Circulating Metabolite Biomarkers of Glycemic Control in Youth-Onset Type 2 Diabetes

Type 2 diabetes is rising at an alarming rate in children (1). Compared with adults with type 2 diabetes, children with type 2 diabetes have more rapid β-cell failure, worse insulin resistance, and earlier and more severe complications (2–4). Treatments for children with type 2 diabetes, however, are limited and less effective (4), with newer agents (e.g., sodium–glucose transport protein 2 inhibitors) only recently approved (5). The TODAY (Treatment Options for Type 2 Diabetes in Adolescents and Youth) study was a multicenter longitudinal study of the treatment and natural history of youth-onset type 2 diabetes (6). In TODAY, nearly half of the participants reached the primary outcome of glycemic failure, despite random assignment to either metformin monotherapy, metformin and rosiglitazone combination therapy, or metformin plus intensive lifestyle intervention. In the observational phase of the study (TODAY2), cumulative incidence of any microvascular complication was 80% after 15 years of follow-up, independent of treatment group (7). Given the severe phenotype and poor treatment response in youth-onset diabetes, investigations into molecular markers of glycemic failure are needed to elucidate pathophysiology and improve therapies.

Metabolomics is a powerful high-throughput tool to identify novel biomarkers and uncover underlying disease pathways. Elevated levels of circulating branched-chain amino acids (BCAAs) and aromatic amino acids predict diabetes more than a decade before clinical diagnosis (8,9). These associations with diabetes, along with changes in other amino acids, oligosaccharides, and specific lipid species, have been replicated in numerous adult population cohorts (10,11). Moreover, citrulline levels are higher in adults with type 2 diabetes treated with metformin (12), suggesting that metabolites may also reflect pharmacologic exposures. Differences in circulating levels of adenosine monophosphate, phospholipids, and bile acids in adults at high risk of developing type 2 diabetes were also associated with different rates of diabetes progression after either metformin or lifestyle interventions (13). Metabolomics, therefore, can be used to identify biomarkers of diabetes and treatment effects, but it has yet to be widely applied in youth-onset type 2 diabetes.

Metabolomic studies in youths with type 2 diabetes are limited but have identified unique signatures. In a small cross-sectional study comparing youths with normal weight, obesity, or diabetes, BCAA, alanine, and arginine levels were lower in those with diabetes (14), in direct contrast with the positive associations observed in adults (10). Furthermore, short- and medium-chain acylcarnitines were lower in youths with type 2 diabetes, contrary to associations reported in adults. Other pediatric studies have reported higher plasma levels of BCAAs in obesity, insulin resistance, and type 2 diabetes (9,15).

Our study represents one of the first to longitudinally profile metabolomic signatures in youths with type 2 diabetes, with the goal of identifying predictors of treatment response and loss of glycemic control. We hypothesized that TODAY participants with subsequent loss of glycemic control would have distinct metabolomic signatures and that these metabolites would be associated with measures of glycemia and β-cell function. Secondarily, we hypothesized that specific metabolites would be predictive of treatment response in the TODAY trial. Our findings have the potential to identify novel biomarkers that predict disease progression and treatment failure and to uncover pathogenic pathways in type 2 diabetes.

The TODAY study protocol and results have been published (6). Briefly, TODAY was a multicenter randomized clinical trial that studied treatment with metformin alone (MET), metformin plus rosiglitazone (MET + ROS), and metformin plus lifestyle intervention (MET + LI). A total of 699 youths aged 10–17 years within <2 years from diagnosis were recruited from 2004 to 2009. Participants were observed for a mean of 3.8 years during the original TODAY study. Of these participants, 572 were observed for up to 15 additional years (mean 10.2 years) in the TODAY2 postintervention observational follow-up study (7). Study protocols were reviewed and approved by the institutional review board of each clinical center; all study participants and their legal guardians provided written assent or consent, respectively.

The primary outcome of the TODAY clinical trial, and for our current analyses, was time to treatment failure or loss of glycemic control, defined as HbA_1c ≥8% for at least 6 months, inability to discontinue insulin therapy within 3 months after initiation, or need to resume insulin after it was discontinued within 3 months (6).

