Type 2 diabetes is a progressive multifactorial disease which presently affects >400 m worldwide, with numbers expected to increase to >700 m by 2045. Biomarkers for the disease, which provide a deeper understanding of the disease process, are therefore eagerly sought. Importantly, their identification may improve prediction and personalized approaches to disease treatment.
Whilst many studies have examined associations between circulating biomarkers and incident disease, to date few studies have explored changes associated with glycaemic deterioration after the development of diabetes. Published studies have established that faster glycaemic deterioration is seen in those who are diagnosed younger, are more obese at diagnosis, have lower HDL, and higher HbA1c. A few studies have investigated genetic variants associated with more rapid progression with small and variable results, although a report on a Hong Kong Chinese population reported a replicated finding that a high polygenic risk score consisting of 123 T2D risk variants was associated with increased progression to insulin requirement. To date, no studies that have adopted a multi-omic approach to biomarker discovery, or reported systematically how metabolites of different classes impact on progression. Such associations have the potential to be clinically useful in terms of prediction, as well as providing biological insights into the processes that drive glycaemic deterioration in T2D.
In a collaboration based around the EU Innovative Medicines Initiative-2 Risk Assessment and ProgreSsiOn of Diabetes (RHAPSODY) we have undertaken here to identify, in three large European cohorts, biomarkers of diabetes progression of three molecular classes: charged small molecules (metabolites), lipids and proteins. In this way, we identify species and, in the case of two of the identified proteins, provide evidence through functional studies in preclinical models for previously unidentified mechanisms of action in disease-relevant tissues.
Individuals from three cohorts, DCS, GoDARTS and ANDIS were included. In a subset, molecular characterization was performed of which characteristics are shown in Table S1. The characteristics across the cohorts were comparable (Table S1). Male subjects were more abundant in the cohorts (>55%), and the average age ranged from 61–67 years with a BMI of 30–32 kg/m 2. Glycated haemoglobin (HbA1c) levels were on average lowest in DCS (median 47.08 mmol/mol), followed by GoDARTS (55.54 mmol/mol) and ANDIS (60.06 mmol/mol). The time from diagnosis to sampling time ranged from 0 to 2.63 years. Three phenotypic models were explored in the included cohorts which showed concordance with BMI, use of glucose-lowering drugs being risk factors and age, HDL and C-peptide being protective (Table S2).
Out of the 19 small metabolites examined, five were associated with disease progression with nominal significance in the base model (age, sex, BMI adjusted, P< 0.05) in the meta-analysis of three cohorts (Fig. 1). These were homocitrulline (Hcit), aminoadipic acid (AADA), isoleucine (Ile), glycocholic acid (GCA), taurocholic acid (TCA). Out of the five, the association of two remained significant after multiple testing adjustment, including aminoadipic acid (AADA, HR = 1.11, 95% CI = 1.01–1.22, p FDR = 0.03) and homocitrulline (Hcit, HR = 1.12, 95% CI = 1.00–1.25, p FDR = 0.04, Fig. 1, Table S3). Of note, for AADA higher levels were observed at baseline for incident insulin users versus non-insulin users, but not for homocitrulline (Supplementary Fig. 2). Furthermore, homocitrulline showed a modest interaction with BMI (P = 0.03) which could, to some extent, mask the differences in levels at baseline. For AADA, however, an interaction with C-peptide was observed (P = 0.01). Both metabolites showed associations in the same direction in the replication cohorts, but non-significant with attenuated effect sizes (AADA, HR = 1.03, 95% CI = 0.96–1.11; Hcit HR = 1.03, 95% CI = 0.88–1.21). In external validation cohorts, Hcit showed a trend as a risk factor for incident diabetes (HR = 1.05, 95% CI = 0.74–1.48) in MDC. Based on a logistic model in DESIR, Hcit was a risk factor for prevalent diabetes (OR = 1.32, 95% CI = 1.05–1.66), but not incident diabetes (HR = 0.97, 95% CI = 0.73–1.30). AADA has previously been associated with a higher risk of incident type 2 diabetes in Wang et al. (OR = 1.60, 95% CI = 1.19–2.16). Finally, the most consistent risk factor for time to insulin was isoleucine level, which was nominally significant in the discovery cohort (HR = 1.09, 95% CI = 0.96–1.25), a risk factor for incident diabetes in MDC (HR = 1.48, 95% CI = 1.26–1.74) and DESIR (OR = 23.88, 95% CI = 3.13–182.31) as well as prevalent diabetes (OR = 10.94, 95% CI = 3.94,30.32). In addition, isoleucine levels showed a modest interaction with BMI (P = 0.02). Finally, GCA and TCA were modest risk factors for time to insulin requirement, with hazard ratios of 1.09 (95% CI = 0.91–1.31) and 1.06 (95% CI 0.99–1.15), respectively. In the replication set both TCA and GCA were in the same direction, but no longer significant with hazard ratios of 1.09 (95% CI = 0.91–1.31) and 1.04 (95% CI = 0.94,1.12).
