Discovery of biomarkers for glycaemic deterioration before and after the onset of type 2 diabetes: descriptive characteristics of the epidemiological studies within the IMI DIRECT Consortium

The derivations of cohort 1 and cohort 2 are shown in Fig. 1. The sampling frame for cohort 1 comprised four existing prospective cohort studies: Metabolic Syndrome in Men (METSIM, Finland) [8]; Relationship between Insulin Sensitivity and Cardiovascular disease (RISC) [9], Hoorn Meal Study (HMS) and New Hoorn Study (NHS) [10] (Netherlands); Health2010 [11], Health2006 [12], Danish Study of Functional Disorders (DanFunD) [13] and Gut, Grain and Greens (GGG) [14] studies (Denmark) and Malmö Diet and Cancer (MDC) study (Sweden) [15]. Participants for cohort 2 were identified through general practice and other registries, as described previously [7].

After excluding participants who did not meet the inclusion criteria or whose data failed quality control, a total of 2127 participants at risk of diabetes and 789 participants with type 2 diabetes were retained in cohort 1 and cohort 2, respectively. In cohort 1, emphasis was placed on recruiting participants deemed at high risk of type 2 diabetes according to ADA 2011 HbA_1c criteria (HbA_1c 40–48 mmol/mol [5.7–6.4%]) [2, 7]. As anticipated during the design phase, the sampling frame contained too few participants that fulfilled this criterion; thus, we proceeded to enrol participants with progressively lower HbA_1c concentrations, who were also considered at highest risk of glycaemic deterioration according to the DIRECT-DETECT risk algorithm (based on age, BMI, waist circumference, use of antihypertensive medication, smoking status and parental history of type 2 diabetes) applied to the parent cohort sampling frame [7, 16].

In cohort 1, 1989 (93%) participants enrolled at baseline also attended the first major follow-up visit at a mean of 18.6 (SD 1.4) months and 1729 (81%) attended the final follow-up visit 30.8 (SD 1.3) months later. In cohort 2, 668 (85%) participants enrolled at baseline attended their first major follow-up visit at 18.2 (SD 0.6) months and 552 (70%) attended the final follow-up visit 18.2 (SD 1.0) months later.

Plasma glucose, insulin and C-peptide assays for cohort 1 were carried out centrally at the University of Eastern Finland (Kuopio, Finland), where plasma glucose was analysed using the enzymatic glucose hexokinase method and photometric measurement on Konelab 20 XT Clinical Chemistry analyser (Thermo Fisher Scientific, Vantaa, Finland). In cohort 2, plasma insulin and C-peptide were analysed using chemiluminometric immunoassay (CLIA) (Liaison Insulin [DiaSorin, Saluggia, Italy] and Liaison C-peptide [DiaSorin]). The instrument used was DiaSorin Liaison Analyser (DiaSorin Deutschland, Dietzenbach, Germany). Plasma glucose, insulin and C-peptide assessments for cohort 2 were carried out centrally at the University of Exeter (Exeter, UK). Assessments of HbA_1c, blood lipids, alanine aminotransferase (ALT) and aspartate aminotransferase (AST) for both cohorts were carried out centrally at the University of Exeter. Glucose was measured by the enzymatic colorimetric assay GOD-PAP using Roche MODULAR P analysers (Hoffmann-La Roche, Basel, Switzerland). Insulin was measured by electrochemiluminescence using Roche E170 Analysers (Hoffmann-La Roche). C-peptide concentrations in plasma and urine were measured by electrochemiluminescence using Roche E170 Analysers (Hoffmann-La Roche). HbA_1c was measured by ion-exchange high-performance liquid chromatography using Tosoh G8 analysers (Tosoh Bioscience, San Francisco, CA, USA). Each biochemical assay was performed using validated standard methods. Reference samples were included in all procedures to control for inter-assay variation and laboratories regularly participated in international external quality assessment schemes. In addition, a subset of samples was assayed for C-peptide, insulin and glucose on both sites to assess inter-laboratory variation.

