fig. 2D). KEGG pathway enrichment analysis showed that the Amino sugar and nucleotide sugar metabolism pathway was the most significantly altered in MetS, with the highest enrichment degree, whereas the Bile secretion pathway had the greatest number of differential metabolites (Fig. 2E). The KEGG regulatory network delineates the interconnected metabolic and physiological processes that are dysregulated in MetS (Fig. 2G). Within this network, central metabolic pathways are highlighted in red, identifying them as key hubs of metabolic disruption in the disease state.Fig. 2Comprehensive analysis of differential metabolites in Metabolic Syndrome.A Volcano plot displaying the distribution of all metabolites. Significantly altered metabolites are highlighted in red (up-regulated) and blue (down-regulated). B Metabolites filtered by a VIP score ≥2, indicating high contribution to group separation. C The top 20 most significantly up- and down-regulated metabolites. D Correlation analysis among the top 20 metabolites. E Bubble chart of KEGG pathways by the top 20 differential metabolites. Bubble color and size denote −log10(P-value) and number of metabolites, respectively. F Network diagram integrating differential metabolites and their impacted KEGG pathways. Node legend: Red, KEGG pathway; Green square, Differential metabolite; Yellow, Regulatory enzyme; Light green, Pathway compound; Purple, Molecular module; Blue, Chemical reaction. Comprehensive analysis of differential metabolites in Metabolic Syndrome. A Volcano plot displaying the distribution of all metabolites. Significantly altered metabolites are highlighted in red (up-regulated) and blue (down-regulated). B Metabolites filtered by a VIP score ≥2, indicating high contribution to group separation. C The top 20 most significantly up- and down-regulated metabolites. D Correlation analysis among the top 20 metabolites. E Bubble chart of KEGG pathways by the top 20 differential metabolites. Bubble color and size denote −log10(P-value) and number of metabolites, respectively. F Network diagram integrating differential metabolites and their impacted KEGG pathways. Node legend: Red, KEGG pathway; Green square, Differential metabolite; Yellow, Regulatory enzyme; Light green, Pathway compound; Purple, Molecular module; Blue, Chemical reaction. Metabolic risk score and dietary pattern metabolic signatures associated with MetS To further investigate the clinical relevance of the identified metabolic signatures, we employed LASSO regression to select the most discriminative metabolites for each dietary pattern (Supplementary Table 13 and Supplementary Fig. 8) and for MetS (Supplementary Table 14 and Supplementary Fig. 9). These selected metabolites were subsequently incorporated into logistic regression models to assess their associations with MetS risk. As presented in Table 3, the metabolic signatures derived from dietary patterns showed distinct associations with MetS prevalence. A significant positive association was observed for the meat pattern, with individuals in the highest tertile (T3) of its metabolic signature exhibiting 2.11-fold increased odds of MetS (95% CI: 1.12–3.99; P for trend = 0.027). Similarly, the traditional Chinese pattern signature was associated with elevated MetS odds in T3 (OR = 1.82, 95% CI: 1.02–3.24) groups (P for trend = 0.045). Conversely, the rice and tuber pattern demonstrated a protective association, with a significant 46% reduction in MetS odds in the T3 group (OR = 0.54, 95% CI: 0.30–0.98; P for trend = 0.045). No significant association was found for the protein-sweets pattern. Most strikingly, the composite metabolic risk score displayed a robust association with MetS, showing 23.55-fold increased odds (95% CI: 11.77–47.12) in the highest risk category (P for trend <0.001). Sensitivity analysis using a more stringent metabolite selection criterion (FDR < 0.2) yielded consistent findings, with the metabolic risk score and meat pattern remaining strongly associated with MetS (Supplementary Table 17).Table 3Associations between metabolic signatures of dietary patterns, metabolic risk score, and MetSTertile 1Tertile 2Tertile 3P for trendOR (95% CI)OR (95% CI)OR (95% CI)Metabolic risk score1.002.83 (1.58, 5.07)23.55 (11.77, 47.12)<0.001Metabolic signatures of dietary patterns Meat pattern1.001.63 (0.93, 2.84)2.11 (1.12, 3.99)0.027 Protein-sweets pattern1.001.30 (0.76, 2.23)1.28 (0.74, 2.24)0.349 Traditional Chinese pattern1.001.87 (1.09, 3.22)1.82 (1.02, 3.24)0.045 Rice and tuber pattern1.000.64 (0.37, 1.11)0.54 (0.30, 0.98)0.045 Table 3Associations between metabolic signatures of dietary patterns, metabolic risk score, and MetSTertile 1Tertile 2Tertile 3P for trendOR (95% CI)OR (95% CI)OR (95% CI)Metabolic risk score1.002.83 (1.58, 