figure a

Introduction

Type 2 diabetes is epidemic, affecting the health of millions of people worldwide. The number of years that people with type 2 diabetes are living has increased by 32% over the last few decades due to the rise in age-specific prevalence and population growth and ageing [1]. Consequently, type 2 diabetes ranks sixth among leading causes of the burden of disease globally [2]. Previous studies have shown that type 2 diabetes incidence can be prevented or delayed [3, 4] and that people at risk for developing type 2 diabetes can be identified through measuring common clinical risk factors [5].

Higher fasting plasma glucose levels, even in the non-diabetic range, can predict future type 2 diabetes. Individuals with impaired fasting glucose (IFG [5.6–7.0 mmol/l]) have an annual relative risk of type 2 diabetes of 4.7% (95% CI 2.5, 6.9) compared with normoglycaemic individuals [6]. Impaired glucose tolerance (IGT) and elevated HbA1c (39–46 mmol/mol [5.7–6.4%]) are also associated with an increased risk of type 2 diabetes compared with those with completely normal glycaemia [7, 8]. However, around 5–10% of middle-aged individuals of European descent with normal fasting glucose (NFG [<5.55 mmol/l]) develop type 2 diabetes over a 5–10 year period [6,7,8]. The incidence rate is much higher in middle-aged individuals from other ethnic backgrounds; Asian Indians have one of the highest incidence rates of diabetes, with rapid conversion from normoglycaemia to dysglycaemia (19.4% over a 9 year follow-up period) [9].

Prior studies have identified plasma metabolites that are associated with the development of future type 2 diabetes in individuals with both normoglycaemia and prevalent dysglycaemia (IFG and/or IGT) [10,11,12,13,14,15,16,17,18]. Alterations in these metabolites likely signal changes in relevant biological pathways, including amino acid catabolism [10, 11, 13,14,15,16,17,18], lipid oxidation [12, 13, 15, 17] and hexose metabolism [15, 17]. However, in terms of population-level prediction of future type 2 diabetes, identified metabolites have little value beyond clinical risk factors, such as fasting glucose, one of the most robust predictors of future type 2 diabetes [16]. Further, because previous studies were conducted in a mixture of individuals with normoglycaemia and prevalent dysglycaemia, they were unable to discern whether early dysglycaemia preceded changes in metabolite levels or whether identified metabolites were harbingers of early dysglycaemia. We therefore tested the hypothesis that a metabolomics analysis in people with NFG who developed type 2 diabetes could identify new markers and pathways that elucidate early type 2 diabetes pathogenesis and improve prediction of incident type 2 diabetes beyond clinical risk factors.

Methods

Study participants

We included participants from the Framingham Heart Study (FHS) Offspring cohort, a prospective, observational, community-based cohort including 3799 attendees, age 40–65 years, at the fifth quadrennial examination cycle 1991–1995 (baseline examination) [19]. Participants at the fifth and subsequent quadrennial examination cycles underwent a physician-administered physical examination and medical history and routine laboratory tests. For the current analyses, we excluded individuals without profiling of metabolites (n = 1326) and those with prevalent diabetes or cardiovascular events (n = 346), fasting plasma glucose ≥5.6 mmol/l (n = 967) or 2 h glucose ≥11 mmol/l (n = 10). The final study population included 1150 individuals with NFG and no diabetes. All participants provided written informed consent and the study protocol was approved by the Boston University Medical Center Institutional Review Board.

Metabolite profiling

At baseline, after participants had fasted overnight, plasma samples were collected in EDTA, processed immediately and stored at −80°C until assayed. Plasma samples were collected at the fifth quadrennial examination, which took place between 1991 and 1995, and were processed in 2008. A previous study has documented concordance in several metabolite measures between archived samples from the Framingham Offspring Study and freshly obtained samples [20]. Targeted metabolite profiling was performed using liquid chromatography with tandem mass spectrometry (LC-MS/MS) as previously described [11, 12]. Additional details, including accuracy of the methodology used in analyses, calibration and annotation, are provided in the ESM Methods. Metabolites at high missing rate (>20%) were excluded from this analysis, which includes 220 metabolites.

