Plasma metabolomics reveals progressive alterations across glycemic states

Abstract

BACKGROUND

Diabetes and prediabetes impose a major global health burden, yet conventional glycemic markers provide limited insight into early metabolic disturbances and the metabolic heterogeneity between isolated impaired fasting glucose (IFG) and isolated impaired glucose tolerance (IGT).

AIM

To investigate progressive metabolic alterations and the heterogeneity between isolated IFG and isolated IGT using plasma metabolomics.

METHODS

In this cross-sectional study, 200 participants were classified into four groups: (1) Normal glucose tolerance (NGT); (2) Isolated IFG; (3) Isolated IGT; and (4) Newly diagnosed diabetes (n = 50 each). Plasma metabolomic profiling was performed using ultra-performance liquid chromatography-tandem mass spectrometry. Multivariate analysis, differential metabolite screening, pathway enrichment, least absolute shrinkage and selection operator regression, logistic regression, receiver operating characteristic analysis, bootstrap internal validation, and correlation analysis were performed.

RESULTS

Global metabolic profiles showed a progressive shift from NGT through prediabetic states to diabetes. Eighteen shared core metabolites showed consistent differential abundance and confounder-adjusted trend evidence across the glycemic continuum and were mainly enriched in amino acid-, lipid-, and energy-related pathways. A five-metabolite candidate discriminative panel showed exploratory discriminative performance for isolated IFG, isolated IGT, and diabetes versus NGT, with area under the curve values of 0.820, 0.791, and 0.867, respectively. Exploratory analyses further suggested module-level and pathway-level differences between isolated IFG and isolated IGT.

CONCLUSION

These findings provide metabolomics-based evidence for progressive metabolic remodeling across glycemic states and identify candidate metabolic features that may help characterize early dysglycemia heterogeneity, pending validation in independent cohorts.

Key Words: Diabetes; Prediabetes; Metabolomics; Impaired fasting glucose; Impaired glucose tolerance

Core Tip: Conventional glycemic markers may not fully capture early metabolic disturbances. In this cross-sectional study, untargeted plasma metabolomics profiled 200 participants across different glycemic stages, including normal glucose tolerance, isolated impaired fasting glucose, isolated impaired glucose tolerance, and newly diagnosed diabetes. We identified 18 core metabolites consistently altered across stages. A five-metabolite candidate panel showed exploratory potential to distinguish early dysglycemia. Weighted gene co-expression network analysis suggested module-level differences between impaired fasting glucose and impaired glucose tolerance metabolic modules, which remain exploratory and require validation in independent cohorts. These findings provide candidate features for early stratification of dysglycemia.

INTRODUCTION

Diabetes is a metabolic disorder primarily characterized by chronic hyperglycemia, and its development and progression are closely linked to impaired pancreatic β-cell function, insulin resistance, or both[1]. In recent years, diabetes has become one of the most important chronic non-communicable diseases worldwide, markedly reducing quality of life and placing a substantial burden on public health systems. According to the International Diabetes Federation, an estimated 537 million adults were living with diabetes in 2021, representing nearly 10% of the global adult population, and this number is projected to increase to 783 million by 2045[2]. China bears one of the heaviest diabetes burdens worldwide, with the largest number of affected individuals globally, and this burden continues to rise[2,3]. In addition, rapid population aging has further intensified the burden of diabetes and related metabolic disorders.

In addition to individuals with overt diabetes, a substantial proportion of the population remains in an intermediate glycemic state, in which blood glucose levels are elevated above normal but do not yet meet the diagnostic criteria for diabetes, commonly defined as prediabetes[4]. Both prediabetes and diabetes are closely associated with insulin resistance and confer significantly increased risks of adverse outcomes, including cardiovascular disease[5-7], kidney disease, hypertension, and cognitive impairment[8-11]. In addition, these conditions impose substantial direct and indirect economic burdens, including long-term treatment costs, disability-related expenses, and productivity losses due to premature mortality[12-14]. Importantly, many individuals with prediabetes remain undetected and do not receive timely intervention[15], thereby increasing their risk of progression to diabetes and the development of related complications[16,17]. Therefore, identifying key biological alterations in the early stages of dysglycemia is essential for improving risk stratification, enabling early diagnosis, and guiding targeted intervention strategies.

Traditional clinical indicators, including fasting plasma glucose (FPG), 2-hour plasma glucose (2-h PG) during the oral glucose tolerance test (OGTT), and glycated hemoglobin (HbA1c), are widely used to classify glycemic status. However, these measures primarily reflect the phenotypic consequences of disrupted glucose homeostasis and provide limited insight into the earlier and more complex metabolic alterations that occur during disease initiation and progression. Metabolomics, through the systematic profiling of dynamic changes in small-molecule metabolites, offers a valuable approach for elucidating the molecular basis of dysglycemia[18]. Previous studies have shown that the development and progression of prediabetes and type 2 diabetes are accompanied by disturbances in multiple metabolites, particularly those involved in amino acid, lipid, and energy metabolism[19]. Importantly, prediabetes is not a homogeneous condition and mainly comprises two subtypes: (1) Impaired fasting glucose (IFG); and (2) Impaired glucose tolerance (IGT)[20]. Emerging evidence suggests that these two subtypes may reflect distinct pathophysiological mechanisms, with IFG being more closely associated with hepatic insulin resistance and IGT more strongly linked to reduced insulin sensitivity in peripheral tissues[21,22]. These findings suggest that IFG and IGT may differ not only in clinical phenotype but also in their underlying metabolic pathway alterations. However, most metabolomics studies to date have treated prediabetes as a single entity, and relatively few have distinguished between isolated IFG and isolated IGT, thereby limiting a more comprehensive understanding of heterogeneity in the early stages of diabetes.