HbA_1c and 2-h oral glucose tolerance tests (OGTTs) were assessed every 6 months during the first year of TODAY and annually until the participant reached the primary outcome. Indices for insulin sensitivity and β-cell function were calculated from OGTT-derived values. Fasting insulin, fasting C-peptide, and 1/fasting insulin were used as measures of insulin sensitivity (16). The C-peptide index (incremental ΔC-peptide/Δglucose over the first 30 min of the OGTT) was used as a measure of β-cell function. The C-peptide oral disposition index (C-oDI; insulin sensitivity × C-peptide index) was used to quantify β-cell reserve in the context of insulin sensitivity (17–19).

Metabolomic profiling was conducted in available fasting plasma samples from 397 TODAY participants at baseline and 312 at 36 months; 304 were included at both time points. High-throughput liquid mass spectrometry methods have been described (20). A HILIC method profiled 361 targeted amino acids, carnitines, and phospholipids and their derivatives. A targeted AMIDE method was used to measure 155 nucleic acids, organic acids, and carbohydrates. Quality-control pools inserted every 20 samples were used to normalize intensity trends across batches and calculate the coefficient of variation for each compound. The median coefficient of variation was 7.3% for the HILIC compounds and 9.2% for the AMIDE compounds. Compounds that had >10% missingness at either timepoint were excluded from the analysis, leaving 480 compounds (HILIC n = 350; AMIDE n = 130). Compounds with <10% missingness were imputed to half the minimum detected value. Standard curves using isotope-labeled standards of α-ketoglutarate and pyruvate measured in a random sampling of 20 TODAY participants demonstrated good linearity in our mass spectrometry measurement range and high correlation between reported mass spectrometry intensities and absolute concentrations that were within the expected endogenous range (21,22) (Supplementary Fig. 1).

Compound concentrations were natural log (ln) transformed to correct for nonnormal distributions and standardized (mean 0; SD 1 for each compound) to account for heteroskedasticity. Δ_compound, defined as the difference in ln(36-month compound levels) − ln(0-month compound levels), was also standardized (mean 0; SD 1). Concentrations at 0 months (i.e., baseline) and Δ_compound between 0 and 36 months were associated with baseline glycemic, insulin, and β-cell measurements using linear regression models adjusted for age, sex, race/ethnicity, treatment group, and BMI.

Metabolite associations with the primary outcome were modeled using Cox proportional hazards adjusting 1) for age, sex, race/ethnicity, and BMI and 2) additionally for HbA_1c. Effect modification by treatment was assessed using a treatment group × compound concentration interaction. Δ_compound associations with time to treatment failure were assessed using similar models but restricted to individuals who had not developed treatment failure by 36 months of follow-up (n = 185). We defined the statistical significance threshold as a Benjamini-Hochberg false discovery rate (q) of <0.05 to address multiple-hypothesis testing (23). Mediation analyses (24,25) were conducted to assess if changes in metabolites over 36 months mediated baseline metabolite associations with treatment failure or with glycemic traits previously associated with treatment failure in TODAY (16,17).

Prediction models were created by selecting metabolites with elastic net regularization using 10-fold cross validation after splitting the cohort into 60% for training and 40% for testing. Two separate sets of prediction models were created based on the compounds included for selection. One set was selected from all 480 circulating measured compounds. The other was selected from the 33 compounds associated with treatment failure in the primary analysis. Prediction models for these two sets of compounds at baseline and with change over 36 months were separately compared with a clinical prediction model including only baseline clinical measures (i.e., HbA_1c, C-oDI, C-peptide index, and maternal diabetes history), the model with the addition of age and sex to these clinical measures, and a model that included all clinical traits and metabolites together. The areas under the curve (AUCs) and receiver operating characteristic (ROC) curves from the testing data set are reported. The statistical package R (version 4.3.2) was used for all analyses.

Clinical and demographic characteristics of TODAY participants included in this analysis (n = 304) are shown in Supplementary Table 1. Participants were equally distributed across the three treatment groups. Although most clinical and demographic characteristics were similar, there was a difference in BMI, with the highest mean value in the MET group and lowest in MET + LI (P = 0.009); we therefore adjusted for BMI in all analyses. In contrast with the TODAY primary study findings (N = 699) where MET + ROS had a reduced risk of glycemic failure, there was no difference in treatment failure hazards between the three treatment groups in our subset (n = 304) (6). Besides a slightly higher percentage of youths belonging to the “other” category for self-identified race/ethnicity, there were no differences in clinical or demographic characteristics among participants who did or did not have metabolomic profiling (Supplementary Table 2). A majority of participants in our analyses were in Tanner stage 4 or 5 (88%), and Tanner stage did not significantly modify associations.