Ascertainment of incident type 2 diabetes

The primary endpoint of this study was incident type 2 diabetes. Incident type 2 diabetes was ascertained during the follow-up at every quadrennial examination and was defined as follows: fasting glucose ≥7 mmol/l, non-fasting blood glucose ≥11 mmol/l or the use of glucose-lowering medications, including insulin. Time to type 2 diabetes incidence was derived from the time of the baseline examination. Chart review was conducted to identify and exclude two participants with type 1 diabetes mellitus.

Clinical covariates

Demographic, lifestyle and clinical characteristics were assessed at baseline. BMI was calculated as weight divided by height squared (kg/m2). The HOMA-IR was calculated [21] and was log-transformed due to a skewed distribution. Total cholesterol, HDL-cholesterol (HDLc) and triacylglycerols (TAGs) were measured, in individuals who had fasted overnight, using standard methods. LDL-cholesterol (LDLc) was indirectly calculated using the Friedewald formula when TAG concentrations were lower than 4.52 mmol/l [22]. We used conventional type 2 diabetes risk factors to estimate risk of new onset of type 2 diabetes for each participant, including sex and parental history of diabetes as categorical variables and age, fasting glucose, BMI, HDLc, TAG and blood pressure as continuous variables. We also considered HOMA-IR and 2 h glucose as continuous variables.

Statistical analysis

Differences in clinical characteristics between participants with and without incident type 2 diabetes were analysed in generalised estimating equations models accounting for familial correlation among participants.

The analytical plan flow-chart for metabolite selection, prediction performance and complementary analyses is summarised in Fig. 1. First, plasma metabolite concentrations were log-transformed and standardised. Next, a random binomial variable was used to split the sample into a testing dataset and a training dataset (4:6) and to serve as an internal validation and avoid inflation of the discrimination estimates. For the retained training dataset (60% of the sample), we conducted least absolute shrinkage and selection operator-penalised regressions (LASSO) with tenfold cross validation to select metabolites predictive of type 2 diabetes incidence based on the criteria giving minimum mean cross-validated error [23]. We then assessed the predictive capability of type 2 diabetes risk factors alone (including age, sex, parental history of diabetes, fasting glucose, BMI, HDLc, TAG and blood pressure) and the predictive capability of type 2 diabetes risk factors plus selected metabolites in the testing set (40% of the sample) by generating the area under the receiver operator characteristic (ROC) curve. We used a nonparametric approach (DeLong’s test) to compare the discriminatory capability of the two highly correlated ROC curves [24]. We repeated this process 100 times and accumulated the selection frequency across 100 iterations for each metabolite separately and used a cut-off of ten selections to prioritise the top predictors of incident type 2 diabetes. Next, we evaluated the capability of the metabolites selected ten or more times in 100 iterations to improve prediction of type 2 diabetes over conventional type 2 diabetes risk factors in the entire cohort. As a sensitivity analysis, we included HOMA-IR and 2 h glucose in the model for type 2 diabetes risk factors and repeated the same methodological approach. These analyses were performed using glmnet (https://cran.r-project.org/web/packages/glmnet/index.html) and pROC (https://cran.r-project.org/web/packages/pROC/index.html) packages implemented in R v3.2.0 program (https://www.r-project.org/). We took two-sided p < 0.05 to denote evidence against the null hypothesis of no type 2 diabetes risk prediction improvement when adding metabolites to the prediction model.