5.07)23.55 (11.77, 47.12)<0.001Metabolic signatures of dietary patterns Meat pattern1.001.63 (0.93, 2.84)2.11 (1.12, 3.99)0.027 Protein-sweets pattern1.001.30 (0.76, 2.23)1.28 (0.74, 2.24)0.349 Traditional Chinese pattern1.001.87 (1.09, 3.22)1.82 (1.02, 3.24)0.045 Rice and tuber pattern1.000.64 (0.37, 1.11)0.54 (0.30, 0.98)0.045 Associations between metabolic signatures of dietary patterns, metabolic risk score, and MetS Development and validation of MetS risk assessment models To develop robust risk assessment tools for MetS, we employed multiple ML approaches incorporating dietary patterns and metabolite biomarkers. The study cohort was randomly partitioned into training (n = 280) and testing (n = 120) sets, with no significant differences in baseline characteristics between the two groups (Supplementary Table 23). We established models using dietary patterns and their associated metabolite biomarkers (Supplementary Table 24). In the training set, most algorithms demonstrated strong discriminatory power, with Random Forest (RF), K-Nearest Neighbors (KNN), and eXtreme Gradient Boosting (XGBoost) achieving nearly perfect performance (area under the receiver operating characteristic curve [AUC] = 0.999). However, model performance substantially declined in the testing set, with AUC values ranging from 0.640 to 0.709 across different algorithms (Fig. 3A, B).Fig. 3Development, evaluation, and interpretation of MetS risk assessment models based on dietary patterns and metabolites.ROC curves of the model utilizing dietary patterns and their key metabolites in the A training and B testing sets. ROC curves of the model utilizing dietary patterns and MetS metabolic biomarkers in the C training and D testing sets. E–H SHAP plots for interpreting the best-performing model. E Bar chart of the top 15 features. F Beeswarm plot of the top 15 features. G Scatter plots depicting the association between the top two features and the SHAP value. H Force plot for a single participant. Development, evaluation, and interpretation of MetS risk assessment models based on dietary patterns and metabolites. ROC curves of the model utilizing dietary patterns and their key metabolites in the A training and B testing sets. ROC curves of the model utilizing dietary patterns and MetS metabolic biomarkers in the C training and D testing sets. E–H SHAP plots for interpreting the best-performing model. E Bar chart of the top 15 features. F Beeswarm plot of the top 15 features. G Scatter plots depicting the association between the top two features and the SHAP value. H Force plot for a single participant. Subsequently, we enhanced the models by incorporating both dietary patterns and MetS-specific metabolic biomarkers (Supplementary Table 25). This integrated approach yielded significant improvements in model performance. During the training phase, multiple algorithms, including RF XGBoost, and LightGBM) maintained exceptional performance (AUC = 0.999), while Logistic Regression achieved substantial improvement (AUC = 0.892). Most importantly, these models demonstrated markedly better generalizability in the testing set, with AUC values ranging from 0.759 to 0.820 (Fig. 3C, D). The Support Vector Machine (SVM) algorithm showed the most consistent performance across both datasets (training AUC = 0.871, testing AUC = 0.820). To further validate the robustness of our predictive models, we conducted a sensitivity analysis using metabolites selected under the stricter FDR < 0.2 criterion by LASSO (Supplementary Tables 15–16). After incorporating both dietary patterns and these MetS-specific metabolic biomarkers, logistic regression and SVM continued to demonstrate the highest AUC (Supplementary Tables 26–27 and Supplementary Fig. 12). The Shapley additive explanations (SHAP) bar plot and beeswarm plot (Fig. 3E, F) identified the five most influential features in the SVM model: 2-[6-(1H-benzimidazol-2-yl)-2-pyridyl]-1H-benzimidazole, PC (3:0/16:2), 6-chloro-3-[(dimethylamino)methylidene]chroman-4-one, coniferyl alcohol, and lysine butyrate. Furthermore, the SHAP dependence plot (Fig. 3G) compared the associations between actual values and SHAP values for the top two features. Participants with higher values of both 2-[6-(1H-benzimidazol-2-yl)-2-pyridyl]-1H-benzimidazole and PC (3:0/16:2) exhibited positive SHAP values, indicating a higher risk of the “MetS” classification. The SHAP force plot (Fig. 3H) visualized the cumulative contribution of each feature to the risk assessment for a single participant (the 10th case in the dataset), with the specific value and corresponding SHAP value for each feature indicating its respective positive or negative impact on the final risk assessment. Discussion This study systematically investigated the complex associations