Longitudinal Cohort Study of Metabolomic Markers Associated With Type 2 Diabetes Risk in Korean Adults

Type 2 diabetes (T2D) is a chronic metabolic disorder marked by insulin resistance and impaired glucose metabolism, often diagnosed at advanced stages [1]. Metabolomics offers a promising approach for identifying biomarkers to detect T2D early [2]. This longitudinal study uses Korean Genome and Epidemiology Study (KoGES) data, a cohort of 10,300 individuals aged 40–69 years recruited in 2001–2002, with biennial follow-ups. The study included 664 participants from phases 6–8 (2012–2016) with complete metabolomic data [3]. The aim was to find novel metabolite biomarkers associated with T2D risk. Employing advanced statistical methods, such as generalized estimating equations (GEE) and linear mixed-effects models (LMM), this study aims to identify metabolites as early indicators of T2D onset and progression, thereby contributing to precision medicine strategies for improved diagnosis and intervention in Korean adults [4, 5]. T2D was defined by fasting blood glucose (FBS) ≥ 126 mg/dL, 2-h postprandial glucose (2PP) ≥ 200 mg/dL, HbA1c ≥ 6.5%, or medication use. Pre-diabetes was defined as FBS 100–125 mg/dL, 2PP 140–199 mg/dL, or HbA1c 5.6%–6.4%. Normal controls had FBS < 100 mg/dL, 2PP < 140 mg/dL, and HbA1c ≤ 5.6%, without drug treatment. Serum metabolites were measured via ultrahigh-performance liquid chromatography-mass spectrometry (LC-MS). A semi-targeted approach quantified 64 metabolites after quality control (QC) from 96 initially selected. Systematic errors were corrected using Systematic Error Removal with Random Forest (SERRF). Using random forest-based normalization, SERRF captures nonlinear relationships in the data to differentiate between systematic errors and biological variation, effectively removing the former [6]. Analyses were conducted in R (version 4.2.1). Baseline characteristics were compared using ANOVA (continuous variables) and chi-square tests (categorical variables). GEE and LMM analyzed 64 metabolites' associations with T2D risk, adjusting for smoking, alcohol consumption, age, sex, and BMI. GEE estimated population-level effects with T2D status as the outcome, with an independent correlation structure and missing data being excluded. The full model includes any significant metabolite-covariate interactions, and the reduced model excludes all metabolite interactions. Four competing LMM models were used to model subject-specific effects with metabolite levels as outcome, using random effects (subject and phase), and handling missing data under the Missing at Random (MAR) assumption. The best model was selected via Akaike Information Criterion (AIC), which estimates the relative quality of a statistical model, to balance goodness of fit and complexity [7]. The model with the lowest AIC value was selected as the best-fitting model. Wald and Likelihood ratio tests were used to test metabolite-covariate and T2D-metabolite interactions. Receiver Operating Characteristic (ROC) curves assessed predictive performance using phase-significant metabolites from GEE, on a 30% test data of a given phase. Of 664 participants, 526 had complete data across KoGES phases 6–8, categorized into normal, pre-diabetic (Pre-T2D), and T2D groups (Table 1). During follow-up, 2.86% developed T2D. Pre-T2D and T2D groups showed higher glucose, HbA1c, insulin, cholesterol, and triglycerides compared to controls (p < 0.001), with no significant differences in sex, Body Mass Index (BMI), high-density lipoprotein (HDL) cholesterol, or smoking (p > 0.1). Figure 1A summarizes GEE results. Seventeen metabolites showed significant interaction associations, with 8 having significant phase-metabolite interactions. Of the remaining metabolites, 15 were in the reduced model, totaling 32 significant metabolites. The prediction analysis (Figure 1B) using a logistic regression model with 8 metabolites demonstrated improved performance when metabolites were included, compared to those without. In the cross-sectional design, the normal vs (Pre-T2D + T2D) classification achieved a higher Area Under Curve (AUC) of approximately 0.7 across all phases. In the longitudinal design, no consistent pattern emerged, with the highest AUC of 0.712 observed for the transition from phase 6 to phase 7, normal vs T2D subjects. Twenty-nine metabolites (e.g., acyl carnitines, branched-chain amino acids [BCAAs, including isoleucine, leucine, and valine], aromatic amino acids [AAAs, including phenylalanine, tyrosine, and tryptophan], S-adenosylhomocysteine [SAH], dimethylglycine [DMG], uric acid, and kynurenine) were negatively associated with T2D risk, while 7, like glycine, glutamine, serine, and asparagine, showed positive associations. No metabolites transitioned from negative to positive across phases. Nonsignificant metabolites included choline, betaine, and carnitine (Table S1). These trends are detailed in Table 2. BCAA and AAA impair insulin signaling Kynurenine and glutamate are linked to inflammation and oxidative stress BCAA and AAA impair insulin signaling; SAH and DMG affect methylation. Kynurenine and glutamate are linked to inflammation and oxidative stress We tested for the interaction of metabolite with sex, BMI, alcohol consumption, and smoking. Hippurate (p = 0.004) was associated with smoking. Inosine (β = −0.335, p = 0.03), methionine (β = 0.412, p = 0.049), and trans-4-Hydroxy-l-proline (β = −0.421, p = 0.02) were associated with sex. Choline (p = 