Fig. 1
figure 1

Flow-chart summarising the main steps in the analytical plan. We first split the sample into a testing dataset and a training dataset (at a 4:6 ratio). For the retained training dataset (60% of the sample, red dashed line), we conducted LASSO regression with tenfold cross validation [22] to select metabolites predictive of type 2 diabetes incidence based on the criteria giving minimum mean cross-validated error. Next, we assessed the predictive capability of type 2 diabetes risk factors alone and the predictive capability of type 2 diabetes risk factors plus selected metabolites in the testing set (40% of the sample, blue dashed line) by generating the AUC of the ROC curve in each of the 100 iterations. Next, we evaluated the capability of the final set of metabolites selected ten or more times in 100 iterations to improve prediction of type 2 diabetes when added to a model that included traditional type 2 diabetes risk factors, now using the entire sample. Although we did not have an entirely independent data sample as a testing set, the 95% CIs around the AUC from 100 permutations in the 40% testing set gives the lower bounds that might be expected in testing in an independent sample. Testing predictive improvement in the entire sample also permitted better power for use of several complimentary approaches. These included multivariable-adjusted Cox proportional hazard models to evaluate model robustness, and pathway enrichment analysis and Mendelian randomisation to investigate biological insights and test causality in the metabolite–type 2 diabetes risk association

Full size image

Finally, Cox proportional hazard models were used to investigate the association between prioritised metabolites and type 2 diabetes risk after adjusting for age, sex, BMI, fasting glucose and fasting TAG at baseline. SAS v9.3 (SAS Institute, Cary, NC, USA) was used for the association analyses. We took Bonferroni-corrected threshold for significance at two-sided p < 2.63 × 10−3 (0.05/19 metabolites) to denote evidence against the null hypothesis of no association between prioritised metabolites and type 2 diabetes risk.

Bioinformatics methods

Pathway analysis

We applied pathway enrichment analysis and metabolite set enrichment analysis to identify enriched metabolic pathways using MetaboAnalyst 3.0 [25] for the set of 19 prioritised metabolites. Pathway enrichment analysis at the false discovery rate of 5% was set for significance.

Mendelian randomisation

Mendelian randomisation was conducted for causal inference analyses between components of the nitrogen metabolism pathway and type 2 diabetes risk. Genetic determinants of plasma metabolites were extracted from the MAGNETIC Consortium (n = 24,925) [26]. In the MAGNETIC Consortium, we identified genetic variants associated with glycine and phenylalanine at genome-wide significance (p < 5 × 10−8) (taurine was not available, but all metabolite meta-analysis results are available through www.computationalmedicine.fi/data/NMR_GWAS/), For each independent variant, we gathered summary-level association results for type 2 diabetes from the GoT2D diabetes dataset (www.type2diabetesgenetics.org/projects/got2d; n = 11,645 cases and 32,769 controls) since these variants were not available in other type 2 diabetes genetics consortia [27]. The Mendelian randomisation overall instrumental estimated effect size of the exposure on the outcome, referred to as the inverse-variance weighted (IVW) estimator [28], was calculated using the Genetics ToolboX package (GTX; available at http://cran.r-project.org/web/packages/gtx) (detailed in the ESM Methods). Instrumental heterogeneity was assessed using the Q statistic and reported as a heterogeneity p value. The presence of unbalanced horizontal pleiotropy was assessed by using Mendelian randomisation–Egger when the set of variants in the genetic instrument allowed us to conduct the analysis [29]. We used individual-level data from FHS participants to estimate the variance explained in metabolite levels by the genetic variants. We used genotyped variants with genotyping success rate ≥0.95 and variants in Hardy–Weinberg equilibrium (p > 1 × 10−4). When not directly genotyped, we included variants at high-quality imputation ratio (r2 value threshold of 0.85, representing an approximate correlation with the true genotype higher than 0.9). A linear mixed-effect model with covariates age, sex and random effects to account for familial correlation, including five variants for glycine and three variants for phenylalanine fit individually in an additive genetic model, was used to estimate the variance in plasma metabolite concentrations explained by genetic variants.

Results

Over a 20 year follow-up period, 95 individuals with NFG (8.3%) developed type 2 diabetes. Baseline characteristics according to type 2 diabetes incidence are presented in Table 1. Individuals who developed type 2 diabetes did not differ in age and parental history of type 2 diabetes distribution from those who did not develop type 2 diabetes, but diabetes incidence was higher in men and in individuals who had significantly higher BMI and slightly higher glycaemic trait measurements (fasting and 2 h glucose) at metabolomics sampling. Still, individuals with NFG who progressed to type 2 diabetes were normoglycaemic at baseline, as indicated by 2 h glucose and HbA1c values being in the normal range.

