All Times EDT
Keywords: PGx PRS, Drug response, Predictive score, Bayesian regression, Shrinkage prior
Polygenic risk score (PRS), by combining many small prognostic genetic effects, is a promising prognostic biomarker for the prediction of complex traits in genome-wide association studies (GWAS). In efficacy-based pharmacogenomics (PGx) studies, applying the PRS built from disease GWAS directly to PGx data from randomized clinical trials relies on a stringent assumption, which largely may not be true. A violation of such assumption makes disease PRS explain less heritability of drug responses and thus reduce power in predicting them. Here, we propose the shift from disease PRS to PGx PRS approaches by simultaneously modeling the prognostic and predictive effects. Two PRSs, served as prognostic and predictive biomarkers, respectively, are constructed for drug response prediction and patient stratification in PGx. We make this shift possible by developing a series of novel PRS-PGx methods using clumping and thresholding, penalized regression, and Bayesian regression, separately. In the framework of Bayesian regression (i.e., PRS-PGx-Bayes), we propose a polygenic prediction method that infers posterior prognostic and predictive effect sizes simultaneously using PGx GWAS summary statistics and an external linkage disequilibrium (LD) reference panel. By introducing global-local continuous shrinkage priors on variant effect sizes, our proposed PRS-PGx-Bayes method is more robust to varying relationships between the genotype main and genotype-by-treatment interaction effects. Extensive simulation studies show that PRS-PGx methods generally outperform the current disease PRS methods across a wide range of genetic architectures and PRS-PGx-Bayes is superior to all other PRS-PGx methods. We further apply the PRS-PGx methods to the IMPROVE-IT PGx GWAS data. The drug response prediction results demonstrate the great improvement of PRS-PGx-Bayes in both the prediction accuracy and the treatment-specific predictive ability of patient stratification over alternative methods.