Thirty-three baseline plasma metabolite measurements were associated with treatment failure, independently of treatment randomization, age, sex, race/ethnicity, or baseline BMI (Fig. 1A and Supplementary Table 3). Associations without BMI adjustment are shown in Supplementary Table 3. Seven of these associations remained significant after further adjusting for baseline HbA_1c, including the positive associations of glucose/fructose/galactose (hazard ratio [HR] 1.51; 95% CI 1.22–1.89; q = 1.39 × 10⁻²), α-ketoglutaric acid (1.38; 1.15–1.66; q = 2.37 × 10⁻²), and lactic acid (1.33; 1.12–1.58; q = 2.96 × 10⁻²) and the inverse associations of creatinine (0.77; 0.68–0.88; q = 1.01 × 10⁻²), CAR DC5:0 (0.75; 0.63–0.88; q = 4.79 × 10⁻²), xanthurenic acid (0.71; 0.61–0.82; q = 1.48 × 10⁻³), and N-acetyl-l-methionine (0.67; 0.57–0.79; q = 3.35 × 10⁻⁴) (Fig. 1B). Additionally, 17 remained significant after adjusting for C-peptide index, 18 for C-oDI, and 26 for presence of maternal diabetes (Supplementary Table 3). N-acetyl-l-methionine (HR 0.64; 95% CI 0.54–0.76; q = 2.73 × 10⁻⁵) was associated with treatment failure in a model adjusting for multiple traditional clinical risk factors, including age, sex, race/ethnicity, BMI, HbA_1c, and maternal diabetes.

Creatinine was inversely associated with treatment failure in all models. Estimated glomerular filtration rate (eGFR) calculated using the full age spectrum equation (26) did not differ across treatment groups (Supplementary Table 1) or according to treatment failure. In sensitivity analyses in which we adjusted for eGFR (Supplementary Table 4), we found that HRs were strikingly concordant, suggesting that metabolite associations with loss of glycemic control are not explained by differences in renal filtration (Supplementary Fig. 2).

In cross-sectional analyses, associations of baseline compounds with baseline fasting plasma glucose (FPG), HbA_1c, insulin, C-peptide, C-peptide index, and C-oDI were modeled using linear regression (27). Thirty-six compounds were associated with FPG (q < 0.05), 26 with HbA_1c, 16 with insulin, 38 with C-peptide, three with C-peptide index, and four with C-oDI (Supplementary Table 5). The glycemic and insulin trait associations for the 33 biomarkers of treatment failure are illustrated in Fig. 2 as a heatmap, with compounds ordered by their hazards for treatment failure. A total of 17 of the 33 were significantly associated (q < 0.05) with FPG, 15 with HbA_1c, three with insulin, six with C-peptide, two with C-peptide index, and three with C-oDI.

Associations of baseline metabolites with changes in clinical traits over 36 months are listed in Supplementary Table 6. Associations of metabolite changes with clinical trait changes are listed in Supplementary Table 7. Hexose was the only compound associated with both baseline FPG and HbA_1c and changes in FPG and HbA_1c. Of the 33 treatment failure–associated compounds, four baseline metabolites and 13 changes in metabolites were associated with changes in FPG, 15 baseline levels were associated with changes in HbA_1c, eight were associated with changes in C-peptide, and change in 3-hydroxybutyric acid was associated with change in insulin levels. We also assessed if changes in metabolites mediated the change in glycemic, insulin, or β-cell traits. Baseline insulin, C-peptide, C-peptide index, C-oDI, and HbA_1c values, but not FPG, were associated with their 36-month values (after adjusting for age, sex, race/ethnicity, treatment randomization, and treatment outcome; P < 0.002). Changes in palmitic acid (P = 0.003) and 1,5-AG/1-deoxyglucose (P = 0.02) levels were found to mediate changes in C-peptide levels at 36 months. We repeated the analysis to examine change in clinical values over 48 months (Supplementary Table 8) and noted a trend toward changes in taurodeoxycholic acid mediating changes in C-peptide.