Table 1 Baseline characteristics of participants
Full size table

Overall, 67 metabolites were selected at least once in the training set for type 2 diabetes incidence classification (ESM Table 1). Among them, two metabolites, sphingomyelin C24:0 and diacylglycerol C36:1, were selected in every one of the 100 iterations. The median change in the AUC in the internal validation set upon adding the metabolites selected within each of the 100 iterations to a model that included traditional type 2 diabetes risk factors alone was 0.088 (p = 0.013). Nineteen metabolites were prioritised by LASSO ten or more times in the training dataset (Table 2). The subset of 19 metabolites significantly improved type 2 diabetes prediction when added to a model that included traditional type 2 diabetes risk factors alone using the entire sample (AUC was 0.810 [95% CI 0.77, 0.86] for type 2 diabetes risk factors and 0.902 [95% CI 0.87, 0.94] for type 2 diabetes risk factors + metabolites, p = 1.1 × 10−4) (Fig. 2). In a sensitivity analysis including HOMA-IR and 2 h glucose as additional type 2 diabetes risk factors, metabolites still significantly improved type 2 diabetes prediction (AUC was 0.828 [95% CI 0.78, 0.88] for type 2 diabetes risk factors and 0.906 [95% CI 0.87, 0.94] for type 2 diabetes risk factors + metabolites, p = 2 × 10−4) (ESM Fig. 1).

Table 2 Metabolites prioritised ≥10 times in the training set for type 2 diabetes incidence differentiation
Full size table
Fig. 2
figure 2

ROC curves for models predicting incident type 2 diabetes with and without selected metabolites. ROC curves were derived from a Cox proportional hazard model in the testing dataset. Graphs plot the sensitivity vs (1 – specificity) for diabetes at each possible model cut point. The area under the ROC curve corresponds to the C statistic of that model. Black line, type 2 diabetes risk factors, including age, sex, parental history of diabetes, fasting glucose, BMI, HDLc, TAG and blood pressure; grey line, type 2 diabetes risk factors + LASSO selected metabolites (selected ten or more times in 100 iterations with tenfold cross validation at minimum mean cross-validated error). AUC was 0.81 (95% CI 0.77, 0.86) for the type 2 diabetes risk score and 0.902 (95% CI 0.87, 0.94) for the type 2 diabetes risk score + metabolites (p = 1.1 × 10−4)

Full size image

Next, we used the set of 19 metabolites to identify enriched metabolic pathways. A significant enrichment for association was observed for the nitrogen metabolism pathway at the false discovery rate of 5% (p = 0.047) (Table 3). This pathway is composed of 39 species, three of which (glycine, taurine and phenylalanine) were prioritised by LASSO in the training set ten or more times (ESM Table 2). In separate Cox proportional hazard models for metabolites in the nitrogen metabolism pathway, type 2 diabetes risk was lower per 1 SD increase in plasma glycine (HR 0.65 [95% CI 0.54, 0.78]) and taurine (HR 0.73 [95% CI 0.59, 0.90]) and higher for 1 SD increase in phenylalanine (HR 1.35 [95% CI 1.11, 1.65]) after adjusting for confounders (Table 3). The associations between other prioritised plasma metabolites or conventional risk factors and type 2 diabetes risk is detailed in ESM Tables 3 and 4.

Table 3 Nitrogen metabolism pathway metabolite associations with type 2 diabetes risk
Full size table

