Polygenic type 2 diabetes prediction at the limit of common variant detection

Abstract

Genome-wide association studies (GWAS) may have reached their limit of detecting common type 2 diabetes (T2D)-associated genetic variation. We evaluated the performance of current polygenic T2D prediction. Using data from the Framingham Offspring (FOS) and the Coronary Artery Risk Development in Young Adults (CARDIA) studies, we tested three hypotheses: 1) a 62-locus genotype risk score (GRSt) improves T2D prediction compared with previous less inclusive GRSt; 2) separate GRS for β-cell (GRSβ) and insulin resistance (GRSIR) independently predict T2D; and 3) the relationships between T2D and GRSt, GRSβ, or GRSIR do not differ between blacks and whites. Among 1,650 young white adults in CARDIA, 820 young black adults in CARDIA, and 3,471 white middle-aged adults in FOS, cumulative T2D incidence was 5.9%, 14.4%, and 12.9%, respectively, over 25 years. The 62-locus GRSt was significantly associated with incident T2D in all three groups. In FOS but not CARDIA, the 62-locus GRSt improved the model C statistic (0.698 and 0.726 for models without and with GRSt, respectively; P < 0.001) but did not materially improve risk reclassification in either study. Results were similar among blacks compared with whites. The GRSβ but not GRSIR predicted incident T2D among FOS and CARDIA whites. At the end of the era of common variant discovery for T2D, polygenic scores can predict T2D in whites and blacks but do not outperform clinical models. Further optimization of polygenic prediction may require novel analytic methods, including less common as well as functional variants.

© 2014 by the American Diabetes Association.

Figures

Figure 1

T2D-associated genetic loci. Loci on the x-axis are ordered by inclusion in published 17-, 40- and 62-SNP GRS. Black bars (left y-axis) indicate published DIAGRAMv3 odds ratio (OR) for T2D per risk allele at each locus. The black line plots the T2D OR in the FOS per allele increase in a GRS containing the loci up to that point on the x-axis. Points with error bars plot the C statistics (95% CI) from pooled logistic regression models for T2D in FOS including 17-, 40-, and 62-SNP GRS in demographic and clinical models. Loci used in separate β-cell and IR GRS in the present analyses are also indicated.

Figure 2

Receiver operating characteristic (ROC) curves for models predicting incident T2D with and without a 62-locus GRS among the FOS (A) and white (B) and black (C) young adults in the CARDIA Study. Graphs plot the sensitivity vs. (1 − specificity) for diabetes at each possible model cut point. The area under a ROC curve corresponds to the C statistic of that model. The full clinical model is adjusted for age, sex, parental diabetes (yes vs. no), BMI, systolic blood pressure, fasting glucose, HDL cholesterol, and triglyceride levels.