Therefore, in the present study, we used an ultra-performance liquid chromatography-tandem mass spectrometry (UPLC-MS/MS) platform to systematically compare plasma metabolic profiles across four glycemic states: (1) Normal glucose tolerance (NGT); (2) Isolated IFG; (3) Isolated IGT; and (4) Newly diagnosed diabetes. We aimed to identify core metabolites and metabolic pathways that change consistently during dysglycemic progression and to further characterize the metabolomic differences between IFG and IGT. By doing so, we sought to provide new insight into early diabetes subtyping and to identify candidate metabolic biomarkers associated with early dysglycemia.

MATERIALS AND METHODS

Study participants and clinical grouping

This study was conducted at China-Japan Friendship Hospital from September 2019 to June 2021 and enrolled a total of 200 participants aged 18-70 years. The exclusion criteria were as follows: (1) Pregnant or lactating women; (2) Patients with severe psychiatric disorders; (3) Patients diagnosed with or treated for malignant tumors within the past 5 years; (4) Patients with active infectious diseases; (5) Patients with severe cardiovascular or cerebrovascular diseases (including severe arrhythmia and cardiac insufficiency); (6) Patients with severe hepatic or renal dysfunction; (7) Patients with hematological abnormalities; and (8) Those with other clinical conditions that might affect glycemic assessment, metabolomics measurements, or study compliance. The recruited participants had comparable age and sex distributions across the four groups. An overview of the study design and analytical workflow is provided in Supplementary Figure 1. No formal a priori sample size calculation was performed because this was an exploratory metabolomics discovery study. The sample size was determined based on participant availability and balanced group recruitment during the study period.

All enrolled participants underwent a standard 75 g OGTT and were subsequently classified into four clinical groups according to glycemic status (50 participants per group): (1) NGT group (group A), defined as FPG < 6.1 mmol/L and 2-h PG < 7.8 mmol/L during OGTT; (2) Isolated IFG group (group B), defined as 6.1 mmol/L ≤ FPG < 7.0 mmol/L and 2-h PG < 7.8 mmol/L; (3) Isolated IGT group (group C), defined as FPG < 6.1 mmol/L and 7.8 mmol/L ≤ 2-h PG < 11.1 mmol/L; and (4) Diabetes group (group D), defined as FPG ≥ 7.0 mmol/L or 2-h PG ≥ 11.1 mmol/L. Participants in the diabetes group were all newly diagnosed cases and had not received any glucose-lowering medications before enrollment. The study was approved by the Ethics Committee of China-Japan Friendship Hospital (No. 2019-63-K43), and written informed consent was obtained from all participants.

Metabolomics measurement and data preprocessing

Plasma metabolomics profiling was performed using a UPLC-MS/MS platform. After protein precipitation and metabolite extraction, samples were subjected to instrumental analysis, and an internal standard was added to each sample to monitor extraction efficiency. To evaluate platform stability and data quality, quality control (QC) samples were prepared by pooling equal volumes of all study samples and were inserted regularly throughout the analytical sequence for periodic assessment. System stability was evaluated by monitoring fluctuations in internal standard signals in the QC samples, and a median relative standard deviation (RSD) of internal standard signals below 5% across all QC samples was considered indicative of acceptable analytical quality. At the metabolite level, 959 metabolites were stably detected in pooled QC samples with a fill rate greater than 80%, among which only two metabolites showed an RSD > 30%, corresponding to 99.8% of metabolites with an RSD < 30%. In addition, principal component analysis (PCA) was used to visualize the clustering pattern of QC samples and to further assess the analytical stability and reproducibility of the metabolomics platform.

Raw mass spectrometry data were processed using proprietary platform software for peak extraction, peak alignment, and metabolite identification, followed by manual inspection to ensure data quality. Metabolite identities were assigned by matching experimental features against a reference library using retention time, accurate mass, and UPLC-MS/MS fragmentation spectra. The matching criteria included retention time variation < 0.1 minute, accurate mass variation < 10 ppm, and UPLC-MS/MS spectra with both forward and reverse matching scores > 75%. Annotation confidence was classified according to the evidence supporting metabolite identification: (1) Metabolites without annotation symbols were high-confidence annotations confirmed using authentic standards, corresponding to metabolomics standards initiative (MSI) level 1; (2) Metabolites marked with a single annotation symbol were putative annotations corresponding approximately to MSI level 2; (3) Metabolites marked with two annotation symbols corresponded approximately to MSI level 3; and (4) Partially characterized metabolites corresponded approximately to MSI level 4. MSI level 3/4-equivalent features were not selected as candidate markers or emphasized in the biological interpretation. During data preprocessing, missing values in the metabolite peak area matrix were imputed using the minimal detected value of each metabolite across all samples. The data were subsequently log2-transformed and normalized to reduce systematic variation prior to downstream statistical analyses and machine learning modeling.