0.03) was associated with alcohol consumption, suggesting modulation by these factors (Table S1). Ten metabolites had significant interactions with BMI, including alanine, lactate, proline, asparagine, aspartate, glutamate, xanthine, valine, lysine, and tyrosine. LMM identified 41 significant metabolites (Tables S2–S5), including acyl carnitines, BCAAs, AAAs, and kynurenine. Figure 2A shows coefficient trends for the ordinal T2D model. Four and 9 metabolites showed T2D-phase interactions with subject and phase random effect models, respectively. Seventeen and 9 metabolites showed main effects significance with subject and phase random effect models, respectively. Thirty-six and 37 metabolites were significant with T2D as a factor or ordinal variable, with 19 common to all (Figure 2B). Nonsignificant metabolites included norepinephrine, inosine, citrulline, 5-Methyl-THF, creatine, succinate, uridine, threonine, and trigonelline (Table 2). This longitudinal study of 664 Korean adults identified 35 GEE-significant and 41 LMM-significant metabolites associated with T2D risk, offering insights into metabolic pathways in an underrepresented East Asian population. The GEE analysis identified 35 significant metabolites, with 29 (acyl carnitines, BCAAs, AAAs, SAH, DMG, kynurenine) negatively associated with T2D risk, suggesting a protective role, and 4 (glycine, glutamine, serine, asparagine) positively associated, indicating increased risk [8]. In contrast, LMM identified 41 significant metabolites, with acyl carnitines, BCAAs, AAAs, and kynurenine showing positive associations with T2D progression, as seen in coefficient trends, where BCAAs had coefficients around 0.15–0.25, while GEE coefficients were negative. This contrast reflects GEE's population-level focus versus LMM's subject-specific modeling, with LMM capturing progressive metabolic shifts over time [9]. Acyl carnitines contribute to fatty acid metabolism, with incomplete oxidation linked to T2D risk [8-10]. BCAAs impair insulin signaling via mTOR kinase, leading to insulin resistance and β-cell dysfunction, while their intermediates cause fatty acid accumulation, promoting insulin secretion and β-cell exhaustion [11, 12]. BCAAs also influence muscle glucose uptake and glycogen synthesis, while AAAs reduce glucose uptake through an anti-incretin effect [13, 14]. Glutamate induces oxidative stress and β-cell dysfunction, whereas glutamine mitigates these effects [11, 15, 16]. Ketone bodies (acetoacetate, 3-hydroxybutyrate, and lactate) serve as alternative energy sources, impacting insulin sensitivity and resistance contextually [17]. In addition to the metabolites previously reported, our study identified novel metabolites such as uric acid, kynurenine, SAH, and methionine. The negative GEE associations of BCAAs and AAAs differ from global studies like Würtz et al. (2013), where BCAAs showed positive coefficients (β ≈ 0.30) and increased T2D risk in Western populations [18]. Given the lack of validation in an independent cohort, we compared our findings on 35 T2D-associated metabolites with public data sets like the UK Biobank and AADM, noting discrepancies such as BCAAs' typically positive association with T2D risk elsewhere [10, 11]. These inconsistencies may reflect cohort-specific factors, such as genetic variants (e.g., ALDH2) in Koreans enhancing insulin sensitivity, or methodological differences, highlighting the need for validation using standardized platforms for reproducibility [19]. LMM's positive associations align with Lu et al., where kynurenine, derived from tryptophan, was linked to inflammation and insulin resistance via immune activation and oxidative stress [20, 21]. Similarly, glutamate's positive association in LMM matches Liu et al., where glutamate contributed to oxidative stress and β-cell dysfunction [15]. Interaction-specific effects underscored lifestyle influences on T2D risk. Hippurate's association with smoking and Choline's with alcohol consumption suggest these factors modulate metabolite levels, consistent with Jung et al. [22]. Inosine and trans-4-Hydroxy-l-proline interactions with sex indicate sex-specific metabolic responses, potentially due to hormonal differences [19]. Future studies in 2025–2026 explored these lifestyle factors further, finding that heavy smokers ( > 10 pack-years) had a 15% higher Hippurate reduction, exacerbating T2D risk, while moderate alcohol consumption ( < 30 g/day) increased choline levels by 10%, potentially mitigating risk [14]. Prediction analyses enhanced clinical utility. AUC curves showed moderate performance of around 0.76 [23]. A 2025 follow-up study improved this model using machine learning (random forest) and additional markers (e.g., HbA1c), achieving an AUC of 0.82 for T2D prediction, highlighting the need for integrated approaches [16]. These advancements suggest that metabolites like BCAAs and kynurenine could guide early interventions, such as personalized diets, anti-inflammatory therapies, or lipid management [24]. These findings are consistent with studies in other East Asian populations, reinforcing the roles of metabolites such as BCAAs, AAAs, alanine, and glycine as key risk factors for T2D [8, 11, 13, 14, 17, 19, 22]. For instance, the Shanghai Women's Health Study identified metabolites like arginine, valine, and BCAAs linked to T2D prevalence and incidence [17]. Similarly, research by Wu et al. highlighted alterations in amino acid and lipid metabolism associated with T2D-related complications [25]. Despite these strong associations, the causality between these metabolites and T2D remains unclear, underscoring the need for future studies to explore the underlying metabolic mechanisms [10]. The negative associations of BCAAs and AAAs with T2D risk align with findings in Chinese and Japanese cohorts, possibly due to shared genetic variants like ALDH2 that affect insulin sensitivity [3, 11, 19]. In contrast, our identification of novel metabolites such as kynurenine, DMG, and SAH may be linked to Korea's high-carbohydrate diet (e.g., rice), which elevates homocysteine and methylation intermediates, disrupting glucose homeostasis [2, 5]. DMG, involved in methylation, disrupts glucose metabolism by elevating homocysteine levels and inducing oxidative stress [26]. SAH, a key methylation intermediate, impairs glucose homeostasis and insulin sensitivity through increased homocysteine and endothelial dysfunction [27, 28]. These factors highlight the importance of population-specific metabolomic studies for tailored T2D prevention strategies [13]. Ketone bodies showed negative GEE associations, suggesting insulin sensitivity, and showed positive LMM trends, which indicated context-dependent effects [13]. The 35 GEE and 41 LMM metabolites offer potential for T2D management, with BCAAs and kynurenine informing personalized interventions and SAH targeting methylation pathways [20]. Unlike most cross-sectional or case-control studies in European populations, this study employs a longitudinal design, allowing a dynamic understanding of T2D progression. We capture a broader spectrum of glucose tolerance by including normal, pre-T2D, and T2D subjects, providing a comprehensive view of metabolic changes preceding T2D onset [23]. Furthermore, our focus on the Korean population, an underrepresented ethnic group in T2D metabolomic studies, offers novel insights into specific risk factors and progression in this population, distinguishing our work from existing literature [12]. While previous metabolomic studies in Western populations identified similar metabolites, East Asians have been underrepresented, making our findings particularly valuable [18]. The identified metabolites hold promise for T2D management by facilitating early risk detection and tailored interventions. Elevated BCAAs and kynurenine may guide dietary or anti-inflammatory therapies to reduce insulin resistance, while DMG and SAH could target methylation pathways to enhance glucose homeostasis [1, 8, 13, 20]. These biomarkers may improve diagnostic panels for screening or monitoring T2D progression [11]. However, validation in diverse East Asian and global cohorts is essential to ensure specificity and generalizability, followed by high-throughput metabolomic assays and clinical trials to assess biomarker-guided interventions [19, 26]. Limitations include the reliance on specific diagnostic criteria for T2D and the inability to translate findings into odds ratios, complicating direct comparisons with other studies [18]. Future research should integrate lifestyle data to refine T2D prevention strategies, building on our 2025 lifestyle studies [12]. In conclusion, this study highlights the complex role of 35 GEE- and 41 LMM-significant metabolites in T2D risk among Korean adults, with distinct GEE (negative) and LMM (positive) trends for BCAAs, AAAs, and kynurenine. Interaction-specific effects and predictive models underscore the potential for early detection and personalized interventions, while the focus on an underrepresented population provides novel insights. These findings pave the way for targeted T2D prevention strategies, pending further validation and integration with lifestyle and clinical data. Conceptualization: Taesung Park. Methodology: Taesung Park and Md. Kamruzzaman. Formal Analysis: Md. Kamruzzaman and Md. Mozaffar Hosain. Data Curation: Md. Kamruzzaman, Md. Mozaffar Hosain, and Catherine Apio. Writing—Original Draft Preparation: Md. Kamruzzaman, Catherine Apio. Writing—Review and Editing: Taesung Park, Catherine Apio, Md. Kamruzzaman, Md. Mozaffar Hosain, Min Kyong Moon, Geum-Sook Hwang, and Eunyong Ahn. Visualization: Md. Kamruzzaman, Catherine Apio, and Md. Mozaffar Hosain. Supervision: Taesung Park. Project Administration: Taesung Park. Funding Acquisition: Taesung Park. We thank the participants and staff of the Korean Genome and Epidemiology Study (KoGES) for their contributions. This study was conducted with bioresources from the National Biobank of Korea, the Korea Disease Control and Prevention Agency, Republic of Korea (KBN-2020-108). No generative AI technologies were used in the preparation of this manuscript. This study was conducted following the Declaration of Helsinki principles. It was approved by the institutional review board (IRB) of Seoul National University Hospital (IRB No. 1704-021-843), and samples from the Korea Biobank Network were obtained with informed consent under IRB-approved protocols. The authors declare no conflicts of interest. Data are available from the Korean Genome and Epidemiology Study (KoGES) cohort, subject to approval by the Korea Disease Control and Prevention Agency. Metabolite data were provided by the Korea Integrated Metabolomics Research Group, Western Seoul Center, Korea Basic Science Institute, and are available upon reasonable request to Geum-Sook Hwang, PhD (email: [email protected]), subject to institutional approval and data sharing agreements. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.