Type 2 diabetes mellitus (T2D), a disease characterized by persistent hyperglycemia, is a common and heritable complex disease affecting the health of millions of people worldwide1. Estimates from the World Health Organization indicate that 8.5% of the adult population had T2D in 2016, and this prevalence has been steadily increasing during the last three decades2.

Prospective epidemiological studies have demonstrated that the risk of T2D starts even in the normal fasting glucose range and exponentially increases in pre-diabetic ranges3,4,5,6,7. Relevant physiological perturbations producing a slow utilization of fasting glucose are likely to be present at stages of the disease as early as a decade before diagnosis8. The etiological causes of early glycemic perturbations are likely to be triggered by environmental and lifestyle factors9,10, but the precise biological mechanisms underpinning why people differently progress to hyperglycemia are unknown.

Recent large-scale genetic association meta-analyses have uncovered genetic variants cross-sectionally associated with T2D and related glycemic traits11,12,13,14,15,16. However, prospective data for genetic variant association discovery are scarce and findings have been inconsistent17,18,19. In this study, we conducted the largest genome-wide association study (GWAS) to date to identify genetic variants associated with fasting glucose changes over up to 14 years in 13,807 non-diabetic participants of European descent from nine cohorts.


We included a total of 13,807 participants of European descent and free of diabetes at baseline and during the entire follow-up period with repeated fasting glucose levels measured at least at two time points over up to 14 years from nine cohorts. Characteristics of the study sample and follow-up, phenotype, and genotype information are presented in Table S1. The participants’ average age at baseline for each cohort ranged from 41 to 70 years old. The follow-up time varied by cohort, on average ranging from 5 to 25 years. The slope of fasting glucose was calculated based on available fasting glucose measurements during the follow-up and then the slopes were inverse normal transformed within each cohort, for harmonization. We refer to this transformed slope as fasting glucose change over time. At baseline, cohort-specific average fasting glucose ranged from 4.9 to 6.2 mmol/L in men and 4.8 to 5.9 mmol/L in women, respectively. Our primary analysis included all available participants. In an exploratory analysis, we investigated whether stratifying our sample by cohorts with long-term follow-up [≥10 years] or short-term follow-up [<10 years] identified pertinent signals.

In a genome-wide association meta-analysis we did not find evidence of genetic variants associated with fasting glucose change over time at genome-wide significance level (P < 5 × 10−8) (Supplemental Figs 13) nor evidence of inflated signals (Supplemental Figs 46) in the primary analysis including the entire sample or the sensitivity analysis. For the analysis with all samples, the most significant association with fasting glucose changes over time was an intronic variant at the ODZ4 locus (rs7114256; P = 8.78 × 10−7, Table 1, Fig. 1). There were five other suggestively associated (P < 5 × 10−6) variants for fasting glucose changes over time in three loci whose closest reference genes including ALLC (rs606243), NUDT12 (rs17496593, rs17496653, rs17562893) and ODZ4 (rs7103693) (Table 1, Supplemental Figs 78). In our exploratory analysis stratifying cohorts by follow-up time, there were a few suggestively associated variants with fasting glucose changes over time (Supplemental Tables 56). These included four loci whose closest reference genes were SNX16, BEGFA, GATA3 and CDKAL1 from short follow-up analysis with sample size up to 8,195 and ten loci whose closest reference genes were HCRTR2, WRN, SEPT9, SLC35B3, FAM84A, GRM8, MPP6, BAMBI, SSB, and C8orf31 from long follow-up analysis with sample size up to 3,669.

Table 1 Genome-wide association results for genetic variants with an association p-value < 5 × 10−6.
Figure 1
figure 1

Regional association plot of rs7114256. Results from 500 kb regional associations for fasting glucose change over time, centered at rs7114256. The x axis denotes genomic position build 36 and the y axis denotes the −log(P-value) and recombination rate (blue line). The purple diamond symbol represents the most-associated SNP within the region, rs7114256. The color of each symbol indicates the LD value with rs7114256 based on the HapMap2 CEU sample.

