Predictive Capabilities of Polygenic Scores in an East-Asian Population-based Cohort: The Singapore Chinese Health Study
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Existing polygenic scores (PGS) are derived primarily from studies performed in European populations. It is still unclear how these perform in improving risk predictions in East-Asians.
Methods
We generated 2,173 PGSs from 519 traits and assessed their associations with 58 baseline phenotypes in the Singapore Chinese Health Study (SCHS), a prospective cohort of 23,622 middle-aged and older Chinese residing in Singapore. We used linear regression to evaluate PGS performances for quantitative traits by calculating the explained variance (r²). For dichotomized phenotypes, we employed logistic regression to estimate the area under the receiver operating characteristic curve (AUC) in predictive models.
Results
Overall, traits with higher heritability scores exhibited stronger associations with PGSs, while behavioural traits, for example sleep duration and hours spent watching TV, showed weaker associations. Height and type 2 diabetes (T2D) exhibited the largest SNP-based heritability estimates with the largest increments in explained variance and AUC, respectively, compared to phenotypic models. We explored the effect of T2D risk factors on the association between the T2D PGS (PGS003444) and incident T2D. The PGS association was significantly mediated and modified by hypertension ( P indirect =1.56×10 −18 , P interaction =1.11×10 −6 ) and body mass index (BMI, P indirect =1.25×10 −36 , P interaction =2.10×10 −3 ). The prediction ability of PGS003444 for incident T2D was stronger was stronger among individuals who were non-overweight without hypertension (AUC=0.774) than in overweight individuals with hypertension (AUC=0.709).
Conclusions
In conclusion, our study demonstrated the divergent ability of PGSs in predictions of complex traits, and showed that for certain traits, such as T2D, PGSs may have the potential for improving risk prediction and personalized healthcare.
