Introduction

Prediabetes, an intermediate stage of glucose dysregulation that blood glucose levels are elevated but lower than in diabetes, has become a burgeoning global health emergency¹. Prediabetes affected approximately 720 million individuals worldwide in 2021, with a project to 1 billion people by 2045². Approximately 5% to 10% of people with prediabetes progress to having diabetes each year and the lifetime conversion rate to diabetes could be as high as 70%^3,4. Therefore, preventing or delaying diabetes development among people with prediabetes will have substantial clinical and public health benefits.

Although lifestyle modification and medical therapy have been proven to be effective in preventing or delaying the diabetes onset among people with prediabetes^5–7, the substantial cost of modification programs and medications as well as drug-related side effects limit the widespread delivery of such interventions in this large high-risk population^8,9. Notably, the progression from prediabetes to diabetes is highly heterogeneous, and a fraction of individuals with prediabetes may regress to normoglycemia without treatment.¹⁰ Therefore, identifying targeted population who are at high risk of developing diabetes is the key step to tailor precise and efficient interventions. Glycemic indicators alone for risk stratification are deficient, with fasting glucose and glycosylated hemoglobin A1c (HbA1c) being convenient but less sensitive, while post-load glucose tolerance being sensitive but unfeasible in practice on a large scale^11,12. In addition, several risk assessment models based on conventional clinical variables have been developed, but most of which had comparatively low performance and failed to take follow-up time into account^13–15.

Plasma metabolomics using high-throughput techniques could provide a comprehensive profiling of small-molecule metabolites in a specific physiological period, which might yield valuable information for risk prediction. Previous studies have implied that incorporating circulating metabolites into basic models with conventional risk factors could improve prediction of diabetes risk ^16–18. However, we are aware of only one study that has assessed the relationship between metabolomic profiling and the progression to diabetes among individuals with prediabetes and investigated the predictive values of metabolites¹⁹. Nevertheless, it was limited by a nested case-control study design with a relatively short follow-up (median 5 years) and small sample size (n=∼300). Whether addition of metabolic biomarkers improves the ability in predicting the progression from prediabetes to diabetes in prospective settings remains largely unknown.

To address these knowledge gaps, in the current study we aimed to examine the longitudinal associations of circulating metabolic biomarkers, quantified using high-throughput nuclear magnetic resonance (NMR), with the risk of incident diabetes among individuals with prediabetes from the UK Biobank. Moreover, we evaluated whether metabolic signature adds anything to prediction models for diabetes development and risk stratification.

Methods

Study design and participants

The UK Biobank is a large population-based prospective cohort study enrolling more than 500,000 community-dwelling adults from 22 assessment centers across the UK between 2006 and 2010^20,21. Participants completed touchscreen questionnaires and physical measurements and provided blood samples at baseline. The study was approved by the Northwest Multicenter Research Ethics Committee (REC reference for UK Biobank 11/NW/0382), and all participants provided informed consent.

For the identification of metabolomic biomarkers associated with the progression from prediabetes to diabetes, the current study focused on participants with prediabetes at baseline with available circulating metabolite data. The diagnosis of prediabetes was defined by an HbA1c level of 5.7% to 6.4% (39 to 47 mmol/mol) in participants without diabetes, according to the American Diabetes Association (ADA) criteria²². After excluding individuals who developed diabetes or died within 1 month from the baseline, 13,489 participants with prediabetes were included in the final analyses.

Metabolite quantification

The metabolomics analysis of approximately 118,000 non-fasting ethylenediaminetetraacetic acid (EDTA) plasma samples at baseline were performed using the high-throughput NMR platform in Nightingale Health’s laboratories of Finland. Details of the metabolic profiling platform and experimentation have been described elsewhere^23–25. In brief, the EDTA samples were collected and stored at -80°C. Before preparation, frozen samples were slowly thawed at +4°C overnight and were centrifuged (3,400 g) for 3 minutes. Each sample was analyzed with a spectrometer and the metabolic biomarkers were quantified using Nightingale Health’s proprietary software. The quality control procedures were implemented during the whole process and only samples and biomarkers that underwent the quality control process were stored in the UK Biobank dataset. The coefficient of variations was below 5% for most of the biomarkers and there were no batch effects²⁶.