between dietary patterns, circulating metabolites, and MetS risk in a rural Chinese population. Our analysis identified four distinct dietary patterns-meat, protein-sweets, traditional Chinese, and rice and tuber-that demonstrated significant associations with MetS prevalence. More importantly, we objectively validated three dietary exposures (meat, traditional Chinese, and rice and tuber) their unique serum metabolic signatures, providing molecular evidence for their biological plausibility in MetS pathogenesis. The construction of a metabolomic risk score revealed profound systemic metabolic disturbances underlying MetS, offering enhanced capability for risk stratification beyond traditional factors. Furthermore, our ML approach demonstrated that integrating dietary patterns with MetS-specific metabolic biomarkers significantly improved risk assessment accuracy compared with models using dietary patterns and their metabolites. These findings establish a critical connection between external dietary exposures and internal metabolic responses, advancing our understanding of diet-metabolome interactions in MetS development. Our study identified four distinct dietary patterns with varying associations with MetS risk. The meat pattern, characterized by high intake of fish, poultry, organ meats, and red meat, was significantly associated with increased MetS risk, consistent with multiple domestic and international studies27,28. This pattern shares similarities with Western dietary patterns rich in red and processed meats, which systematic reviews have linked to an increased prevalence of MetS29. Conversely, the Protein-sweets pattern demonstrated a protective association with MetS. This may be attributed to nutrient-rich components such as fresh milk, dairy products, nuts, and eggs, which provide high-quality protein, unsaturated fats, and essential micronutrients. For example, multiple studies consistently indicate that milk and dairy consumption are associated with a reduced risk of MetS, attributable to components such as calcium, magnesium, and whey protein, which may improve lipid profiles, lower blood pressure, and enhance insulin sensitivity30,31. Furthermore, eggs, which are rich in choline and phospholipids32, and nuts, which are rich in unsaturated fats33, dietary fiber, and antioxidants, have also been shown to lower MetS risk, potentially through mechanisms that improve lipid metabolism and reduce inflammatory responses. The traditional Chinese pattern, although containing vegetables and legumes, was positively associated with MetS. This may be attributed to a high consumption of refined wheat flour, pork, and alcohol. A study among Chinese adults linked high wheat intake to increased MetS risk and elevated BMI34. In addition, a Mendelian randomization study confirmed a causal role of alcohol intake in the development of MetS35. Although vegetables and soy products may offer metabolic benefits, these advantages appear offset by the adverse effects of refined grains and pork in this pattern. In contrast, the rice and tuber pattern exhibited significant protective effects. Its major components (rice and tubers) provide complex carbohydrates, dietary fiber, vitamins, and minerals. Other Chinese studies have reported similar benefits from a “green water pattern” rich in rice and vegetables36. Rice protein has been shown to modulate lipid metabolism and reduce triglycerides, while tubers exhibit anti-diabetic and anti-inflammatory properties37,38. These components may collectively contribute to improved metabolic health. While numerous studies have established physiologically plausible associations between single food items or dietary patterns and MetS39, limitations inherent in dietary assessment tools, such as recall bias and intercorrelations among food groups, coupled with interindividual variations in metabolic health, underscore the need for enhanced rigor in nutritional research on MetS40. For the meat-based pattern in our study, metabolites such as Phe-Pro, N-Acetyl-L-carnosine, and steroid hormones were significantly up-regulated, whereas lysophosphatidic acids were down-regulated. These findings are consistent with previous reports linking processed meat intake to elevated levels of phenylalanine and proline41. In addition, phospholipids such as LPE 20:3 were reduced in the meat pattern, aligning with studies that have shown high-fat diets led to decreased LPE levels in mice42. However, conflicting results exist regarding the direction of LPE changes under different dietary condition43. These inconsistencies highlight that LPEs as dietary biomarkers still require further validation in future studies to clarify their specific roles and responses to dietary intake. In the protein-sweets pattern, phospholipids