Metabolic profiling, differential metabolite screening, and functional enrichment analysis

PCA was performed as an unsupervised dimensionality reduction approach to characterize the overall metabolic profiles of the samples and to assess distribution trends among groups. Orthogonal partial least squares-discriminant analysis (OPLS-DA) was subsequently applied as a supervised method, and variable importance in projection (VIP) values were calculated to evaluate the contribution of individual metabolites to group discrimination. VIP values were primarily used to identify and rank discriminatory metabolites and served as an auxiliary feature evaluation metric. For visualization, top-ranked metabolites in the VIP plots were color-coded according to their metabolic categories. To assess model robustness and exclude overfitting, permutation tests were performed for the OPLS-DA models.

Welch’s t-test was used for pairwise comparisons between groups. To control the false-positive rate, P values were adjusted for multiple testing using the Benjamini-Hochberg (BH) method. In this study, a BH-adjusted P value < 0.05 was defined as the primary statistical criterion for identifying significantly differential metabolites, together with VIP values > 1 in the OPLS-DA models as an auxiliary feature evaluation criterion. Mean abundance ratios and log2 fold changes (log2FCs) were calculated and reported as effect-size measures. Volcano plots were used to visualize the distribution of metabolites, with the X-axis showing log2FC and the Y-axis showing -log10(BH-adjusted P value). Each point in the plot represented a metabolite, and its position reflected both the magnitude of FC and the statistical significance of the difference.

A random forest model was applied to rank the importance of differential metabolites, using the mean decrease in accuracy as the variable importance metric. UpSet plots were used to perform intersection analysis of differential metabolites that met the criterion of BH-adjusted P value < 0.05 across group comparisons, in order to identify core metabolites showing stable alterations during the progression of dysglycemia. The selected core metabolites were standardized and visualized using heatmaps to illustrate their expression patterns across different groups. In addition, the standardized relative abundance levels of these core metabolites across clinical groups were calculated, and trajectory plots were generated to depict their dynamic patterns throughout the progression of dysglycemia. Functional enrichment analysis of the selected core metabolites was conducted using the Kyoto Encyclopedia of Genes and Genomes (KEGG) database to explore their potential biological functions.

Exploratory candidate metabolite panel selection, discriminative performance evaluation, and clinical correlation analysis

To reduce redundancy and identify an optimal feature combination, least absolute shrinkage and selection operator (LASSO) regression was performed on the core metabolites identified through intersection analysis (UpSet plot) for the comparison between the NGT group (group A) and the isolated IFG group (group B). The optimal penalty parameter (lambda) was determined by 10-fold cross-validation, and the metabolites corresponding to lambda.1se were selected to construct the final metabolite panel.

A logistic regression model was then developed using the selected metabolites to discriminate the NGT group from the isolated IFG group (A vs B). Model discrimination was evaluated using receiver operating characteristic curves and the area under the curve (AUC). Internal validation was performed using 1000 bootstrap resamples to estimate the optimism-corrected AUC. The derived metabolite panel was further applied to stage-wise comparisons to evaluate its discriminative performance for distinguishing the NGT group from the isolated IGT group (A vs C) and the diabetes group (A vs D).

In addition, boxplots were used to visualize the distribution patterns of the selected metabolites across the four clinical groups in order to compare their variation across different glycemic states. Spearman rank correlation analysis was conducted to assess the associations between the selected metabolites and clinical indicators, including glycemic traits (HbA1c, FPG, and 2-h PG), insulin-related measures [fasting insulin and homeostasis model assessment of insulin resistance (HOMA-IR)], lipid parameters [triglycerides (TG) and high-density lipoprotein cholesterol (HDL-C)], blood pressure [systolic blood pressure (SBP) and diastolic blood pressure (DBP)], liver function markers [alanine aminotransferase (ALT) and aspartate aminotransferase (AST)], renal function markers (blood urea nitrogen and creatinine), and body mass index (BMI). The correlation results were presented as correlation coefficients with their corresponding statistical significance and visualized using heatmaps.

Exploratory analysis of metabolic differences between IFG and IGT subtypes

To explore coordinated metabolic differences between the isolated IFG group (group B) and the isolated IGT group (group C), a weighted gene co-expression network analysis (WGCNA)-based metabolite co-abundance module analysis was performed using all quantified metabolites in these two groups. Metabolite abundance data were log₂-transformed, standardized, and used to construct a signed co-abundance network. Metabolite modules were identified by hierarchical clustering and dynamic tree cutting, and module eigengenes were correlated with IFG/IGT status coded as isolated IFG = 0 and isolated IGT = 1. Module-trait association P values were adjusted using the BH method.

In addition, to aid biological interpretation, an exploratory KEGG pathway enrichment analysis was conducted based on metabolites selected by nominal statistical significance (P < 0.05) and effect size (mean ratio > 1.25 or < 0.75). This pathway analysis was considered hypothesis-generating and interpreted cautiously as a complementary pathway-level visualization.

Statistical analysis

Continuous variables were expressed as median and interquartile range, and comparisons among multiple groups were performed using the Kruskal-Wallis H test. Categorical variables were presented as n (%), and group comparisons were conducted using the χ² test or Fisher’s exact test, as appropriate.

To provide quantitative evidence for the 18 shared core metabolites, multivariable linear regression analyses were performed. Standardized log2-transformed metabolite abundance was used as the dependent variable, and glycemic status was coded ordinally from NGT to diabetes to estimate adjusted trend β values. Models were adjusted for major non-glycemic clinical covariates, including age, sex, BMI, TG, SBP, DBP, ALT, and AST. P values for trend were adjusted using the BH method.