A total of 249 metabolic biomarkers (168 directly measured and 81 ratios of these), spanning lipids, lipoprotein subclass, fatty acids, amino acids, ketone bodies, and glycolysis metabolites were quantified for each sample. In the present study, we analyzed 168 metabolic biomarkers that were directly measured (Supplementary Table S1). The values of all metabolites were transformed using natural logarithmic transformation (ln[x+1]) followed by Z-transformation.

Covariate collection

Information on covariates was collected through a self-completed touchscreen questionnaire or verbal interview, including age, sex, ethnicity (white people or others), Townsend deprivation index, household income (high: ≥ £52,000, medium: £18,000-£51,999, and low: < £18,000), education (college/university degree or others), employment status (current working, retired, or other), smoking status (never, previous, or current smoking), moderate alcohol (alcohol intake > 0 g and ≤ 14 g/day for women and alcohol intake > 0 g and ≤ 28 g/day for men), physical activity, healthy diet score, healthy sleep score, family history of diabetes (yes or no), history of cardiovascular disease (CVD, yes or no), history of hypertension (yes or no), history of dyslipidemia (yes or no), history of chronic lung diseases (CLD, e.g. chronic bronchitis, emphysema, and chronic obstructive pulmonary disease, yes or no), and history of cancer (yes or no). The Townsend deprivation index is a composite measure of area-level socioeconomic deprivation, with a higher score indicating higher levels of socioeconomic deprivation. Physical activity was measured by the metabolic equivalent task (MET) (sum of days performing walking, moderate activity, and vigorous activity)²⁷. A healthy diet score was calculated based on the intake of vegetables (≥ median), fruits (≥ median), fish (≥ median), red meat (< median), and processed red meat (< median)²⁸. One point was given for each favorable diet factor and the total diet score ranges from 0 to 5. A healthy sleep score was evaluated based on insomnia (sometimes or never), sleep duration (7∼8 h), chronotype (morning person), daytime sleepiness (sometimes or never), and snoring (no)²⁹. Each favorable sleep factor was given a score of 1, with the total sleep score ranging from 0 to 5.

Physical measurements including blood pressure, height, weight, waist circumference (WC), and hip circumference (HC) were measured using calibrated instruments with standard protocols by trained nurses. Blood pressure was measured using the Omron automatic digital monitor and two measurements were obtained at a few minutes’ intervals. We calculated the mean systolic (SBP) and diastolic blood pressure (DBP) from two measurements. Body mass index (BMI) was calculated as weight in kilograms divided by the square of height in meters (kg/m²). The HbA1c level was measured by high-performance liquid chromatography with the VARIANT II Turbo analyzer (Bio-Rad Laboratories). Missing covariates were imputed by the median value for continuous variables and a missing indicator for categorical variables.

Ascertainment of diabetes

Incident diabetes was ascertained from hospital inpatient records, death registers, and primary care records, according to the International Classification of Diseases, 10th revision (ICD-10) codes. Detailed information about the linkage procedure is available from https://content.digital.nhs.uk/services. The follow-up time was calculated from the baseline to the occurrence of diabetes, death, or the censoring date (30 March 2023), whichever came first.

Statistical analyses

Baseline characteristics were presented as numbers (percentages) for categorical variables and means (standard deviations, SDs) for continuous variables, respectively. Continuous variables were assessed for statistical differences using t-test and categorical variables were evaluated using the χ² test. Overall schematic workflow of the study is shown in Graphical abstract.