LPC 17:0 and LPC 12:0 was up-regulated. Experimental studies have shown that LPC 17:0 activates GLP-1 and promotes insulin secretion, thereby improving glucose metabolism and insulin resistance44. For the traditional Chinese pattern, Isobutyryl carnitine was upregulated and showed a positive association with MetS. Although this metabolite has not been widely studied in dietary contexts, elevated levels have been linked to stroke and heart failure45,46, aligning with the adverse metabolic profile of this pattern. Moreover, FAHFA (22:4/18:0), which was elevated in this pattern, has been reported to be higher in omnivores than in vegetarians or vegans47. The rice and tuber pattern was characterized by elevated levels of trans-4-Hydroxy-L-proline, which is abundant in rice and has been associated with lower cardiometabolic risk in Asian Indians with prediabetes48. We identified a panel of nutrient-related metabolites that reflect the composition of self-reported dietary patterns and interact with metabolic pathways broadly implicated in MetS. This approach aligns with emerging international efforts to use molecular measurements as biomarkers of dietary exposure and chronic disease risk. Several studies have linked metabolites to a priori dietary patterns such as the Healthy Eating Index-2015 (HEI-2015) and the Mediterranean diet and have further evaluated the association of these metabolic signatures with cardiovascular and cardiometabolic health18,49. Additionally, a UK study developed a Mediterranean diet metabolic score using targeted metabolomics and found it to be inversely associated with dyslipidemia and insulin resistance50. To our knowledge, this study is the first in China to construct dietary pattern-specific metabolic signatures in a rural population and systematically examine their associations with MetS. Using an objective metabolomic approach, we validated that the metabolic signatures of the meat-based, traditional Chinese, and rice and tuber patterns remained strongly associated with MetS. This study identified 35 key metabolites significantly associated with MetS through comprehensive metabolomic profiling. Among these, Monoolein and MAG (18:3), both glycerol esters, emerged as crucial biomarkers. These findings align with randomized trials indicating that glycerol esters containing saturated and monounsaturated fatty acids are closely associated with clinical glycemic parameters, and that diacylglycerols and triacylglycerols can predict changes in hemoglobin, insulin, and homeostasis model assessment for insulin resistance (HOMA-IR)51. Protective metabolites identified in our study included 4-Hydroxyisoleucine, which has been confirmed in animal studies to effectively alleviate chronic inflammation and related MetS issues induced by high-fat diets52. These findings were incorporated into a metabolic risk score that showed significant positive correlations with MetS and its components, particularly hyperglycemia and dyslipidemia. This approach is supported by international studies demonstrating that comprehensive metabolic risk scores can better describe disease metabolic status than single metabolites53. We developed a MetS risk assessment model that integrates dietary patterns with key metabolic biomarkers, demonstrating superior predictive performance compared to models based solely on dietary patterns. While dietary patterns have been incorporated into chronic disease risk assessment54,55, their standalone predictive capacity remains limited for complex conditions like MetS. The addition of metabolomic biomarkers enables more comprehensive evaluation of metabolic heterogeneity and reveals how dietary components influence health through specific pathways. This approach is supported by growing evidence for metabolomics in disease prediction. A UK cohort demonstrated that metabolomic data significantly improved the 10-year prediction of 15 disease outcomes compared to traditional predictors56. At the same time, Loic et al. reported a 4.5% enhancement in diabetes prediction accuracy beyond clinical biomarkers57. Among the tested algorithms, SVM demonstrated the most consistent performance across training and testing datasets, slightly outperforming logistic regression. This aligns with previous applications of SVM to metabolomic data for disease prediction58, with its strength in handling high-dimensional data being particularly valuable for capturing complex metabolite-MetS associations. This study combines dietary patterns with metabolomic profiling to identify objective biomarkers and construct interpretable ML models for MetS risk assessment, with a specific focus on a Chinese rural population that has been underrepresented in previous research. Several limitations should be acknowledged. Due to