Triacylglycerol was measured by quantitative determination with glycerol blanking. HDL-cholesterol was measured directly using polyethylene glycol-modified enzymes and dextran sulphate. When cholesterol esterase and cholesterol oxidase enzymes are modified by polyethylene glycol, they show selective catalytic activity towards lipoprotein fractions, with reactivity increasing in the order LDL < VLDL ≈ chylomicrons < HDL. Total cholesterol was measured by an enzymatic, colorimetric method. LDL-cholesterol was calculated from the total cholesterol, HDL-cholesterol and triacylglycerol concentrations using the Friedewald formula: LDL = total cholesterol – HDL-cholesterol – (triacylglycerol/2.2). ALT and AST were measured by UV absorbance without pyridoxal phosphate activation. ALT, AST, cholesterol, glucose, triacylglycerol and HDL-cholesterol were measured using a Roche MODULAR P analyser (Roche Diagnostics, Indianapolis, IN, USA). Insulin and C-peptide were measured using a Roche E170 analyser (Roche Diagnostics).

Plasma concentrations of glucagon-like peptide-1 (GLP-1) were determined by drawing blood samples collected at two different time points (0 and 60 min) during the 75 g frequently sampled OGTT (fsOGTT)/mixed-meal tolerance test (MMTT) (baseline samples only). P800 tubes (Becton Dickinson, Wokingham, UK) were used to provide immediate protection from intrinsic proteolysis. Quantitative determination of active GLP-1 was achieved using MSD GLP-1 active kit (product code K150JWC; Meso Scale Diagnostics, Rockville, MD, USA). Total GLP-1 was assayed using MSD GLP-1 total kit (product code K150JVC; Meso Scale Diagnostics).

The volume of adipose tissue was measured in litres using MRI, as described elsewhere [17]. Total abdominal adipose tissue (TAAT) may be separated into intra-abdominal adipose tissue (IAAT), also known as ‘visceral’ fat, and abdominal subcutaneous adipose tissue (ASAT). IAAT is the volume of adipose tissue within the abdominal cavity. TAAT is the sum of IAAT and ASAT. Liver and pancreas fat and iron (T2*) were derived simultaneously using a multiecho MRI technique, as previously described [17, 18]. This method has the advantage over single voxel MR spectroscopy in that regional differences in ectopic fat distribution can be measured. Furthermore, it is often possible to obtain a single slice quantification of the liver and pancreas, allowing simultaneous measurement of fat and iron within two separate organs. A biexponential curve-fitting model was used to derive the relative signal contributions from fat and water from the many images normally obtained with the multiecho sequence. Briefly, tissue with no fat infiltration generates a very smooth decay curve, whereas tissue containing a higher level of fat is characterised by significant oscillations throughout the decay curve [18]. A further output from the multiecho technique is T2* tissue values; as changes in these are indicative of iron content, this provides a clinically relevant additional measurement. Tissue iron concentration (mg/g dry weight tissue) was estimated from T2* using a validated model [19].

Self-reported dietary intake was assessed by a 24 h multi-pass method, using food habit and 24 h recall questionnaires. Analysis of these diet data was undertaken using Dietplan-6 (version 6.70.43, 2013; Forestfield Software, Horsham, UK). The specific analysis methods are described in detail elsewhere [7]. We also objectively assessed diet using discriminative metabolite signatures, an approach described in detail elsewhere [20]. Briefly, each participant’s serum metabolite profile was obtained using a targeted metabolomics assay (AbsoluteIDQ p180 Kit; BIOCRATES Life Sciences, Innsbruck, Austria), which simultaneously quantifies 188 metabolites. In a previously published diet intervention study [20], serum samples had been collected in 19 participants who had undergone a metabolic ward-based supervised diet intervention. We assayed these blood samples using the BIOCRATES AbsoluteIDQ p180 Kit and derived diet-discriminatory metabolomic signatures, using previously described methods [20]. These data were then used to predict the dietary characteristics of the IMI DIRECT study participants.