Metabolite selection

We first used Cox proportional hazards (CPH) model to assess the associations between individual metabolites and risk of diabetes progression with adjustment for sociodemographic covariates (age, sex, ethnicity, education, Townsend Deprivation Index, employment status and household income), family history of diabetes, health conditions (history of CVD, hypertension, dyslipidemia, CLD and cancer), physical measurements (BMI, WC, HC, SBP and DBP), lifestyle factors (smoking status, moderate alcohol, healthy diet score, healthy sleep score and physical activity), and HbA1c. The potential confounders were selected based on prior knowledge of the risk factors for diabetes. Metabolites that were significantly associated with incident diabetes (P <0.05/168) were retained.

Secondly, we performed priority-Lasso to deal with multicollinearity in high dimensional data and to retain variables with nonzero coefficients. Priority-Lasso is a Lasso-based intuitive analysis strategy, which uses prior knowledge regarding the outcome by defining the blocks of different types of predictor variables³⁰. In this study, we defined the 24 covariates as block 1, while all metabolites significantly associated with diabetes risk in the CPH model were defined as block 2. The penalization parameter λ was determined as values with maximum partial-likelihood in a 10-fold cross-validation.

Thirdly, three machine learning models including supporting vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost) were adopted to further evaluate the importance of the Lasso-selected metabolites, as they can model nonlinear and nonadditive relations more flexibly³¹. Models were built by 10-fold cross-validation through the “caret” package. Common signals detected across diverse approaches are more likely to represent the strongest and true patterns in the data. We chose the intersection set of the top 20 most important variables selected by the three machine learning models.

Model development

Participants were randomly subclassified into a training set and a test set at a ratio of 8:2 and two common algorithms for survival data including CPH model and random survival forest (RSF)³² were adopted for model development. RSF, as a machine learning method, is designed to be used specifically for survival outcome prediction and has shown promising results in various settings^33,34. It builds many decision trees using split points based on the log-rank test to identify different survival statuses and produces the predicted probability for an individual derived from the average prediction across all trees³⁵. The RSF model was fitted using the “randomForestSRC” package and the gird search method was used for hyperparameter tuning (number of trees, number of variables to possibly split at each node, and minimum size of terminal node).

Model evaluation

The model performance was assessed in the test set. The time-dependent area under the receiver operating characteristic curve (AUROC) was used to evaluate the model’s discrimination ability. Continuous net reclassification improvement (NRI), and absolute integrated discrimination improvement (IDI) were used to assess whether adding the selected metabolites could improve risk discrimination and reclassification for the risk of progression from prediabetes to diabetes over the basic model that was built on 10 conventional clinical variables (age, sex, Townsend Deprivation Index, family history of diabetes mellitus, BMI, WC, HC, SBP, DBP, and HbA1c)³⁶. The calibration ability of the model was estimated using calibration curve.

Furthermore, we used decision curve analysis (DCA) to assess the clinical usefulness of prediction model-based guidance for prediabetes management. To facilitate risk stratification, we classified participants into two risk groups according to the predictive value using “surv_cutpoint” function in the “survminer” R package³⁷. We also divided participants into three categories according to the tertiles of probability. In addition, we included 90,688 participants with normal glucose from the UK Biobank and divided them into the training and test sets using an 8:2 ratio to further investigate the additive value of the selected metabolites in diabetes prediction among participants with normoglycemia. All analyses were conducted in R software (version 4.2.2). A two-sided P value < 0.05 was considered statistically significant. To control for the false discovery rate in the association between multiple metabolic biomarkers and incident diabetes, Bonferroni correction for P value (P <0.05/168) was used.

Results

Baseline characteristics

Among the 13,489 participants with baseline prediabetes, the mean age was 59.6 (SD, 7.1) years, and 6,166 (45.7%) were males. During a median follow-up of 13.6 (12.3-14.6) years, 2,525 (18.7%) participants progressed to diabetes. Baseline characteristics of the study population stratified by incident diabetes are summarized in Table 1. Participants who developed diabetes were more likely to be male, non-White, less educated, more deprived, and smokers. They also tended to have a family history of diabetes, comorbidities such as CVD, hypertension, dyslipidemia and CLD, and higher levels of BMI, WC, and HC.