Next, we investigated whether genetic variants available in our GWAS previously associated with T2D prevalence (82 SNPs)15 and cross-sectional glycemic traits based on our primary analysis with all available sample, including fasting glucose (32 SNPs)13, and HbA1c (58 SNPs)16, associated with fasting glucose change over time13,15,16. For T2D associated genetic variants15, we showed evidence of a nominal significant association between a variant at CDKN2A/B loci (rs10965248, P = 0.0192) and longitudinal fasting glucose change. In addition, two loci previously associated with cross-sectional FG in the latest GWAS13 for fasting glucose associated with longitudinal fasting glucose changes (GRB10; rs6943153, P = 0.0019 and PDX1; rs11619319, P = 0.0114), as well as five loci previously associated with HbA1c in the latest GWAS for HbA1c16 including TMC6 (rs2073285, P = 0.0019), CDH3 (rs4783565, P = 0.0057), ABO (rs579459, P = 0.0082), PDX1 (rs11619319, P = 0.0114), and HK1 (rs10823343, P = 0.0390) (Table 2, Supplemental Tables 24). After Bonferroni correction for conduct of 82, 32, and 58 tests for T2D risk, fasting glucose, and HbA1c, respectively none of these signals remained significant.

Table 2 Association results of the genetic variants showing a nominal significant signal for fasting glucose change (p < 0.05) in known T2D or glycemic trait locia.

We conducted a power analysis with a sample size of 13,807 at the genome-wide significant threshold (5 × 10−8) to detect a genetic variant explaining at least 0.05% to 0.5% of the variation in the fasting glucose change over time. The results showed that our study has 80% power to detect genetic variants, which explain at least 0.28% of variation in change of fasting glucose over time (Fig. 2). This is equivalent to detect the genetic variants with minor allele frequency of 0.05 or 0.25 whose minimum effect corresponds to 0.17 or 0.09 standard deviation unit difference in the change of fasting glucose over time, respectively.

Figure 2
figure 2

Power analysis. The relationship between power and variation explained in the trait of interest by a genetic variant with a sample size of 13,807 at a significance level of 5 × 10−8. The y-axis represents the power and the x-axis the variance explained by a genetic variant. The horizontal red line represents the power of 80%. This Figure shows that we had 80% power to detect a genetic variant that explained at least 0.28% of variation in fasting glucose change over time.


We tested whether common genetic variants were associated with fasting glucose change over time in a GWAS including 13,807 initially non-diabetic participants from nine cohorts of European descent with repeated fasting glucose measures over up to 14 years, We found three suggestive associated variants at sub genome significance level (near ODZ4, ALLC, and NUDT12), and eight nominally associated previously known T2D-glycemia GWAS loci (CDKN2A/B, GRB10, PDX1, TMC6, CDH3, ABO, PDX1, and HK1) but none reached genome-wide significance for association or survived adjustment for multiple testing. We have placed the GWAS results data sets from this analysis on line at the Meta-Analyses of Glucose and Insulin-related traits Consortium (MAGIC) website ( and T2D knowledge portal ( to provide a genomic resource to further combine with futures studies and evaluate shared genetic links with T2D and other metabolic risk traits.

To date, more than 120 genetic loci have been identified to be associated with cross-sectional glycemic outcomes in successive waves of large-scale genetic association studies12,14,15. These risk alleles are associated with glycemic phenotypes and predict incident T2D when aggregated into a genetic risk score12,20,21. However, findings from our study do not support that single common risk alleles have a substantive impact on longitudinal fasting glucose changes. If there are effects on glucose changes, they are likely to be small. This observation is in agreement with evidence from other intermediate phenotypes such as genetic variants associated with deterioration of lipid levels22,23,24,25, lung function26,27 or change in BMI28,29,30,31, where modest effects have been attributed to genetic factors on longitudinal trait changes, relative to single cross sectional trait measures. Regarding glycemic trajectories, no single genetic variants associated with glucose deterioration over time were detected in previous studies of European descent indivuals17 or in Han individuals19, although Han Chinese carrying a higher number of T2D increasing-risk variants showed a greater increase in FG over time compared with those carrying a lower number of T2D increasing-risk variants. One small single-cohort study identified five genomic regions associated at genome-wide significance with longitudinal change in fasting glucose (GCKR, G6PC2, GCK, SLC30A8, MTNR1B), but the study included diabetic individuals taking hypoglycemic medications, which almost certainly introduced confounding into genotype – glucose change associations18.

Our results may highlight the importance of environmental determinants of glycemic deterioration. Nevertheless, genetic determinants of changes in glycemia may remain relevant for people with rapid transition from pre-diabetes to diabetes, which the design of our study was not able to capture. A potential future strategy to identify these loci, if they exist, would be to focus on pre-diabetic individuals who progress to T2D and adjust for the effects of all known variants affecting cross-sectional blood glucose inter-individual variability32. A limitation to our study is that the meta-analysis only involved European participants, so the results may not be generalizable to other ancestry groups. The power of our study was relatively limited even though this is the largest existing meta-analysis of fasting glucose change. In addition, we identified challenges posed by phenotypic heterogeneity, e.g. different follow-up duration or different numbers of longitudinal data points. Larger sample sizes from new cohorts will be key to help confirm or refute the current findings. An alternative approach, if data are available in the future, would be to study longitudinal glycemia in large and homogeneous populations, with more homogeneous phenotypes especially with more consistent number of follow-up visits and similar follow-up duration to gain more statistical power.