Finally, we investigated whether genetically increased metabolites in the nitrogen metabolism pathway have a causal role in type 2 diabetes risk (ESM Table 5). Using the IVW estimator method, we found that for every 1 SD genetically increased glycine, the odds of type 2 diabetes was reduced by 11% (OR 0.89 [95% CI 0.80, 0.99]; heterogeneity p = 0.08, Fig. 3a). The genetic variance in glycine metabolite levels was 11.1% in FHS. The genetic variance attributed to CPS1 was 10% and the allele associated with higher glycine concentrations is also associated with lower risk of type 2 diabetes. The adjusted causal effect estimate was similar when applying the bootstrap method in Mendelian randomisation–Egger regression, showing a trend towards statistical significance (OR 0.87 [95% CI 0.74, 1.00], p = 0.074) (ESM Table 6). The estimate for the intercept in the Mendelian randomisation–Egger regression suggested no evidence of presence of unbalanced pleiotropy (βintercept = 0.01; SE = 0.02; p = 0.118). In a Mendelian randomisation analysis for phenylalanine, which included three phenylalanine risk-increasing variants (variance explained = 16.5% in FHS), the estimate using the IVW estimator method was 1.6 type 2 diabetes higher odds per 1 SD genetically increased phenylalanine (95% CI 1.08, 2.04; heterogeneity p = 0.19) (Fig. 3b). We did not conduct Mendelian randomisation–Egger regression for phenylalanine given the low number of variants in this analysis.

Fig. 3
figure 3

Mendelian randomisation estimate of raised glycine and phenylalanine on type 2 diabetes. Effect of glycine- and phenylalanine-raising genetic variants on type 2 diabetes odds obtained from publicly available data from MAGNETIC Consortium (n = 24,925) and GoT2D (n = 44,414). Each white circle represents a known locus for glycine (a) or phenylalanine (b). The association of each variant with type 2 diabetes [base e logarithmic OR transformation, ln(odds)] is denoted on the y-axis (white circle and grey error bars around the white circles) while the association with glycine and phenylalanine is denoted on the x-axis. The black lines illustrate regression of glycine and phenylalanine on type 2 diabetes and the dashed grey lines show the 95% CI. The effect of the polygenic instrument comprising all five glycine-increasing variants reduced the odds of type 2 diabetes risk per 1 SD increase in metabolite concentration (OR 0.89 [95% CI 0.80, 0.99]). Genetically driven phenylalanine increased odds of type 2 diabetes risk per 1 SD increase in metabolite concentration (OR 1.6 [95% CI 1.08, 2.37]). T2D, type 2 diabetes

Full size image

Discussion

We conducted a population-based prospective study in individuals with NFG at baseline, of whom 95 progressed to type 2 diabetes during the follow-up period of 20 years. Using information from a discrete set of 19 metabolites associated with type 2 diabetes incidence, we improved the capability of predicting incident type 2 diabetes beyond the predictions made using only the clinical risk factors obtained in routine care. A significant biological finding is the enrichment in the nitrogen metabolism pathway with type 2 diabetes risk. Further, the genetic approach provides additional evidence that markers identified in the nitrogen metabolism pathway—glycine and phenylalanine—may be causal rather than only associative, suggesting that alterations in this pathway and its components may contribute to the earliest stages of type 2 diabetes pathogenesis.

While prior clinical metabolomics studies have focused on a mixture of individuals with normoglycaemia and prevalent dysglycaemia [10,11,12,13,14,15,16,17,18], our work is novel in its study of participants who were normoglycaemic yet progressed to type 2 diabetes. By studying individuals who progressed from having normal glucose metabolism to having type 2 diabetes, we eliminate confounding of our results by processes that occur in response to development of dysmetabolism (IFG or IGT). The main clinical finding of the present study is that selected metabolites substantially improved the ability to predict type 2 diabetes, beyond the prediction achieved using conventional risk factors, in individuals classified as normoglycaemic based on fasting glucose. Our findings are slightly different from those of previous metabolomics studies, which were not able to show that metabolites materially improved type 2 diabetes risk prediction over other clinical risk factors [11, 15, 16]. A possible explanation might be related to the study population: in normoglycaemic individuals, blood glucose and other traditional type 2 diabetes risk factors may not be as strong predictors of future type 2 diabetes as in those with dysglycaemia [30]. Thus, novel risk factors like metabolites could show stronger predictive capability. Nevertheless, the AUC of clinical risk factors in normoglycaemic individuals was still >80%, slightly higher than reported in other recent studies [31, 32]. This suggests that differences in the length of follow-up or the inclusion of different ethnic groups could affect the predictive capability of traditional type 2 diabetes risk factors. Another possible explanation for the increased predictive ability of metabolites might be related to the methodological approach implemented in this study. The predictive capability observed in this study is aligned with prediction performance observed in recent studies using similar machine learning approaches to prioritise metabolites [31, 32].