the cross-sectional design of this study, causal inference between dietary patterns, metabolites, and MetS cannot be established. In particular, the temporal sequence and directionality of these associations remain unclear. Future prospective and longitudinal studies are warranted to validate these findings and clarify potential causal relationships. The sample was drawn from a specific rural population in Liaoning Province, which may limit generalizability to other regions or urban populations, highlighting the need for validation in more diverse populations. The FFQ used in this study has not been formally validated against established dietary assessment methods, such as repeated 24-h dietary recalls or dietary biomarkers. This may introduce measurement error and potential misclassification of dietary intake. Although the FFQ was developed based on commonly used food group classifications in Chinese nutritional epidemiology studies59–62, the lack of validation in the target population should be considered when interpreting the results. Future studies are warranted to formally validate dietary assessment tools in this population. Furthermore, as this study employed an untargeted LC-MS approach, the identified metabolic features should be considered putatively annotated rather than definitively identified. Although most metabolites reached MSI Level 2, these annotations remain uncertain. Therefore, further validation using authentic reference standards, targeted metabolomics approaches, and independent cohorts is required before these metabolites can be reliably interpreted with respect to their biological roles or potential clinical applications. Although we adjusted for numerous potential confounders, residual confounding from unmeasured factors such as genetic background cannot be excluded, suggesting the value of incorporating genomic data in future studies. Finally, while the metabolomic approach provides comprehensive coverage, some potentially relevant metabolites may remain undetected due to methodological constraints. This study provides a comprehensive investigation into the association between diet and MetS by integrating dietary patterns with deep metabolomic phenotyping. We not only identified the “meat”, “traditional Chinese”, and “rice and tuber” dietary patterns, associated with MetS risk in a rural Chinese population but also objectively validated these associations through their unique serum metabolomic signatures. The construction of a metabolomic risk score further captured the metabolic disturbances associated with MetS, offering a potential molecular tool for risk stratification. Crucially, integrating dietary patterns with MetS-specific metabolomic biomarkers using SVM improved risk assessment accuracy compared with conventional approaches. This multimodal strategy bridges the gap between dietary exposure and the internal metabolic response, offering insights for future research toward personalized risk assessment and targeted interventions to prevent MetS. Methods Study population This cross-sectional study u sed a random-cluster sampling method to select participants from 33 natural villages across 4 towns in Liaoning Province, China, in March 2021. All participants provided written informed consent before data collection, which included questionnaires, physical examinations, and blood sample collection. Inclusion criteria were: (1) age ≥35 years; (2) residents living in the area for ≥5 years; and (3) voluntary participation with signed informed consent. Exclusion criteria included: (1) physical disability affecting mobility; (2) inability to communicate normally; and (3) pregnancy. The study was approved by China Medical University’s ethics committee (Approval No. [2018] 083). Dietary assessment Dietary intake was assessed using a semi-quantitative food frequency questionnaire (FFQ) comprising 24 food groups. Dietary information was derived from a semi-quantitative FFQ63, including 24 food items listed in Supplementary Table 1. As part of the FFQ, nine frequency categories were included, ranging from “almost never eat” to “three times or more per day”. In addition, in the FFQ, participants reported how often they ate each food item over the course of a year and how much of each they consumed. The daily intake for each food group was calculated, and energy intake was estimated using the Chinese Food Composition64. Dietary patterns were derived using PCA with varimax rotation. The suitability of data for factor analysis was confirmed by a Kaiser–Meyer–Olkin (KMO) value ≥ 0.5 and a significant Bartlett’s test of sphericity (P < 0.05). Factors were retained based on eigenvalues > 1.25, scree plot inspection, and interpretability. Food groups with factor loadings ≥|0.20| were considered significant