Evaluation of genetic risk scores for prediction of dichotomous outcomes
© 2015, E-Century Publishing Corporation. All rights reserved. Substantial uncertainty exists as to whether combining multiple disease-associated single nucleotide polymorphisms (SNPs) into a genotype risk score (GRS) can improve the ability to predict the risk of disease in a clinically relevant way. We calculated the ability of a simple count GRS to predict the risk of a dichotomous outcome under both multiplicative and additive models of combined effects. We then compared the results of these simulations with the observed results of published GRS measured within multiple epidemiologic cohorts. If the combined effect of each disease-associated SNP included in a GRS is multiplicative on the risk scale, then a count GRS score should be useful for risk prediction with as few as 10-20 SNPs. Adding additional SNPs to the GRS under this model dramatically improves risk prediction. By contrast, if the combined effect of each SNP included in a GRS is linearly additive on the risk scale, a simple count GRS is unlikely to provide clinically useful risk prediction. Adding additional SNPs to the GRS under this model does not improve risk prediction. The combined effect of SNPs included in several published GRS measured in several well-phenotyped epidemiologic cohort studies appears to be more consistent with a linearly additive effect. A simple count GRS is unlikely to be clinically useful for predicting the risk of a dichotomous outcome. Alternative methods for constructing GRS that attempt to identify and include SNPs that demonstrate multiplicative gene-gene or gene-environment interactive effects are needed.
International Journal of Molecular Epidemiology and Genetics
Yoo, Wonsuk; Smith, Selina A.; and Coughlin, Steven S., "Evaluation of genetic risk scores for prediction of dichotomous outcomes" (2015). Neurobiology. 893.