Discriminative performance of the five-metabolite candidate panel was evaluated using logistic regression models. The baseline model included standard clinical indicators and potential confounders, including age, sex, BMI, TG, SBP, DBP, ALT, and AST, and the augmented model additionally incorporated the five metabolites. Model discrimination was assessed using AUCs with 95%CIs. The incremental discriminative value of the metabolite panel was quantified using continuous net reclassification improvement and integrated discrimination improvement, with corresponding 95%CIs and P values. Variance inflation factors (VIFs) were calculated for predictors in the final five-metabolite logistic regression model to assess multicollinearity.

All statistical analyses were performed using R version 4.5.2, and a two-sided P < 0.05 was considered statistically significant.

RESULTS

Clinical characteristics of the four study groups

The clinical characteristics of the four study groups are summarized in Table 1. The four groups were comparable in age and sex, with no significant differences observed. Similarly, BMI, waist circumference, AST, blood urea nitrogen, creatinine, total cholesterol, and low-density lipoprotein cholesterol did not differ significantly across groups. In contrast, SBP, DBP, and ALT showed significant differences among the four groups (P < 0.05). TG levels were generally higher in the dysglycemic groups than in the NGT group, whereas HDL-C showed an overall decreasing trend.

Table 1 Clinical characteristics of participants across the four study groups, n (%)/median (interquartile range).

Item	Group A	Group B	Group C	Group D	P value
n	50	50	50	50
Male (%)	19 (38)	28 (56)	26 (52)	30 (60)	0.133
Age (years)	47.00 (42.00, 55.50)	52.50 (40.50, 56.25)	45.00 (40.00, 56.00)	53.00 (46.00, 58.25)	0.1
BMI (kg/m²)	24.50 (22.15, 27.25)	24.30 (22.83, 26.65)	24.90 (22.80, 26.90)	25.90 (23.50, 28.70)	0.161
WC (cm)	86.75 (78.50, 94.00)	88.50 (80.90, 93.00)	90.00 (81.25, 94.40)	91.25 (86.92, 98.50)	0.057
SBP (mmHg)	117.00 (113.00, 124.50)	119.50 (111.25, 132.00)	119.00 (110.00, 131.00)	127.00 (123.00, 137.00)	0.001
DBP (mmHg)	75.00 (69.50, 82.00)	76.50 (70.25, 85.75)	80.00 (71.00, 85.00)	82.00 (75.00, 89.00)	0.033
ALT (U/L)	20.00 (15.00, 27.50)	21.50 (17.00, 31.50)	22.00 (16.00, 34.50)	29.00 (20.75, 42.00)	0.02
AST (U/L)	20.00 (17.50, 26.00)	21.00 (18.00, 26.00)	19.50 (17.00, 25.75)	22.50 (19.75, 27.00)	0.16
BUN (mmol/L)	4.77 (3.73, 5.52)	4.93 (4.62, 6.10)	4.64 (4.06, 6.01)	4.85 (3.92, 5.58)	0.53
Cr (μmol/L)	61.30 (55.35, 78.00)	73.45 (58.02, 82.70)	65.55 (58.92, 73.92)	67.45 (58.35, 73.55)	0.356
TC (mmol/L)	4.98 (4.42, 5.94)	4.84 (4.30, 5.53)	5.04 (4.54, 5.73)	4.81 (4.38, 5.47)	0.399
TG (mmol/L)	1.23 (0.94, 1.64)	1.33 (0.92, 1.76)	1.56 (1.15, 2.04)	1.85 (1.19, 2.62)	0.001
HDL-C (mmol/L)	1.38 (1.23, 1.56)	1.23 (1.08, 1.52)	1.21 (1.10, 1.46)	1.17 (1.06, 1.44)	0.003
LDL-C (mmol/L)	3.13 (2.61, 3.83)	3.00 (2.62, 3.54)	3.10 (2.59, 3.71)	2.96 (2.71, 3.53)	0.795
HbA1c [% (mmol/mol)]	5.6 (5.2, 5.7), 38 (33, 39)	5.9 (5.6, 6.0), 41 (38, 42)	5.7 (5.4, 5.8), 39 (36, 40)	6.2 (6.0, 6.6), 44 (42, 48)	< 0.001
FPG (mmol/L)	5.63 (5.31, 5.88)	6.58 (6.35, 6.73)	5.57 (5.24, 5.79)	6.83 (6.33, 7.51)	< 0.001
Fasting insulin (μIU/mL)	7.71 (6.19, 10.69)	9.82 (7.44, 12.74)	10.56 (6.92, 14.61)	13.67 (9.63, 20.68)	< 0.001
2-h PG (mmol/L)	6.81 (6.19, 7.49)	6.94 (6.17, 7.41)	9.52 (8.82, 10.66)	12.72 (12.07, 15.38)	< 0.001
HOMA-IR	2.01 (1.50, 2.66)	2.89 (2.13, 3.88)	2.50 (1.59, 3.89)	4.18 (2.63, 6.66)	< 0.001

Group A: Normal glucose tolerance; Group B: Isolated impaired fasting glucose; Group C: Isolated impaired glucose tolerance; Group D: Diabetes. Categorical variables are presented as n (%), and continuous variables are presented as median (interquartile range). Comparisons among groups were performed using the Kruskal-Wallis H test for continuous variables and the χ² test or Fisher’s exact test for categorical variables. A two-sided P < 0.05 was considered statistically significant. BMI: Body mass index; WC: Waist circumference; SBP: Systolic blood pressure; DBP: Diastolic blood pressure; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; BUN: Blood urea nitrogen; Cr: Creatinine; TC: Total cholesterol; TG: Triglycerides; HDL-C: High-density lipoprotein cholesterol; LDL-C: Low-density lipoprotein cholesterol; HbA1c: Glycated hemoglobin; FPG: Fasting plasma glucose; 2-h PG: 2-hour plasma glucose; HOMA-IR: Homeostasis model assessment of insulin resistance.