We conducted exploratory analyses to determine if treatment randomization interacted with associations between metabolites and treatment failure. Of the 33 baseline compounds associated with the primary outcome, treatment group modified the associations of hexose and N4-acetylcytidine (P_{interaction term} < 0.05). Individuals with hexose levels in the highest quartile had an increase in treatment failure risk of >11-fold compared with those in the lowest quartile when they were randomly assigned to MET + ROS, whereas those randomly assigned to MET had an increase of ∼50% (Table 1). Individuals with N4-acetylcytidine in the highest quartile had a decrease of ∼50% in treatment failure risk when compared with those in the lowest quartile, except when randomly assigned to MET + LI, where there seemed to be no difference. Treatment randomization modified treatment failure associations for an additional 23 baseline metabolites (Table 1).

We next used Cox models to determine if changes in metabolite concentrations over 36 months (Δ_compound) were associated with treatment failure, adjusting for age, sex, race/ethnicity, treatment randomization, and BMI. We excluded all individuals who developed treatment failure before 36 months of follow-up, which was when the second metabolomic measurements were obtained. This reduced our total number to 185 individuals, with 20 developing treatment failure subsequently (mean follow-up time of 45 months). Changes in d-gluconic acid (HR 5.60; 95% CI 2.91–10.77; q = 7.41 × 10⁻⁵), glucose/fructose/galactose (5.56; 2.89–10.71; q = 7.41 × 10⁻⁵), and 1,5-AG/1-deoxyglucose (0.23; 0.13–0.42; q = 2.56 × 10⁻⁴) were associated with treatment failure and remained significant after further adjusting for baseline HbA_1c (Supplementary Table 9). No metabolites remained significant after further adjusting for C-peptide index, C-oDI, or maternal diabetes status. Because the Δ_compound values reflected treatment effect on the metabolome, we assessed whether metabolite changes mediated the baseline metabolite association with treatment failure; however, none were significant. We also explored whether metabolite changes mediated the association of glucose, HbA_1c, insulin, C-peptide index, or C-oDI with treatment failure and found that changes in N-acetylglucosamine/N-acetylgalactosamine, hexose, aconitic acid, LPE 18:0, LPC 15:0, and LPC 19:0 mediated the association of glucose with treatment failure at a nominal significance level (P < 0.04).

Metabolite prediction models using baseline concentrations were created using elastic net and were compared with known clinical predictors, including HbA_1c, C-oDI, C-peptide index, maternal diabetes history, and a full clinical model (age, sex, maternal diabetes history, HbA_1c, and BMI) (Fig. 3 and Supplementary Table 10). A model with 35 compounds that were selected using elastic net from the 480 compounds that we measured had an AUC of 0.67 (Fig. 3A and Supplementary Table 11). The AUC was 0.77 for a model including baseline HbA_1c only. There was also no improvement in AUC when these metabolites were added to a clinical model that included age, sex, and baseline HbA_1c (0.74 vs. 0.78). Metabolite prediction models (for both baseline and change in levels) created by selecting from the 33 treatment failure–associated compounds had AUCs comparable to the clinical-only prediction models (Fig. 3B–D and Supplementary Table 10). Models that included the change over 36 months of two metabolites, 1,5-AG/1-deoxyglucose and d-gluconic acid, had AUCs that were greater than those for prediction models including change in HbA_1c, C-oDI, and C-peptide alone (Supplementary Table 10). The addition of changes in these two metabolites also increased the AUC of the full clinical risk model (from 0.82 to 0.87) (Supplementary Table 10). Given the association of N-acetyl-l-methionine with treatment failure in the fully adjusted Cox model, we tested it as a single metabolite predictor in addition to the full clinical model of age, sex, race/ethnicity, baseline BMI, baseline HbA_1c, and maternal diabetes history but found there was minimal improvement in AUC (AUC 0.73 vs. 0.71).