Objective measures of physical activity were derived from triaxial accelerometers (ActiGraph GT3X+/GT3X+w/GT3X+bt; ActiGraph Co., Pensacola, FL, USA) as described previously [7]. Raw data files (.gt3x) were converted to comma separated value (.csv) format storing rawest possible accelerations for each axis at a resolution of 30 Hz using ActiLife 6 (version 6.11.5; ActiGraph Co.). All inferred measures of physical activity were calculated using PAMPRO (version uploaded 21 Oct 2015; MRC Epidemiology unit, Cambridge, UK), custom open source software available under public license (https://​github.​com/​Thomite/​pampro). Data from each axis of acceleration was auto-calibrated to local gravity. Non-wear was inferred as a vector magnitude SD of less than 4 mg for a consecutive period greater than 60 min. All measures presented here have been adjusted for diurnal rhythm to account for bias from non-wear removal. However, due to the wear method (non-dominant wrist fastened using the manufacturer’s non-removable hospital band), intermittent non-wear time was rare. The main physical activity estimates presented here are high-pass-filtered vector magnitude (hpfVM), which infers intensity of participants’ movement in any direction at any given time (here, averaged during wear period). Time spent in established physical activity intensities by physical activity energy expenditure was estimated using calculated hpfVM cutpoints: sedentary (<48 mg hpfVM), light (48–154 mg hpfVM), moderate (154–389 mg hpfVM) and vigorous (>389 mg hpfVM). The methods used to infer these measures have been validated and described in detail elsewhere [21].

DNA extraction was carried out using Maxwell 16 Blood DNA purification kits and a Maxwell 16 semi-automated nucleic acid purification system (Promega, Southampton, UK). Genotyping was conducted using the Illumina HumanCore array (HCE24 v1.0) and genotypes were called using Illumina’s GenCall algorithm. Samples were excluded for any of the following reasons: call rate <97%; low or excess mean heterozygosity; sex discordance and monozygosity. Genotyping quality control was then performed to provide high-quality genotype data for downstream analyses using the following criteria: call rate <99%; deviation from Hardy–Weinberg equilibrium (exact p < 0.001); variants not mapped to human genome build GRCh37 and variants with duplicate chromosome positions (a total of 30,318 markers were excluded). A total of 3032 samples and 517,958 markers across the two studies passed quality control procedures. We took autosomal variants with MAF >1% that passed quality control and constructed axes of genetic variation using principal components analysis implemented in the GCTA (version 1.24.4, downloaded from https://​cnsgenomics.​com/​software/​gcta/​#Download) software to identify ethnic outliers defined as non-European ancestry using the 1000 Genomes samples as reference. We identified six individuals as ethnic outliers.

Biomarker discovery analyses using these data also employ additional measures (including ‘omic’ measures), which are outside the scope of this cohort description. Additional measures that are not described here include transcriptomics (RNA sequencing from fasting whole blood), microbiomics (DNA isolation and deep sequencing in faecal samples), proteomics (targeted array in fasting plasma) and metabolomics (targeted and untargeted assays in fasting plasma). GAD/islet antigen-2 assessments from fasting serum samples were also undertaken. Data from the Recent Physical Activity Questionnaire (RPAQ) and sleep diaries were also collected in subcohorts.

A detailed section on sample size and power for the study is available in the previously published rationale and design paper [7]. Briefly, statistical power will vary depending on a number of factors specific to the analysis to be carried out, such as biomarker effect sizes, variance/frequency of outcome and biomarker, statistical modelling methods employed, number of tests (multiple testing adjustment) and of course sample size, available for the relevant variables included in the model. The dataset will therefore be well powered for some analyses while it may be underpowered for other analyses and will thus be covered in detail for the specific scenarios in subsequent analyses.