Among all variables, glycemia-related indicators exhibited the most marked intergroup differences. HbA1c, FPG, 2-h PG, fasting insulin, and HOMA-IR all differed significantly among the four groups (all P < 0.001). The isolated IFG group was characterized predominantly by elevated FPG, whereas the isolated IGT group was characterized by elevated 2-h PG. More pronounced abnormalities across multiple glycemic indicators were observed in the diabetes group.

Overall metabolic profiling of the four study groups

The PCA score plot demonstrated the overall distribution of metabolic profiles across the four groups (Figure 1). A certain degree of separation was observed among the groups, with the diabetes group (group D) showing a more pronounced shift in the principal component space relative to the NGT group (group A), whereas the isolated IFG group (group B) and the isolated IGT group (group C) were located between the normal and diabetes groups. Although PC1 and PC2 together explain only 30.1% of the total variance, this PCA plot provides an exploratory visualization of overall group trends rather than a full representation of all sample-level variation. These findings indicate a progressive alteration in the global metabolic profile with worsening glycemic status. Additional pairwise PCA comparisons are provided in Supplementary Figure 2. In addition, QC samples were tightly clustered in the PCA space, supporting the analytical stability and reproducibility of the metabolomic profiling workflow (Supplementary Figure 3).

Figure 1 Principal component analysis score plot of the four study groups. The principal component analysis score plot illustrates the overall distribution of samples across the four study groups: (1) Group A: Normal glucose tolerance; (2) Group B: Isolated impaired fasting glucose; (3) Group C: Isolated impaired glucose tolerance; and (4) Group D: Diabetes. The X-axis indicates the first principal component, explaining 23.3% of the variance, and the Y-axis indicates the second principal component, explaining 6.8% of the variance. First principal component and second principal component together explained 30.1% of the total variance; therefore, this plot was used as an exploratory visualization of overall group trends rather than a complete representation of all sample-level variation. Each point represents an individual sample, and ellipses indicate the confidence regions for each group. PC1: First principal component; PC2: Second principal component.

To further characterize metabolic differences among the glycemic groups, OPLS-DA models were constructed for all pairwise comparisons (Supplementary Figure 4). The OPLS-DA model metrics are summarized in Supplementary Table 1. For the main comparisons with the NGT group, IFG vs NGT, IGT vs NGT, and diabetes vs NGT showed R²Y(cum) values of 0.330, 0.827, and 0.801, Q²(cum) values of 0.270, 0.244, and 0.458, and CV-ANOVA P values of 2.35 × 10^-7, 1.75 × 10^-4, and 1.09 × 10^-10, respectively. The strongest model performance was observed for diabetes vs NGT, and the supervised model results were interpreted together with CV-ANOVA metrics and permutation testing, which supported the robustness of the OPLS-DA models and suggested no obvious overfitting (Supplementary Figure 5). VIP analysis indicated that the metabolites contributing most strongly to group discrimination were mainly related to amino acid, lipid, and energy metabolism.

Differential metabolite screening and integrative analysis of core metabolic features

To systematically identify metabolites associated with dysglycemia, pairwise comparisons were performed between groups, and volcano plots were used to visualize the distribution of differential metabolites (Supplementary Figure 6). The results showed substantial variation in the number of differential metabolites across comparisons, with the most pronounced metabolic differences observed between the NGT and diabetes groups, whereas the differences between the NGT and isolated IGT groups were relatively less marked. Overall, the number of differential metabolites increased progressively with worsening glycemic status, suggesting that metabolic dysregulation intensified during disease progression.

Based on these differential analysis results, a random forest model was further applied to rank the importance of differential metabolites across all participants (Figure 2A). Metabolites such as valine, spermine, and alpha-ketoglutarate (α-KG) showed relatively high importance scores, indicating a strong contribution to the discrimination of different glycemic states. UpSet analysis was then performed on the differential metabolites identified in the comparisons between the NGT and isolated IFG groups (A vs B), the NGT and isolated IGT groups (A vs C), and the NGT and diabetes groups (A vs D) (Figure 2B). Eighteen core metabolites that were significantly altered in all three comparisons were identified, suggesting that these metabolites may represent relatively stable metabolic features during the progression of dysglycemia.

Figure 2 Differential metabolite screening and functional enrichment analysis. A: Variable importance analysis of differential metabolites in the overall study population based on the random forest model; B: UpSet plot showing the intersections of differential metabolites identified in the comparisons between the normal glucose tolerance (NGT) and isolated impaired fasting glucose groups (A vs B), the NGT and isolated impaired glucose tolerance groups (A vs C), and the NGT and diabetes groups (A vs D); C: Heatmap showing the expression patterns of the 18 shared differential metabolites across the four groups (NGT, impaired fasting glucose, impaired glucose tolerance, and diabetes); D: Kyoto Encyclopedia of Genes and Genomes pathway enrichment analysis based on the 18 shared differential metabolites. The X-axis represents the enrichment ratio, bubble size indicates the number of metabolites enriched in each pathway, and bubble color represents the adjusted P value.