contributors to a pattern. Factor scores for each dietary pattern were calculated, and to assess adherence, participants were categorized into tertiles (T1-T3) based on these scores, with T3 representing the highest adherence. Metabolomic profiling For metabolomic profiling, a case-control design was adopted. To enhance comparability between MetS cases and healthy controls, PSM was performed using IBM SPSS Statistics 21.0. Propensity scores were estimated using a logistic regression model including the following covariates: age, sex, ethnicity, education status, marital status, annual income, physical activity, smoking status, and family history of hypertension and diabetes. Participants were matched at a 1:1 ratio using the nearest neighbor algorithm with a caliper of 0.01. Following matching, simple random sampling was conducted among the matched pairs to select participants for non-targeted metabolomic analysis. Specifically, 200 matched pairs (200 MetS cases and 200 matched controls) were randomly selected using SPSS’s random number generator, ensuring that each matched pair had an equal probability of inclusion. Serum samples were stored at −80 °C until analysis. For metabolite extraction, 100 μL of each serum sample was placed in an EP tube and mixed with 400 μL of 80% methanol. The mixture was vortexed and incubated on ice for 5 min to facilitate protein precipitation, then centrifuged at 15,000 × g and 4 °C for 20 min. Subsequently, an aliquot of the supernatant was diluted with mass spectrometry-grade water to a final methanol concentration of 53%. The solution was then subjected to a second centrifugation at 15,000 × g and 4 °C for 20 min. Finally, the resulting supernatant was collected and injected into the LC-MS system for analysis65–67. Chromatographic separation was conducted on a Hypesil Gold column (C18, 100 × 2.1 mm, 1.9 μm) with a column temperature of 40 °C and a flow rate of 0.2 mL/min. In positive mode, mobile phase A was 0.1% formic acid, and mobile phase B was methanol. In negative mode, mobile phase A was 5 mM ammonium acetate (pH 9.0), and mobile phase B was methanol. The 12-min gradient elution program was: 0–1.5 min, 2% B; 1.5–3 min, 2–85% B; 3–10 min, 85–100% B; 10–10.1 min, 100–2% B; and 10.1–12 min, 2% B. Mass spectrometry was performed using an ESI source with a scan range of m/z 100–1500. The MS parameters were set as follows: spray voltage, 3.5 kV; sheath gas flow rate, 35 psi; auxiliary gas flow rate, 10 L/min; capillary temperature, 320 °C; S-lens RF level, 60; and auxiliary gas heater temperature, 350 °C. Data were acquired in both positive and negative polarities using data-dependent MS/MS scans (DDA). Raw data were processed using CD3.1 software for peak alignment, extraction, and quantification. Initially, peaks were aligned across samples based on retention time and mass tolerance (ppm). Peak extraction and quantification were then executed using predefined thresholds for S/N ratio, adduction types, and mass accuracy (ppm). The parameters included a retention time tolerance of 0.2 min, a mass tolerance of 5 ppm, a signal-to-noise ratio of 3, and a signal intensity deviation of 30%. Background ions were removed using blank samples. Peak areas were normalized using the formula: Relative peak area = Original peak area/(Total sample peak area/Total QC1 peak area). A total of 1328 metabolites were identified and retained for subsequent statistical analysis. According to the Metabolomics Standards Initiative (MSI) guidelines68, metabolites were categorized into different identification levels based on annotation confidence. Level 2 identification (putatively annotated compounds based on MS/MS spectral matching) was achieved using the mzCloud database (https://www.mzcloud.org/), a high-resolution accurate mass spectral library acquired from authentic standards on Orbitrap platforms, and the mzVault database (Thermo Fisher Scientific), a local spectral library containing MS/MS data optimized for Orbitrap instruments (including approximately 2000 common metabolites and nearly 100,000 lipids). MS/MS spectral matching was performed with a similarity score threshold of ≥40 for mzCloud and ≥50 for mzVault, with precursor ion matching conducted within a mass tolerance of 5 ppm. Level 3 identification (putatively characterized compound classes based on exact mass and formula prediction) was assigned using an in-house MassList database, constructed from authentic standards and purchased ion pairs (sourced from the Laboratory in Dalian Institute of Chemical Physics), covering endogenous metabolite classes including amino acids and derivatives, nucleotides and derivatives, bile acids, oxylipins, hormones, organic acids, and lipids. MassList matching was performed based on precursor ion