The normoglycaemic individuals included in this study were selected not only because of their normal fasting glucose but also because of their normal glucose tolerance and HbA1c, under current accepted definitions. Although different, perhaps more stringent, thresholds for ‘normal’ might have been chosen (especially as there were subtle elevations in fasting glucose among those who developed type 2 diabetes vs those who did not), we think that the data provide insight into early type 2 diabetes pathogenesis among those currently considered clinically normoglycaemic. The convergence of selected metabolites in the nitrogen metabolism pathway, serving as nitrogen donors for the urea cycle [33], suggests that this pathway may influence the early pathogenesis of type 2 diabetes. Data from the Mendelian randomisation experiments further support the notion that components within this pathway may have a causal role in type 2 diabetes development; therefore, confounding effects by obesity or lipid abnormalities are less likely. In our study, genetically increased glycine reduced the odds of type 2 diabetes, consistent with findings from previous epidemiological studies in terms of directionality and effect sizes [15, 34]. However, this conflicts with a previous Mendelian randomisation analysis that showed no association between a single genetic variant for glycine or glycine-to-serine ratio and diabetes-related traits [35]. More precisely estimated effect sizes derived using data from the largest metabolites meta-genome-wide association studies currently available and the increase in the number of genetic variants used to proxy glycine here are likely to explain the difference between these two Mendelian randomisation studies. With regard to phenylalanine, different studies have reported a direct association between this metabolite and type 2 diabetes risk [11, 15, 34], although no Mendelian randomisation analysis for phenylalanine on type 2 diabetes risk has been conducted yet.

In the present study, we also documented that particular biomarkers previously associated with type 2 diabetes risk (e.g. TAGs with lower carbon number or 2-aminoadipate [12, 14]) are likely to be relevant even before initial glycaemic perturbations. In contrast, other metabolites such as branched chain amino acids, associated with type 2 diabetes incidence in previous studies, were not prioritised by our selection algorithm. Notably, our findings, which highlight early metabolite changes in the pathogenesis of type 2 diabetes, are consistent with a recent Mendelian randomisation analysis finding that elevations in branched chain amino acids occur after the development of insulin resistance [36].

We acknowledge that the results of our population-based analysis should be interpreted with caution since several limitations such as unmeasured factors (e.g. changes in lifestyle factors, medications or insulin secretion and resistance over time) might have influenced our findings. Using a Mendelian randomisation approach for available metabolites in the nitrogen metabolism pathway partially mitigates this concern, suggesting that genetically driven glycine and phenylalanine are indeed related to the risk of type 2 diabetes independently of potential confounders. While our results were internally validated, we recognise that they were not confirmed in a separate prospective cohort. The lack of independent validation is due to the lack of availability of comparable cohorts of normoglycaemic individuals who developed type 2 diabetes for whom the necessary data were available. However, the internal validation approach, using 40% of the sample and running 100 iterations, allowed us to rule out other potential conflicting issues, such as compatibility between metabolomics platforms or available standards in libraries, even when similar platforms were used. Five TAGs were prioritised by our methodological approach but we did not find significant enrichment of pathways associated with these species. This might be because the software we used for pathway enrichment analysis may provide poor reporting for lipid classes and pathways or because only five TAGs in a particular metabolic pathway are likely to be less than expected by chance for enrichment. In addition, most prioritised metabolites correlate with baseline clinical risk factors such as BMI, 2 h glucose, HOMA-IR, HDLc and TAGs (ESM Table 7) but associations of metabolites with type 2 diabetes in normoglycaemic individuals remained after risk factor adjustment. Last, we recognise that participants in this study were all of European descent. Further work is needed to determine whether our findings can be replicated in an independent cohort of the same ethnicity and to extend the study to other racial/ethnic groups.

In conclusion, our study identifies a discrete set of metabolites that signal increased risk for type 2 diabetes among normoglycaemic individuals; these metabolites are involved and may play a causal role in the early stages of type 2 diabetes pathogenesis.