To characterize the expression patterns of these core metabolites during disease progression, a heatmap was generated to visualize their abundance across the four groups (NGT, isolated IFG, isolated IGT, and diabetes) (Figure 2C). The heatmap revealed a certain degree of regularity in the expression patterns of the 18 shared differential metabolites. Some metabolites were present at higher levels in the NGT group and decreased in the diabetes group, whereas others showed a gradual increase from the NGT group to the diabetes group. Overall, the isolated IFG and isolated IGT groups exhibited transitional patterns between the normal and diabetes groups, suggesting bidirectional metabolic alterations during the progression of dysglycemia. Subsequently, KEGG pathway enrichment analysis was performed based on these 18 core metabolites (Figure 2D). The results showed that these metabolites were mainly enriched in pathways related to amino acid metabolism and energy metabolism, including amino acid biosynthesis, central carbon metabolism, and branched-chain amino acid metabolism. These findings indicate that the progression of dysglycemia is accompanied by systematic alterations in multiple key metabolic pathways.

In addition, the trajectories of the 18 core metabolites across different clinical groups were further examined (Figure 3). To support these visual patterns with quantitative evidence, Supplementary Table 2 summarizes the VIP ranges, B/A, C/A, and D/A FCs with FDR values, and confounder-adjusted trend statistics for the 18 shared core metabolites. In multivariable analyses, most amino acid-, lipid-, energy-, and nucleotide-related metabolites showed positive adjusted trends across worsening glycemic status, whereas peptide-related metabolites showed inverse or stage-specific patterns. Consistently, metabolites such as valine, leucine, isoleucine, glutarate (C5-DC), and α-KG showed increasing patterns from NGT to diabetes, whereas threonylphenylalanine, phenylalanylglycine, and glu-gly-asn-val showed overall inverse or non-linear decreasing patterns. Other metabolites, such as spermine and adenine, showed early alterations at the isolated IFG or isolated IGT stage and further stage-specific changes in diabetes. These findings suggest that dysglycemic progression is characterized by heterogeneous, quantitatively supported, and stage-specific metabolic alterations.

Figure 3 Trajectories of the 18 shared core metabolites across different glycemic states. Line plots illustrate the standardized relative abundance of the 18 core metabolites across four clinical groups: (1) Normal glucose tolerance; (2) Impaired fasting glucose; (3) Impaired glucose tolerance; and (4) Diabetes. Each panel represents one metabolite, and the data were standardized using Z-score transformation. The X-axis indicates the clinical groups, and the Y-axis indicates the standardized metabolite level. Z-score: Standardized score.

Exploratory candidate metabolite panel selection and clinical correlation analysis

Using the 18 core metabolites identified by UpSet intersection analysis, LASSO regression was performed for variable selection in the comparison between the NGT group (group A) and the isolated IFG group (group B) (Figure 4A and B). As the penalty parameter lambda increased, the coefficients of several metabolites gradually shrank toward zero. Based on model parsimony and robustness, a five-metabolite candidate discriminative panel corresponding to lambda.1se was ultimately retained: (1) Leucine; (2) Valine; (3) Glutarate (C5-DC); (4) Palmitate (16:0); and (5) Threonylphenylalanine.

Figure 4 Least absolute shrinkage and selection operator-based selection of candidate discriminative metabolic biomarkers, discrimination performance, and clinical associations. A: Least absolute shrinkage and selection operator coefficient profiles of candidate metabolites for discriminating between the normal glucose tolerance (NGT) and impaired fasting glucose (IFG) groups; B: Cross-validation curve of the least absolute shrinkage and selection operator model. The dashed lines indicate the lambda values corresponding to the minimum error criterion (lambda.min) and the 1-SE criterion (lambda.1se), respectively. Based on model parsimony and robustness, lambda.1se was selected to identify the final metabolite panel; C: Receiver operating characteristic curve of the final metabolite panel for discriminating between the NGT and IFG groups, with the optimism-corrected area under the curve shown; D: Receiver operating characteristic curves evaluating the stage-wise discriminative performance of the selected metabolite panel for the comparisons of NGT vs IFG, impaired glucose tolerance, and diabetes; E: Boxplots showing the distribution of the selected metabolites across the NGT, IFG, impaired glucose tolerance, and diabetes groups. ^bP < 0.01; F: Spearman correlation heatmap showing the associations between the selected metabolites and clinical indicators; color intensity represents Spearman correlation coefficients. AUC: Area under the curve; HbA1c: Glycated hemoglobin; FPG: Fasting plasma glucose; 2-h PG: 2-hour plasma glucose; HOMA-IR: Homeostasis model assessment of insulin resistance; TG: Triglycerides; HDL-C: High-density lipoprotein cholesterol; SBP: Systolic blood pressure; DBP: Diastolic blood pressure; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; BUN: Blood urea nitrogen; Cr: Creatinine; BMI: Body mass index. A: Normal glucose tolerance; B: Isolated impaired fasting glucose; C: Isolated impaired glucose tolerance; D: Diabetes.

A logistic regression model was then constructed using these five selected metabolites to exploratorily assess their discriminative performance. The apparent AUC for distinguishing the NGT group from the isolated IFG group (A vs B) was 0.850 (95%CI: 0.771-0.929). Following internal validation with 1,000 bootstrap resamples, the optimism-corrected AUC was 0.820 (95%CI: 0.746-0.908). When further applied to stage-wise comparisons, the metabolite panel yielded AUCs of 0.791 (95%CI: 0.701-0.881) for A vs C and 0.867 (95%CI: 0.796-0.938) for A vs D, indicating discriminative performance across different stages of dysglycemia, with the highest AUC observed for diabetes versus NGT (Figure 4C and D).