accurate mass with a mass deviation threshold of 5 ppm. Chemical classification for all identified metabolites was assigned using the ClassyFire tool based on the ChemOnt structural taxonomy, with compounds categorized into four hierarchical levels. Detailed information for all 1328 metabolites, including their MSI levels, is provided in Supplementary Data S1. Quality control To ensure the stability of the analytical system and the reliability of the data, QC samples were prepared by pooling equal aliquots (100 μL) from each experimental serum sample. They injected every 10 samples throughout the run. Analytical stability was assessed by calculating Pearson correlation coefficients among QC samples, which were all greater than 0.95, indicating excellent system robustness (Supplementary Figs. 1, 2). Furthermore, the precision of individual metabolites was evaluated using the CV in the QC samples. In the primary analysis, metabolites with a coefficient of variation >30% in QC samples were excluded, retaining 1328 metabolites for further analysis (Supplementary Data S1). This threshold has been adopted in recent metabolomic studies to balance data quality and metabolite coverage69,70. Sensitivity analyses using a stricter QC threshold (CV < 25%) yielded highly consistent results across all analyses, with the majority of identified metabolites retained (Supplementary Data S2). Definition of MetS According to the Standards of medical care for type 2 diabetes of CDS1, (1) Abdominal obesity: waist circumstance (WC) ≥ 90 cm among the male or ≥85 cm among the female; (2) Hyperglycemia: fasting blood glucose (FBG) was greater than or equal to 6.1 mmol/L and/or diagnosed with diabetes; (3) Hypertension: an individual with hypertension has systolic blood pressure (SBP) over 130 mmHg and/or diastolic blood pressure (DBP) over 85 mmHg, as well as being diagnosed with hypertension; (4) Dyslipidemia: triglyceride (TG) was greater than or equal to 1.70 mmol/L and/or high density lipoprotein (HDL) was less than 1.04 mmol/L. According to MetS, three or more of the above abnormalities must be present. Covariates Face-to-face interviews collected information about participants, including sex, age, ethnicity, marital status, educational status, annual income, smoking status, moderate physical activity, family history of diabetes, family history of hypertension, and energy intake. Age was collected as a continuous variable and divided into two groups (<60 years or >60 years). We categorized the ethnicity into two groups (Han or Mongolian) based on the study’s purpose. Two groups were classified according to their educational status (Middle school or below versus others). Income was the annual household income from all sources reported by the participants and was categorized into two groups (<10,000 yuan or ≥10,000 yuan). Body mass index (BMI) is calculated as weight (kg)/height (m²) and categorized according to Chinese criteria into normal (<24 kg/m2) and overweight or obesity (≥24 kg/m2)71. Smoking statuses were defined as those who had smoking habits. Participants reported engaging in weekly moderate-intensity physical activity, such as brisk walking, running, cycling, stair climbing, and aerobics. Participants were categorized into two groups based on the World Health Organization (WHO) recommendation of ≥150 min of moderate-intensity activity per week: light activity (<150 min/week) and moderate-to-vigorous activity (≥150 min/week)72. A yes/no code was assigned to the family history of diabetes or hypertension. Statistical analysis The baseline characteristics of the study population were summarized according to MetS status. Normality of continuous variables (BMI, WC, FBG, TG, HDL, SBP, DBP, and energy intake) was assessed using the Kolmogorov-Smirnov and Shapiro-Wilk tests. All variables showed significant deviation from normal distribution (all P < 0.001). Therefore, continuous variables were presented as median (interquartile range [IQR]) and compared between groups using the Mann-Whitney U test. Categorical variables were expressed as frequencies and percentages, and compared using the Chi-square test. Dietary patterns were derived from the 24 food groups using PCA. The associations between dietary patterns and MetS were examined using multivariable logistic regression models, with the T1 of each pattern serving as the reference group. Additionally, dietary pattern scores were also analyzed as continuous variables (per 1 SD increase). These associations were assessed through a series of adjusted models: Model 1 was crude; Model 2 adjusted for age and sex; Model 3 further adjusted for ethnicity, marital status, education, income, smoking, physical activity, family history of diabetes and hypertension, total energy intake, and BMI; and