A penalized regression framework for building polygenic risk models based on summary statistics from genome-wide association studies and incorporating external information

Abstract

Large-scale genome-wide association (GWAS) studies provide opportunities for developing genetic risk prediction models that have the potential to improve disease prevention, intervention or treatment. The key step is to develop polygenic risk score (PRS) models with high predictive performance for a given disease, which typically requires a large training data set for selecting truly associated single nucleotide polymorphisms (SNPs) and estimating effect sizes accurately. Here, we develop a comprehensive penalized regression for fitting l ₁ regularized regression models to GWAS summary statistics. We propose incorporating Pleiotropy and ANnotation information into PRS (PANPRS) development through suitable formulation of penalty functions and associated tuning parameters. Extensive simulations show that PANPRS performs equally well or better than existing PRS methods when no functional annotation or pleiotropy is incorporated. When functional annotation data and pleiotropy are informative, PANPRS substantially outperforms existing PRS methods in simulations. Finally, we applied our methods to build PRS for type 2 diabetes and melanoma and found that incorporating relevant functional annotations and GWAS of genetically related traits improved prediction of these two complex diseases.

Keywords: Genome wide association study; Lasso; genetic pleiotropy; genetic risk prediction; polygenic risk score; summary statistics.

Figure 1.

Concordance between Lasso and PANPRS. Figures A, B and C are for a quantitative trait (linear regression). Figures D, E and F are for a binary trait (logistic regression). For Figures A, B, D and E, the x-coordinate is the β values based on Lasso with individual level data; the y-coordinate is the β values estimated based on PANPRS. For Figures C and F, the x-coordinate is the number of nonzero estimates of regression coefficients of Lasso model; the y-coordinate is the correlation between the β values estimated based on Lasso and PANPRS. Figure A and D: Numerical experiment for M = 200 SNPs on chromosome 1. Df denotes the number of non-zero coefficients in Lasso estimates. Figures B and E: Numerical experiment for M = 213,240 SNPs.