To evaluate whether the five-metabolite candidate panel provided incremental discriminative value beyond standard clinical indicators and potential confounders, a baseline logistic regression model including age, sex, BMI, TG, SBP, DBP, ALT, and AST was constructed. An augmented model incorporating the five selected metabolites was then compared with the baseline model. The addition of the five-metabolite panel consistently improved model performance across all three comparisons (Supplementary Table 3). The AUC increased from 0.663 to 0.868 for A vs B, from 0.755 to 0.868 for A vs C, and from 0.842 to 0.945 for A vs D, with significant improvements in AUC, net reclassification improvement, and integrated discrimination improvement across all comparisons. These findings suggest that the five-metabolite candidate panel provided incremental discriminative information beyond standard clinical indicators, including BMI, TG, blood pressure, and liver enzyme markers. VIF analysis showed no substantial multicollinearity among the five metabolites in the final logistic regression model; all VIF values were below 5, including leucine and valine (4.266 and 4.186, respectively, Supplementary Table 4).

The distributions of the five selected metabolites differed significantly across the four groups (Figure 4E). Most showed broadly increasing trends from NGT to dysglycemic states, although certain metabolites exhibited stage-specific non-linear patterns. Correlation analysis further demonstrated significant associations between these metabolites and multiple clinical indicators (Figure 4F). In particular, leucine and valine were positively correlated with HbA1c, FPG, 2-h PG, fasting insulin, HOMA-IR, and TG, while negatively correlated with HDL-C, supporting their close links to dysglycemia and insulin resistance.

Exploratory analysis of module-level and pathway-level differences between IFG and IGT subtypes

To further characterize coordinated metabolic differences between isolated IFG and isolated IGT, we performed a WGCNA-based metabolite co-abundance module analysis using all quantified metabolites in these two groups. Nine metabolite modules were identified, among which the turquoise, pink, and brown modules were significantly associated with IFG/IGT status after false discovery rate (FDR) correction (turquoise: r = -0.279, FDR = 0.024; pink: r = -0.268, FDR = 0.024; brown: r = 0.264, FDR = 0.024, Supplementary Table 5). The turquoise and pink modules were higher in isolated IFG, whereas the brown module was higher in isolated IGT. Functional annotation of these significant modules suggested coordinated alterations involving lipid-related, amino acid-related, peptide-related, xenobiotic-related, nucleotide-related, and cofactor/vitamin-related metabolic categories (Supplementary Table 6). Exploratory KEGG pathway enrichment analysis was performed as a complementary pathway-level visualization to aid biological interpretation (Supplementary Figure 7). This analysis highlighted several pathways related to amino acid, lipid, redox, and energy metabolism, including alanine, aspartate and glutamate metabolism, glycerophospholipid metabolism, glutathione metabolism, sphingolipid metabolism, and arginine and proline metabolism. These findings suggest module-level metabolic differences between isolated IFG and isolated IGT, but should be interpreted cautiously and require further validation.

DISCUSSION

Using a UPLC-MS/MS-based metabolomics platform, the present study systematically characterized metabolic alterations across NGT, isolated IFG, isolated IGT, and diabetes. We observed a gradient pattern of metabolic disturbance across glycemic states, suggesting that progression from normoglycemia to overt diabetes is accompanied by coordinated changes in metabolite abundance and related pathways.

Previous metabolomics studies have examined the metabolic abnormalities associated with prediabetes and type 2 diabetes from multiple perspectives. A systematic review and meta-analysis by Guasch-Ferré et al[19] showed that the most consistent metabolic features of dysglycemia are concentrated in amino acid-related, lipid-related, and glucose-related pathways. In particular, branched-chain amino acids (BCAAs), aromatic amino acids, glycine, glutamine, phospholipids, sphingolipids, and TG were closely associated with the development and progression of prediabetes and type 2 diabetes. The same meta-analysis further demonstrated that elevated isoleucine, leucine, and valine were consistently associated with an increased future risk of type 2 diabetes, whereas glycine and glutamine showed protective associations[19]. In a nested case-control study within the Framingham Offspring cohort, Wang et al[23] reported that BCAAs and aromatic amino acids were significantly associated with fasting insulin, HOMA-IR, and future risk of type 2 diabetes, suggesting that disturbances in amino acid metabolism may arise early in the course of dysglycemia. Würtz et al[24] further showed that BCAAs and phenylalanine predicted fasting and post-OGTT glucose levels, whereas gluconeogenesis-related metabolites, including alanine, lactate, and pyruvate, were more strongly associated with postprandial glucose changes. These findings suggest that alterations in these metabolic processes may already be present at an early stage of dysglycemia[24]. In addition, studies by Li et al[25] and Jabbar Al-Rikabi et al[26] showed that amino acids and acylcarnitines differ significantly across glycemic stages and are involved in pathways related to branched-chain amino acid metabolism, fatty acid metabolism, and energy metabolism, further supporting an important role for multi-pathway metabolic dysregulation in dysglycemia.