In summary, a large GWAS did not identify common genetic variation genome-wide significantly associated with fasting glucose change over time. Such genetic effects, if present, are likely small. The data have been deposited as a public genetic epidemiological resource to aid the hunt for genetic determinants of T2D and its relevant physiology.

Methods and Materials

Study sample

We recruited in total 13,807 individuals of European descent free from T2D at baseline and during the entire follow-up period with repeated fasting glucose measurements at two or more time points from nine cohorts representing three continents (America, Europe and Australia). The participating cohorts include the Bogalusa Heart Study (BHS), the CoLaus study (COLAUS), the Data from the Epidemiological Study on the Insulin Resistance Syndrome study (DESIR), the Erasmus Rotterdam Gezondheid Onderzoek study (ERGO), Framingham Heart Study (FHS), the Helsinki Birth Cohort Study (HBCS), Cooperative Health Research in the Region of Augsburg (KORA), Prevention of Renal and Vascular End-stage Disease study (PREVEND), and the National Institute on Aging (NIA) SardiNIA Study (SARDINIA). The ethnicity information for each individual was based on questionnaires or assessed using genetic data (principal component analysis). Ethnic outliers detected by principal component analysis for European ethnicity were excluded from further analysis. Diabetes was defined as a fasting glucose level >7 mmol/l, or use of glucose lowering medication. The study conformed to the Declaration of Helsinki guidelines. Institutional Review Board and/or oversight committees approved the study in each participating cohort and all participants provided written informed consent (See Supplemental Text).

Genotyping, imputation and quality control

Genotyping was conducted as specified in the Table S1. Each study imputed their genotype to ~2.5 million Phase 2 HapMap CEU SNPs with imputation software, either IMPUTE or MACH33,34. We applied a quality control filter by removing SNPs with a minor allele frequency less than 1% and those with an imputation quality threshold proper_info < 0.4 for cohorts using IMPUTE and r2 > 0.3 for cohorts using MACH. We used imputed allelic dosage in our association analysis.


To calculate longitudinal fasting glucose slopes we used repeated FG measurements available in the longitudinal cohort studies. To harmonize FG measures, FG measures obtained from whole blood were converted to plasma levels by using a coefficient of 1.13 (FG in mmol/l). We then modelled the association for each individual between FG and duration of time between baseline measure and each follow-up measure. The resulting beta coefficients (slopes) were then pooled and inverse normal transformed. The transformed slopes were used as the trait ‘fasting glucose change’ for the genetic association analysis.

Association Analysis and Meta-analysis

For each participating cohort, we conducted genome-wide association analysis with transformed longitudinal fasting glucose changes adjusting for age at baseline, sex, and principal components of genetic variation to account for population stratification using linear regression with additive genetic effects for cohorts with unrelated samples. We performed mixed-effect model analysis with random effect to account for sample relatedness for cohorts with related samples. We then conducted inverse variance weighted meta-analysis of cohort-specific association results using METAL35. In an exploratory analysis, we stratified our analyses by study follow-up time and classified each cohort as having a long follow-up time (≥10 years) or a short follow-up time (<10 years). We applied genomic control correction to control type I error36. SNPs with a meta-analysis p-value ≤ 5 × 10−8 were considered to be genome-wide significant.

Interrogation of Published Loci for Type 2 Diabetes Related Traits

We tested the hypothesis that longitudinal fasting glucose slopes would be associated with previously-identified GWAS variants for T2D (128 SNPs)15, fasting glucose (32 SNPs)13, and HbA1c (60 SNPs)16. If the previously reported most-associated SNP was unavailable in the present analysis, we used a proxy SNP (LD r2 > 0.8) if available. After proxy searches we evaluated 172 loci, including 82 for T2D, 58 for HbA1c and 32 for FG. The Bonferroni corrected p-value threshold for these look-ups was set at 0.0003 (0.05/172).

Post-hoc power calculation

We conducted a post-hoc power analysis using Quanto software to investigate the power to detect 0.05% to 0.5% percent variation in phenotype explained by a genetic variant with a sample size of 13,807 at the genome-wide significant threshold (5 × 10−8).