We identified 33 circulating metabolite biomarkers associated with treatment failure in youths with type 2 diabetes after adjusting for age, sex, race/ethnicity, and treatment. Half of these metabolites were associated with clinical glycemia and insulin measures, but several were not, possibly revealing biologic pathways that contribute to the deterioration of glycemic control that are independent of insulin sensitivity and β-cell function. Of these 33 compounds, the association of hexose and N4-acetylcytidine levels differed depending on treatment randomization. These data suggest that metabolites may help guide treatment choices, but this possibility will require further study. We also found that d-gluconic acid, 1,5-AG/1-deoxyglucose, and the composite measurement of glucose/fructose/galactose are dynamic markers of treatment failure. The change in a subset of metabolites provided increased predictive utility even beyond traditional clinical risk factors such as age, sex, BMI, HbA_1c, and maternal diabetes history. These findings support the need for further study of metabolomic markers in youths with type 2 diabetes to facilitate the optimization and individualization of disease treatment.

We have described the circulating metabolome in youths with type 2 diabetes undergoing treatment, with long-term follow-up to ascertain treatment failure. Because blood glucose was used to define treatment failure, it is not surprising that our top associations, for both baseline and change in metabolite, included the positive association of a composite measurement of circulating 6-carbon sugars and the inverse association of the polyol 1,5-AG/1-deoxyglucose, a validated marker of glycemic control (28). Interestingly, several N-acetylated amino acid species (i.e., N-acetyl-l-glycine, N-acetyl-l-histidine, N-acetyl-l-alanine, and N-acetyl-l-methionine) had reductions in hazards for treatment failure that were comparable to 1,5-AG/1-deoxyglucose. Dietary supplementation with N-acetyl-l-methionine led to a smaller decrease in weight compared with l-methionine in rats (29), suggesting that acetylated amino acids participate in similar pathways as their free amino acids, but likely with differential effects. We previously demonstrated that higher levels of N-acetyl-l-alanine were associated with both incident and prevalent diabetes in adults (20). We now demonstrate that this class of compounds is associated with treatment outcomes in youth-onset type 2 diabetes. What remains unclear is whether these associations are due to acetylated amino acids serving as a proxy for glycemic control (17) or whether they represent independent biomarkers of therapeutic response. Many acetylated species are associated with FPG and HbA_1c, which could suggest the former, but because the association of N-acetyl-l-methionine survived adjustment for HbA_1c (Supplementary Table 5), additional studies to clarify this point are warranted.

Pseudouridine, xanthurenic acid, and creatinine (which survived further adjustment for HbA_1c) were also inversely associated with treatment failure. These metabolites and N-acetyl-alanine have been linked with kidney disease (30–32), but there was no difference in eGFR between treatment groups, nor according to treatment failure during the TODAY study. Several organic acids involved in glycolysis and the citric acid cycle also had positive associations with treatment failure, including butyric acid, pyruvic acid, and lactic acid (which also survived additional adjustment for HbA_1c). Furthermore, their baseline levels and change over 36 months were associated with FPG and HbA_1c in TODAY (Supplementary Tables 5 and 6), consistent with prior reports in adults with diabetes (33,34). Baseline butyric acid and changes in pyruvic acid were also associated with changes in C-peptide (Supplementary Table 6), suggesting they may reflect endogenous insulin production and/or insulin resistance.

Of the 33 metabolites associated with treatment failure, hexose and N4-acetylcytidine had differential associations based on treatment group (Table 1). For hexose, youths in the highest quartile at baseline had the highest risk of treatment failure, but the effect was most pronounced in those assigned to MET + ROS or MET + LI. Hexose is traditionally thought to be a surrogate marker of glycemia; however, baseline HbA_1c did not differ across groups (Supplementary Table 1). These findings suggest that hexose may provide additional information outside of clinically measured glycemia on treatment response. By contrast, higher baseline levels of N4-acetylcytidine, a chemically modified nucleoside that has a role in RNA translation and inflammation (35), were associated with lower risk of treatment failure with MET and MET + ROS, but this trend was not seen with MET + LI. Urinary and circulating N4-acetylcytidine have been associated with inflammation (36,37), cancer (38), and gestational diabetes (39,40).