Table 1.

Identification of metabolic biomarkers for progression to diabetes

After adjusting for covariates and correcting for multiple testing, 94 of 168 metabolic biomarkers were significantly associated with the risk of incident diabetes (Figure 1, Table S2). Concentrations of very low-density lipoprotein (VLDL) particles, particularly larger VLDL particles and composition within larger VLDL, were strongly associated with progression to diabetes. Triglyceride in all lipoprotein subclasses also demonstrated strong positive associations with diabetes risk. In contrast, concentrations of larger high-density lipoprotein (HDL) particles and composition within these particles were inversely associated with incident diabetes. For lipoprotein particle diameter, larger HDL and LDL particle sizes were associated with a lower risk of progression to diabetes, while larger VLDL particle size was associated with a higher risk.

Figure 1.

Monounsaturated fatty acids and saturated fatty acids were positively associated with the risk of diabetes, whereas docosahexaenoic acid and the degree of fatty acid unsaturation were negatively associated with diabetes. Among the amino acids, higher concentrations of alanine, tyrosine, and branched-chain amino acid (BCAA) such as leucine and valine were associated with an increased risk of diabetes, but glutamine and glycine were inversely associated with diabetes. Neither of the ketone bodies showed an association with the risk of diabetes.

Of the 94 metabolites that were significantly associated with diabetes, 17 metabolites were selected by priority-Lasso (Table S3). When further evaluating the importance of these metabolites after adjustment for covariates using three machine learning algorithms, the intersection of the top 20 important predictors identified a total of 9 metabolites, namely cholesteryl esters in large HDL, cholesteryl esters in medium VLDL, triglycerides in very large VLDL, average diameter for LDL particles, triglycerides in IDL, glycine, tyrosine, glucose, and docosahexaenoic acid (Figure 2, Table S4).

Figure 2.

Model development and evaluation

Build upon the selected 9 metabolites and 10 clinical variables, there was no obvious difference in the AUROC obtained from CPH model (1-year: 0.823 [95% confidence interval, CI 0.702, 0.945]; 5-year: 0.830 [0.797, 0.864]; 10-year: 0.801 [0.778, 0.825]) and RSF model (1-year: 0.828 [0.723, 0.933]; 5-year: 0.820 [0.785, 0.855]; 10-year: 0.802 [0.778, 0.826]). Hence, we chose CPH model as the final model because of its simplicity and interpretability. The addition of selected metabolites consecutively outperformed the basic model with conventional clinical variables in diabetes risk prediction from 1 to 10 years (Figure S1). Specifically, the AUROC increased from 0.759 (95% CI 0.608, 0.911) to 0.823 (0.702, 0.945), 0.798 (0.762, 0.834) to 0.830 (0.797, 0.864), and 0.776 (0.750, 0.801) to 0.801 (0.778, 0.825) for 1-year, 5-year, and 10-year diabetes risk, respectively (Table 2, Figure S2). Results from continuous NRI and absolute IDI also demonstrated improvement in the risk prediction for progression to diabetes (Table 2), although the model calibration was not significantly improved (Figure S3). The decision curve analysis showed that the inclusion of the metabolites had a higher net benefit across the threshold probabilities of 0-0.35 for predicting 5-year diabetes risk and 0-0.55 for predicting 10-year diabetes risk (Figure S4).

Table 2.

We further categorized the participants from the test set into low-risk and high-risk groups according to the optimal threshold of the predicted value (1.02) reflecting the best risk difference. Compared with the low-risk group, participants in the high-risk group had a significantly higher cumulative risk of incident diabetes (log-rank P < 0.0001) (Figure 3). When participants were alternatively classified into low-risk, medium-risk, and high-risk groups according to the tertile cut-off point of the predicted value, the high-risk group showed the highest risk of developing diabetes, followed by the medium-risk and low-risk groups (log-rank P < 0.0001). In addition, the predicted risk of diabetes within 1 year (P = 0.001), 5 years (P < 0.001), or 10 years (P < 0.001) was generally higher among participants who progressed to diabetes than those who did not (Figure 4).