Based on ADA 2011 diagnostic criteria, impaired fasting plasma glucose is 5.6–6.9 mmol/l (100–125 mg/dl), impaired glucose tolerance is 2 h 75 g OGTT plasma glucose 7.8–11.0 mmol/l (140–199 mg/dl) and prediabetes HbA_1c is 40–48 mmol/mol (5.7–6.4%) [2]. Accordingly, below and above these cut-offs was considered ‘normal’ and ‘diabetic’ ranges, respectively. Cohort 1 was stratified into two categories of blood glucose level: normal glucose regulation (NGR) and impaired glucose regulation (IGR). NGR was defined as the HbA_1c, fasting glucose and 2 h glucose values being within the normal ranges for each measure. IGR was defined as there being impaired values in at least one of HbA_1c, fasting glucose or 2 h glucose. Cohort 2 was stratified into treatment categories: lifestyle advice only or metformin plus lifestyle advice. Descriptive statistics are presented as mean ± SD. Pairwise Pearson correlations were calculated for all key variables described here. For these analyses, continuous variables were first inverse normal transformed and then adjusted for age, sex and study centre by two-step residual regression. We present the same type of data for anthropometric and glycaemic variables for the main follow-up visits, as well as the differences for these variables between the baseline and the final follow-up visit measures (follow-upΔ = final follow-up value – baseline value). We also calculated pairwise Pearson correlations for the follow-upΔ values; for these analyses, continuous variables were first inverse normal transformed and then adjusted for age, sex, study centre and days since baseline visit by two-step residual regression. All statistics were computed using R software version 3.4.0 [22]. The IMI DIRECT data release version used for the analyses in this article was direct_03-29-2019.

Glycaemic traits were derived from the 75 g fsOGTT (sampling at 0, 15, 30, 45, 60, 90, 120 min) and MMTT (sampling at 0, 30, 60, 90, 120 min) for cohort 1 and cohort 2, respectively. Analyses used a mathematical model that describes the relationship between insulin secretion and glucose concentration [23, 24]. The model expresses insulin secretion as the sum of two components, the first of which represents the dependence of insulin secretion on absolute glucose concentration at any time during the fsOGTT/MMTT, through a dose–response function. Characteristic parameters of the dose–response relationship are the mean slope over the observed glucose range, denoted as glucose sensitivity. The dose–response relationship is modulated by a potentiation factor, which accounts for the fact that during acute stimulation, insulin secretion is higher on the descending phase of hyperglycaemia than at the same glucose concentration on the ascending phase. In participants with NGR and normal insulin secretion, the potentiation factor typically increases from baseline to the end of a 2 h OGTT [25]. To quantify this excursion, the ratio between the 2 h and the baseline value was calculated. This ratio is denoted as potentiation ratio and reflects late insulin release. The second insulin secretion component represents the dependence of insulin secretion on the rate of change of glucose concentration. This component is termed derivative component and is determined by a single parameter, denoted as rate sensitivity. Rate sensitivity reflects early insulin release [25].

The model parameters were estimated from glucose and C-peptide concentrations by regularised least-squares, as previously described [23]. Regularisation involves the choice of smoothing factors, which were selected to obtain glucose and C-peptide model residuals with SDs close to expected measurement error (~1% for glucose and ~4% for C-peptide). Insulin secretion rates were calculated from the model every 5 min. The integral of insulin secretion during the fsOGTT was denoted as total insulin output.

The validity of the fsOGTT and MMTT for the assessment of insulin sensitivity has been shown in the original publications presenting the indices [26‐28]. In the studies, the OGTT/MMTT indices are compared with values obtained by euglycaemic glucose clamp. The validity of the beta cell function model is supported by numerous studies [25]. For beta cell function, it is not possible to validate an OGTT/MMTT method against the classical tests with glucose intravenous infusion, due to the presence of the incretin effect. However, it has been shown that the estimated beta cell dose–response is consistent with the graded glucose infusion test across the spectrum of glucose tolerance [29, 30].

Of 2235 enrolled participants in cohort 1, 2127 passed all inclusion, exclusion and quality control criteria. Of these, 1419 (67%) had IGR according to at least one ADA category for HbA_1c, fasting glucose or 2 h glucose [2] and were thus within the target ‘prediabetes’ range. A total of 693 participants (33% of cohort 1) displayed NGR according to all three glycaemic measures. Participants with prevalent type 2 diabetes (n = 105) or who withdrew from the study (n = 3) were excluded from further analyses.