a Hazards of a time to insulin model in the three discovery cohorts plus two replication sets in two of three discovery cohorts and their respective meta-analyses (Model 1). The figure shows the five nominally significant metabolites, with Hcit and AADA being also significant after multiple testing. Data are presented as hazard ratios with 95% confidence intervals. N = 1,267 individuals for DCS, n = 897 individuals for GoDARTS discovery, n = 699 individuals for GoDARTS validation, n = 811 individuals for ANDIS discovery, n = 1969 individuals for ANDIS validation. b Hazards of incident diabetes in MDC based on a Cox proportional hazards model adjusted for age, sex, and BMI. Data are presented as hazard ratios with 95% confidence intervals. N = 3423 individuals c Odds ratios of incident and prevalent diabetes in DESIR based on a logistic regression model adjusted for age, sex and BMI. Data are presented as odds ratios with 95% confidence intervals. N = 1087 individuals for DESIR.
a Hazards of a time to insulin model in the three discovery cohorts and the meta-analysed hazards (Model 1). The figure shows the nine significant lipids after multiple testing. Data are presented as hazard ratios with 95% confidence intervals. N = 900 individuals for DCS, n = 899 individuals for GoDARTS, n = 809 for ANDIS. b Hazard models of incident diabetes in MDC based on a Cox proportional hazards model. Data are presented as hazard ratios with 95% confidence intervals. N = 3667 individuals.
Among the 162 lipids investigated, the levels of nine reached significance in the base model (Fig. 2). Among these eight lipids were a risk factor for early insulin requirement, and these were all triglycerides (Fig. 2, Supplementary Data 1). These eight lipids were also a risk factor for incident diabetes in MDC (Fig. 2). A single lipid was protective for early insulin initiation (SM 42:2;2, HR = 0.85, 95% CI = 0.73–0.99). Interestingly, SM 42:2;2 was a risk factor for incident diabetes in MDC (HR = 1.16, 95% CI = 1.06–1.27). Further adjustment in the discovery cohort attenuated the effect size but the direction remained the same. Furthermore, in the partly (HDL, C-peptide) and fully adjusted model (additional adjustment for diabetes duration, glucose-lowering drugs) four and three lipids remained significant, respectively (Supplementary Data 1). At baseline, the levels of TAGs were higher in incident insulin users versus non-insulin users (Supplementary Fig. 3). In line with the protective hazard ratio, the levels of SM 42:2;2 were lower in the incident insulin users versus non-insulin users (Supplementary Fig. 3). As observed previously by us based on a previous report, TAG acyl chain length and number of double bonds determined the magnitude of effect of TAGs. In this study, we also observe an almost linear relation between the acyl chain length and the number of double bonds and the hazard ratio, where the highest hazard was observed for the TAGs with the shortest acyl chains and the lowest number of double bonds (Supplementary Fig. 4). Nonetheless, the levels of TAGs were strongly correlated among each other (Supplementary Fig. 5).
In the 1195 investigated plasma proteins, the levels of 98 were nominally associated with time to insulin in the base model. Additional adjustment attenuated the hazard ratios only minimally in both the partly and fully adjusted model. MIC-1/GDF15 –from here onwards referred to as GDF15– was the protein associated with the highest risk of progression (HR = 1.34, 95% CI = 1.01–1.79) and this association was replicated in ACCELERATE (HR = 1.22, 95% CI = 1.04–1.42). Of note, GDF15 did not show a difference in baseline levels, but it should be noted that GDF15 levels are dependent on more factors including age (Supplementary Fig. 6). The protein associated with the second highest risk of progression was the Nogo receptor (NogoR, HR = 1.33, 95% CI = 0.78–2.27, Fig. 3, Supplementary Data 2). In ANDIS, NogoR also replicated (HR = 1.20, 95% CI = 1.07–1.34, Fig. 3). NogoR was also a risk factor for incident (OR = 1.45, 95% CI = 1.15–1.83) and prevalent diabetes in AGES-Reykjavik (OR = 1.77, 95% CI = 1.60–1.95). In the top associated proteins, four were protective including SMAC, coactosin-like protein, testican-1 and HEMK2, of which HEMK2 was the most protective (HR = 0.78, 95% CI = 0.59–1.03). Levels of HEMK2 showed an interaction with C-peptide levels (P = 0.01). In the AGES-Reykjavik study, HEMK2 was also protective for prevalent diabetes (OR = 0.78, 95% CI = 0.72,0.85). Levels of testican-1 and HEMK, SMAC, coactosin-like protein were correlated (Supplementary Fig. 5