Figure 3.

Figure 4.

Among participants with normoglycemia, we also observed a significant improvement in the prediction of diabetes after the addition of metabolic biomarkers to the basic model. The AUROC increased from 0.821 (95% CI 0.736, 0.907) to 0.868 (0.802, 0.934), 0.790 (0.738, 0.842) to 0.811 (0.762, 0.860), and 0.791 (0.765, 0.816) to 0.806 (0.781, 0.831) for 1-year, 5-year, and 10-year diabetes risk, respectively. (Table S5). The increases in NRI and IDI were similar to or slightly lower than those found among participants with prediabetes.

Discussion

By leveraging data from the large UK Biobank cohort, this prospective study provided a comprehensive analysis of the associations of circulating metabolites with the risk of progression to diabetes and predictive ability in participants with prediabetes. We found that lipoprotein particles, lipoprotein particle size and composition, fatty acids, and amino acids were associated with the risk of incident diabetes. More importantly, our findings suggested that adding the selected metabolites could significantly improve the risk prediction of progression from prediabetes to diabetes beyond the conventional clinical variables.

In the present study, the association between diabetes risk and lipid and lipoprotein profile, including VLDL particles and composition with larger VLDL, HDL particles and composition within larger HDL, triglyceride, smaller HDL and LDL particle sizes, and larger VLDL particle sizes, were broadly consistent with previous studies in the general population^38–41. BCAAs have been widely reported to be involved in the pathogenesis of diabetes, which might impair insulin signaling and lead to increased insulin secretion and pancreatic β-cell exhaustion⁴². Furthermore, genetic association studies have shown higher BCAAS resulting from insulin resistance, which may in turn cause diabetes^43,44. Our study confirmed the vital role of these metabolites in the progression to diabetes among individuals with prediabetes.

Several risk assessment models for predicting the risk of progression from prediabetes to diabetes have been reported^13–15. Yokota et al. developed a logistic regression model to predict the risk for conversion from prediabetes to diabetes based on family history of diabetes, sex, SBP, fasting plasma glucose (FPG), HbA1c, and alanine aminotransferase (ALT)¹³. The model derived from a retrospective longitudinal study design achieved an AUROC of 0.80 (0.70–0.87) but did not take follow-up time into account. Similarly, Liang et al developed a predictive model using three glycemic indicators (FPG, 2-h postprandial blood glucose [2-hPG], and HbA1c) alone¹⁵ and obtained a relatively low AUROC of 0.732 (95% CI 0.688-0.776). In a cohort study of 852,454 individuals with prediabetes, a machine-learning model predicting the progression to diabetes within 1-year was established using data from electronic medical records¹⁴. The model built on age, gender, BMI, medication usage, and laboratory results achieved a high AUROC of 0.865 (0.860-0.869). However, the model’s performance over a longer follow-up period was unclear and conventional parameters such as lifestyle, family history of diabetes or comorbidities were not taken into account.

