Using Genetic Risk Score Approaches to Infer Whether an Environmental Factor Attenuates or Exacerbates the Adverse Influence of a Candidate Gene

Abstract

Some candidate genes have been robustly reported to be associated with complex traits, such as the fat mass and obesity-associated (FTO) gene on body mass index (BMI), and the fibroblast growth factor 5 (FGF5) gene on blood pressure levels. It is of interest to know whether an environmental factor (E) can attenuate or exacerbate the adverse influence of a candidate gene. To this end, we here evaluate the performance of "genetic risk score" (GRS) approaches to detect "gene-environment interactions" (G × E). In the first stage, a GRS is calculated according to the genotypes of variants in a candidate gene. In the second stage, we test whether E can significantly modify this GRS effect. This two-stage procedure can not only provide a p-value for a G × E test but also guide inferences on how E modifies the adverse effect of a gene. With systematic simulations, we compared several ways to construct a GRS. If E exacerbates the adverse influence of a gene, GRS formed by the elastic net (ENET) or the least absolute shrinkage and selection operator (LASSO) is recommended. However, the performance of ENET or LASSO will be compromised if E attenuates the adverse influence of a gene, and using the ridge regression (RIDGE) can be more powerful in this situation. Applying RIDGE to 18,424 subjects in the Taiwan Biobank, we showed that performing regular exercise can attenuate the adverse influence of the FTO gene on four obesity measures: BMI (p = 0.0009), body fat percentage (p = 0.0031), waist circumference (p = 0.0052), and hip circumference (p = 0.0001). As another example, we used RIDGE and found the FGF5 gene has a stronger effect on blood pressure in Han Chinese with a higher waistto-hip ratio [p = 0.0013 for diastolic blood pressure (DBP) and p = 0.0027 for systolic blood pressure (SBP)]. This study provides an evaluation on the GRS approaches, which is important to infer whether E attenuates or exacerbates the adverse influence of a candidate gene.

Efficient gene–environment interaction testing through bootstrap aggregating

“…Alternatively, GRS can be constructed internally 3 , 24 , which means that the available data has to be divided into independent training and test data sets. The GRS is constructed using the training data and evaluated on the test data.…”

Estimating the Relative Contribution of Environmental and Genetic Risk Factors to Different Aging Traits by Combining Correlated Variables into Weighted Risk Scores

“…The weights of the predictors will be determined internally in a training sample using regularized regression (explicitly, ridge regression) and a bootstrapping approach to account for the randomness in splitting the data set into training and test set. Estimating the weights by regularized regression in a training sample has been widely used for the construction of polygenic risk scores in general [ 13 , 14 ] and in the context of gene–environment–interaction studies [ 15 , 16 , 17 ]. In the second step, the Lindeman–Merenda–Gold measure of relative importance [ 18 ] is applied to assess the contribution of the composed risk scores to the health outcome in the test sample.…”