Quantifying the utility of type 2 diabetes polygenic risk score for predicting incident diabetes: an analysis of large US-based cohort studies

BACKGROUND: Genome-wide association analyses (GWASs) have identified numerous genetic variants associated with type 2 diabetes, but the utility of polygenic risk scores (PRSs) derived from these associations for predicting future incident diabetes remains uncertain. We analyzed utility of PRSs to predict incident diabetes in longitudinal studies, consisting of African Americans (AfrAm) and European Americans (EurAm) from the publicly available dbGAP resource. METHODS: Data consisted of 3,886 AfrAm and 17,345 EurAm with GWAS data, who were initially without diabetes; there were 688 incident diabetes cases in AfrAm and 2,304 in EurAm. Strength of association was assesed by the hazard ratio (HR), while ability to discriminate between those who did and did not develop diabetes was assessed by change in the C statistic (ΔC), and ability to correctly reclassify diabetes risk was assessed by the net reclassficiation information (NRI). Decision curve analysis was used to model potential benefits of using the PRS to select individuals for a preventive intevention across a range of thresholds for implementation. RESULTS: Among the 8 PRSs evaluated, PGS002308 (1.2 M variants, from a multi-ancestry GWAS) provided the best performance. With adjustment for age, sex, parental history of diabetes, body mass index (BMI) and fasting glucose levels, the HR was 1.50 per 5000 risk alleles (95% confidence interval, 1.41-1.60). The NRI was 0.313 (0.229-0.397) and ΔC was 0.010 (0.007-0.014). In comparison, NRI was 0.401 (0.338-0.464) for BMI and 0.257 (0.188-0.327) for serum lipid levels. The PRS provided a 5.9% (1.6-10.4) improvement in area under the decision curve. For most measures of predictive utility there was little heterogeneity between AfrAm and EurAm. In contrast, measures of predictive utility were stronger in younger, than in older, individuals. CONCLUSIONS: PRSs for type 2 diabetes, as currently constructed, provide utility for clinical prevention purposes that is similar to that provided by commonly used clinical predictors; utility is likely greater in younger than in older individuals. Further studies of diabetes risk assessment and those that aim to determine optimal target populations for diabetes prevention efforts would be strengthened by incorporating PRSs.