Changes in circulating small-molecule metabolites may occur long before the disease onset. Although rapid development in the technology of metabolomics provides a powerful tool for precise disease prediction, few studies have investigated the role of metabolomics-derived metabolic biomarkers in predicting progression from prediabetes to diabetes. To our best knowledge, only one case-control study among 153 individuals with prediabetes and 160 matched controls reported that adding 13 metabolites to conventional clinical variables including BMI, waist-hip ratio, WC, SBP, DBP, triglyceride, LDL, and triglyceride-glucose index improved the risk prediction of diabetes progression within 5 years, with the AUROC increasing from 0.72 to 0.98¹⁹. However, the predictive ability of metabolites in prospective settings with large sample size remains uncertain. In this longitudinal study among 13,489 participants with prediabetes, we comprehensively used multiple machine learning algorithms to identify a panel of 9 circulating metabolites that were associated with diabetes incidence during a median follow-up of 13.6 years. The CPH model integrating conventional clinical variables and the selected metabolic signature achieved a comparatively high AUROC of 0.823, 0.830, and 0.801 for 1-year, 5-year, and 10-year diabetes risk, respectively. Importantly, the addition of the metabolites resulted in a significant improvement in the discrimination ability and risk reclassification of diabetes beyond conventional risk factors. Furthermore, we categorized participants according to the optimal threshold points of the predicted value and found that the high-risk group had a significantly higher cumulative incidence of diabetes than the low-risk group. These results provided novel evidence supporting the value of metabolic biomarkers in risk prediction and stratification for the progression from prediabetes to diabetes. Considering the epidemic proportion of prediabetes worldwide, even a modest improvement in diabetes risk prediction among individuals with prediabetes will have substantial clinical and public health implications. Early detection of individuals with prediabetes who are at high risk of developing diabetes would not only advance targeted screening initiatives, health management and interventions but also facilitate a rational allocation of medical resources while avoiding disproportionate healthcare expenditure, which could finally translate into precise and efficient prevention of diabetes. The value of the selected metabolic biomarkers in diabetes prediction was also confirmed in individuals with normal glucose.

Strengths of our study included large sample size, prospective study design with long-term follow-up, and comprehensive control of covariates. Moreover, we used multiple machine learning algorithms to identify the consistently important metabolic biomarkers based on which we developed the predictive models. The final model exhibited relatively high performance for 1-year, 5-year, and 10-year diabetes risk prediction. However, several limitations of our study should be noted. First, since FPG and 2-hPG were not available in the UK Biobank, we defined prediabetes using HbA1c alone and to what extent our results could be extrapolated to other people with prediabetes determined by multiple glycemic indicators requires further investigation. Second, circulating metabolites were measured at baseline, thus their dynamic change over time could not be captured. However, our models showed stable performance in predicting short-term and long-term progression to diabetes (1 to 10 years), indicating the validity of single measurements of metabolic biomarkers for risk prediction. Third, although incident diabetes cases were ascertained through different data sources, including hospital inpatient records, death registers, and primary care records, some undiagnosed diabetes might have been missed. This misclassification would underestimate the effect of the observed associations between metabolites and diabetes risk. Fourth, in this study the prediction models were established and tested using the UK Biobank dataset, external validation in an independent cohort is warranted to confirm the predictive values of the metabolic biomarkers. Finally, the participants from the UK Biobank were mostly White, which might limit the generalizability of the findings to other populations.

Conclusions

In this large prospective study among individuals with prediabetes, we detected a panel of circulating metabolites that were associated with incident diabetes risk. Use of these metabolites significantly improved the risk prediction of progression from prediabetes to diabetes. Our findings provide evidence that integrating metabolite markers with conventional risk factors is a promising approach to advance effective screening strategies and precise interventions for individuals with prediabetes who are at high risk of developing diabetes.

Availability of data and materials

The data analyzed during this study are available at https://www.ukbiobank.ac.uk/. This research has been conducted using the UK Biobank Resource under application number 77740.

Competing interests

The authors declare no competing interests.

Data Availability

The data analyzed during this study are available at https://www.ukbiobank.ac.uk/. This research has been conducted using the UK Biobank Resource under application number 77740.

Acknowledgements

We thank all participants and staff in the UK Biobank for their dedication and contribution to this study.

Authors’ contributions

B.W. and Y.L. conceived and designed the study. J.L. and Y.Y. performed the statistical analysis and drafted the manuscript. Y.S., Y.F., W.S., and L.C. participated in data collection. B.W., Y.L., X.T., N.W. critically revised the manuscript. All authors read and approved the final manuscript. B.W. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.