The number of participants enrolled into cohort 1 varied between centres, with the Finnish subcohort being the largest (providing 58% [n = 1240]) of the total cohort 1 baseline sample. The other centres in the Netherlands, Denmark and Sweden enrolled 22% (n = 473), 13% (n = 275) and 7% (n = 139) of the total cohort, respectively.

The ratio of men to women varied in each subcohort, with all participants at the Finnish centre being male, and 43%, 45% and 29% being male in the subcohorts from the Netherlands, Denmark and Sweden, respectively.

Detailed participant baseline characteristics for cohort 1 are shown in Table 1 (and stratified by glycaemic category in electronic supplementary material [ESM] Table 1). Figure 2 shows the pairwise correlations between a selection of key phenotypic variables at baseline. Participant characteristics at the follow-up visits and the difference (Δ) between baseline and final follow-up for cohort 1 are shown in Table 2. The pairwise correlations between the baseline to final follow-up difference for anthropometric and glucose-control variables are shown in Fig. 3.

Briefly, at baseline, participants had a mean (SD) age of 62 (6.2) years, BMI 27.9 (4.0) kg/m², HbA_1c 37 (2.9) mmol/mol [5.5 (0.27)%], fasting glucose 5.7 (0.6) mmol/l, 2 h glucose 5.9 (1.6) mmol/l, fasting insulin 10.9 (7.6) pmol/l, glucose sensitivity 113 (55) pmol min⁻¹ m⁻² (mmol/l)⁻¹ and insulin sensitivity (2 h oral glucose insulin sensitivity [OGIS]) 381 (59) ml min⁻¹ m⁻². Participants had a 0–48 month mean (SD) difference in fasting plasma glucose levels of 0.3 (0.2) mmol/l.

Of the participants in cohort 2 at baseline, the UK (Dundee, Exeter, Newcastle), Dutch (Amsterdam), Swedish (Lund) and Danish (Copenhagen) study centres enrolled 21% (n = 167), 18% (n = 141), 21% (n = 166), 21% (n = 167), 12% (n = 96) and 7% (n = 52) of the total cohort, respectively; 52–63% of the subcohort participants were male.

Detailed participant characteristics and key variables for cohort 2 at baseline are shown in Table 1 (and stratified by treatment category in ESM Table 1). Figure 2 shows the pairwise correlation matrix for key variables at baseline adjusted for age, sex and study centre. Participant characteristics for follow-up visits and the difference between the baseline and final follow-up visit (Δ) for cohort 2 are shown in Table 2. A pairwise correlation matrix for the anthropometric and glucose-control Δ variables are shown in Fig. 3.

Briefly, at baseline, participants had a mean (SD) age 62 (8.1) years, BMI 30.5 (5.0) kg/m², HbA_1c 46.5 (5.8) mmol/mol [6.4 (0.53)%], fasting glucose 7.2 (1.4) mmol/l, 2 h glucose 8.6 (2.8) mmol/l, fasting insulin 107 (71) pmol/l, glucose sensitivity 83 (55) pmol min⁻¹ m⁻² (mmol/l)⁻¹ and insulin sensitivity (2 h OGIS) 298 (69) ml min⁻¹ m⁻². Participants had a 0 months to 18 months mean difference in fasting plasma glucose levels of 0.8 (1.9) mmol/l.

As some study centres enrolled participants into both cohorts, we elected to characterise the genetic population substructure across the cohorts by study centre (i.e. pooling both cohorts at a given centre where possible). Genetic substructures closely map to the geographic location of the populations [31], indicating ethnic homogeneity within regions from which the cohorts were recruited, whereas there is far greater heterogeneity between centres, the latter driven mainly by the inclusion of Finnish participants. This is illustrated in Fig. 4 where Finnish participants form a distinct cluster (to the north east) compared with the population from the other cohorts.