Building on previous evidence, the present study adopted a more refined phenotypic framework by systematically comparing four groups: (1) NGT; (2) Isolated IFG; (3) Isolated IGT; and (4) Diabetes. We identified 18 core metabolites showing reproducible alterations across glycemic states. Group-wise trend analyses further showed that several metabolites, particularly valine and leucine, increased progressively with worsening glycemic status, whereas others exhibited less uniform or non-linear patterns across stages. Pathway enrichment analysis indicated that these differential metabolites were mainly involved in amino acid, lipid, and energy metabolism, suggesting that multiple metabolic pathways undergo systematic alterations across glycemic states. In addition, isolated IFG and isolated IGT showed differences in amino acid-related and lipid-related pathway enrichment, with additional involvement of redox-related and energy-related pathways.

Notably, among the core metabolites identified in this study, the BCAAs leucine and valine showed relatively consistent upward trends across glycemic stages, suggesting that altered amino acid metabolism may be an important feature of dysglycemia. This observation is consistent with previous studies reporting associations between elevated BCAAs, insulin resistance, and future diabetes risk[27-30]. Mechanistic studies have proposed several possible explanations, including mTOR pathway activation, impaired BCAA catabolism, accumulation of branched-chain keto acids, mitochondrial dysfunction, oxidative stress, and ectopic lipid deposition[31-34]. However, these mechanisms were not directly tested in the present cross-sectional study and should therefore be interpreted as biologically plausible explanations rather than causal evidence derived from our data.

Other core metabolites may also reflect broader metabolic alterations across glycemic states. α-KG, a tricarboxylic acid cycle intermediate, may indicate changes in energy metabolism and has been implicated in metabolic-epigenetic regulation in previous studies[35-37]. Similarly, the observed alterations in spermine may be related to polyamine metabolism, redox homeostasis, and cellular stress responses, as suggested by prior experimental evidence[38,39]. These interpretations should therefore be considered hypothesis-generating and require confirmation in longitudinal and mechanistic studies.

Adenine, a purine-related metabolite, may reflect alterations in purine metabolism, nucleotide turnover, and cellular energy stress during dysglycemic progression. In addition, glutarate (C5-DC) and fatty acid-derived metabolites, including hydroxy fatty acids and long-chain fatty acids, may indicate altered acylcarnitine/dicarboxylate metabolism, incomplete fatty acid oxidation, and mitochondrial metabolic inflexibility[25-27]. These alterations are biologically consistent with impaired substrate utilization and mitochondrial stress in insulin-resistant states, but should be further validated in longitudinal and mechanistic studies.

Beyond descriptive metabolite profiling, we further refined the 18 core metabolites into a smaller candidate discriminative panel using LASSO regression. The resulting five-metabolite panel showed exploratory discriminative performance not only for identifying IFG, but also for distinguishing more advanced dysglycemic states, suggesting that these metabolites may reflect a continuum of metabolic deterioration rather than a single cross-sectional alteration. However, because this panel was derived and internally validated within the same modest-sized cross-sectional cohort, it should be interpreted as a candidate discriminative panel rather than a validated biomarker model.

Importantly, the selected metabolite panel was dominated by markers related to amino acid and lipid metabolism, particularly leucine and valine, which is in line with the well-established role of BCAAs in insulin resistance and dysglycemia. The correlation patterns further support this interpretation, as these metabolites were closely linked to glycemic, insulin-related, and lipid traits. Taken together, these findings suggest that the identified metabolite signature reflects an integrated metabolic state characterized by impaired glucose handling, insulin resistance, and altered lipid metabolism. In this respect, our study identified stage-related metabolic alterations and summarized them into a parsimonious candidate panel that warrants further evaluation and external validation before potential use in early metabolic stratification.

Notably, this study further compared the two prediabetic subtypes, isolated IFG and isolated IGT. To avoid relying solely on individual metabolite-level significance, we performed a WGCNA-based co-abundance module analysis using all quantified metabolites in these two groups. Three modules were significantly associated with IFG/IGT status after FDR correction, suggesting coordinated module-level metabolic differences between the two subtypes. Functional annotation indicated that these modules mainly involved lipid-related, amino acid-related, peptide-related, xenobiotic-related, nucleotide-related, and cofactor/vitamin-related metabolic categories. Previous studies have shown that IFG is primarily associated with hepatic insulin resistance, whereas IGT more strongly reflects reduced insulin sensitivity in peripheral tissues, particularly skeletal muscle[21,22]. Our findings extend this physiological distinction by suggesting that the divergence between IFG and IGT may also be detectable at the metabolomic module level. Nevertheless, these subgroup findings remain exploratory and require validation in larger independent cohorts.

Several limitations should be acknowledged. First, the cross-sectional design limits causal inference regarding the temporal sequence between metabolic alterations and dysglycemic progression. Therefore, the pathway and mechanistic interpretations should be considered hypothesis-generating rather than causal. Second, the sample size was relatively modest, particularly for the comparison between isolated IFG and isolated IGT, and these subgroup findings should therefore be interpreted with caution. Third, the five-metabolite panel was derived and internally validated within the same modest-sized study population. Although bootstrap optimism correction was applied, external validation in independent cohorts is required before clinical application.

CONCLUSION

In summary, this study systematically characterized metabolic alterations across different glycemic states and identified a candidate metabolite panel with exploratory discriminative value. WGCNA-based module analysis further suggested coordinated metabolic differences between isolated IFG and isolated IGT, providing hypothesis-generating insight into early dysglycemia subtyping and metabolic heterogeneity. These findings require validation in larger independent cohorts.

ACKNOWLEDGEMENTS

The authors thank all participants and staff for their support and assistance in this study.