We also assessed the predictive utility of these circulating biomarkers. Models that only included baseline metabolite levels did not add information beyond known clinical risk factors. However, the inclusion of changes in 1,5-AG/1-deoxyglucose and d-gluconic acid outperformed models that included changes in HbA_1c, C-oDI, and C-peptide index and a full clinical model that included age, sex, race/ethnicity, maternal diabetes history, and changes in HbA_1c and BMI (Fig. 3 and Supplementary Table 10). In regression analyses, the changes in these two compounds over 36 months, along with the composite measurement of glucose/fructose/galactose, were the only changes in compounds associated with treatment failure after 36 months after adjusting for age, sex, race/ethnicity, BMI, and HbA_1c (Supplementary Table 7). These findings were not surprising, given all three measures are considered to be surrogate measures of glycemia, and glycemia was used to define the primary end point. This does, however, demonstrate how circulating metabolites can provide additional information beyond traditional clinical measures for disease prognosis and treatment effect. Additionally, we demonstrate that metabolites may predict treatment failure better than clinical measures of β-cell reserve, such as C-peptide index and C-oDI, which are traditionally considered strong predictors of disease progression in youths with type 2 diabetes (16,17).

Strengths of our study include the use of a randomized control cohort that was racially/ethnically diverse, the breadth of circulating compounds profiled, and longitudinal measures of metabolomics and clinical traits. A limitation to the study, however, was the sample size; although it represents the largest metabolomic analysis among youths with type 2 diabetes, the numbers are small for metabolomic profiling, which reduced our power to detect significant associations. This also limited the power for downstream analyses, such as determining treatment-specific biomarkers. Additionally, because of blood sample availability, by chance our metabolomic subcohort did not replicate the differences in treatment failure rates between treatment groups reported in the original TODAY study. However, besides BMI, baseline characteristics between the treatment groups were not different in our metabolomic subcohort, and there was no difference in baseline characteristics, including the number of youths who had treatment failure by 36 months of follow-up, among youths who did or did not undergo metabolomic profiling. Also, in our association analyses of metabolite changes with treatment failure (Supplementary Table 9), we restricted inclusion to youths who did not have treatment failure before 36 months of follow-up (the second metabolomic profiling timepoint), because treatment failure is associated with metabolite levels. This resulted in the exclusion of 41% of the participants who were profiled, limiting the generalizability of the findings for this one analysis. It is also important to note that although participants were randomly assigned to treatment, they were not randomly assigned to metabolite level exposure, so we cannot conclude that the observed associations were causal. Finally, because of the large number of circulating analytes measured, we were at increased risk of false positive findings but have attempted to control for this by using a false discovery rate of <0.05 for significance.

In conclusion, we conducted in-depth metabolomic profiling of circulating small molecules in youths with type 2 diabetes and found that baseline levels of several compounds, including organic acids involved in glycolysis and the citric acid cycle, and the novel class of acetylated amino acids, were associated with treatment failure. These associations may reflect disease pathways independent of glycemia or β-cell function. In addition, the association of hexose and N4-acetylcytidine levels with treatment failure differed based on treatment randomization. Changes in d-gluconic acid and 1,5-AG/1-deoxyglucose over 3 years predicted treatment failure better than changes in HbA_1c, C-oDI, or C-peptide index alone. Together, our data suggest that plasma metabolite signatures are associated with glycemic outcomes in youth-onset type 2 diabetes and are as good at predicting treatment failure as traditional risk factors, such as age, sex, race/ethnicity, BMI, and HbA_1c, and may be better than clinical measures of β-cell function. Additional metabolomic studies in youth-onset type 2 diabetes will be important to confirm these biomarkers and illuminate additional biomarkers of disease progression and therapeutic response in this challenging disease.

Funding. This project was supported by National Institute of Diabetes and Digestive and Kidney Diseases grants U01 DK061230 (E.I.), K23 DK127073 (Z.-Z.C.), T32DK007260 (C.L.), and P30DK036836 (J.M.D. and E.I.).

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. Z.-Z.C., C.L., and E.I. researched data, wrote the manuscript, and contributed to the discussion. J.M.D., G.T., X.S., S.Z., and D.W. researched data. L.P., P.B., L.E.G., and R.E.G. reviewed and edited the manuscript. E.I. and R.E.G. supervised the research activities. Z.-Z.C. and E.I. are the guarantors of this work and, as such, had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.

Handling Editors. The journal editors responsible for overseeing the review of the manuscript were John B. Buse and Casey M. Rebholz.