Model 4 additionally adjusted for the scores of the other dietary patterns. A linear trend test was performed by treating the dietary pattern tertiles as continuous variables. Meanwhile, we also performed stratified analyses of variables, such as sex, age, and BMI, as well as the interactions between stratified covariates and dietary patterns. For the metabolomics analysis, data was preprocessed with internal standard normalization. Unsupervised PCA was first performed to assess the overall metabolic differences among groups. PLS-DA, a supervised discriminant statistical method, was employed to model the association between metabolite expression levels and dietary and MetS status. The model’s performance was evaluated using R2 (explained variance) and Q2 (predictive ability) obtained through 7-fold cross-validation. Higher R² and Q² values indicate better model fit and predictive performance, respectively. To further validate the model and assess the risk of over-fitting, a permutation test with 200 iterations was conducted by randomly assigning group labels. The model was considered robust and not over-fitted when the permuted R2 values were consistently higher than the Q2 values, and the intercept of the Q2 regression line on the Y-axis was negative. Differential abundance analysis between comparison groups was performed by combining results from PLS-DA and univariate Student’s t-tests. Metabolites were considered significantly altered if they had a variable importance in projection score greater than 1.0, a P-value less than 0.05, and an absolute fold change greater than 1.2. To further assess the robustness of our findings, we conducted a sensitivity analysis using a more stringent criterion of Benjamini-Hochberg false discovery rate (FDR) < 0.2. Pearson’s correlation analysis was performed to evaluate relationships among differential metabolites. The coefficients (r) were calculated for all pairwise combinations, with |r| approaching 1 indicating a strong linear association. Pathway enrichment analysis of these differential metabolites was conducted using the KEGG database and a hypergeometric test. To construct metabolomic risk scores, LASSO regression was employed. For the MetS metabolomic risk score, LASSO logistic regression with 10-fold cross-validation was employed to select a parsimonious set of metabolites for assessing the risk of MetS. The risk score was calculated as the linear combination of the standardized levels of the selected metabolites, weighted by their LASSO regression coefficients. A similar procedure using LASSO linear regression was applied to derive dietary pattern metabolomic signature scores using the continuous dietary pattern factor scores as the outcome. The associations between MetS metabolomic risk score, metabolomic signatures of dietary patterns, and MetS were still examined using logistic regression models. To develop robust risk assessment models, the dataset was randomly split into a training set (70%) and an internal test set (30%). Multiple ML algorithms, including Logistic Regression73, RF74, KNN75, SVM76, XGBoost77, and LightGBM78 were trained and optimized via hyperparameter tuning with cross-validation on the training set. The performance of the final models was evaluated on the held-out test set using the area under the receiver operating characteristic curve (AUC) as the primary metric, with accuracy, sensitivity, and specificity as secondary metrics. To enhance the interpretability of the complex models, SHAP79,80 framework was applied. SHAP is an additive feature attribution method rooted in coalitional game theory that quantifies each feature’s contribution to the individual risk assessment by computing its Shapley value. Specifically, the Shapley value represents the average marginal contribution of a feature across all possible feature combinations, providing a theoretically consistent measure of feature importance. This approach assigns each feature a value representing the change in the expected model output when that feature is observed versus when it is missing, providing both global insights into feature importance and local explanations for individual predictions. In contrast to traditional variable importance metrics, which rank features solely by overall impact, SHAP provides a more nuanced interpretation by indicating whether each feature has a positive or negative relationship with the target outcome. All statistical analyses were conducted using IBM SPSS Statistics 21.0 (IBM, ASiaAnalytics Shanghai), Python 3.5.0 and R 4.0.3, with a two-sided P-value of less than 0.05 considered statistically significant. Supplementary information Supplementary Materials Supplementary Data. Strobe checklist. Supplementary Materials Supplementary Data. Strobe checklist.