Pharmacogenomics polygenic risk score for drug response prediction using PRS-PGx methods

Abstract

Polygenic risk scores (PRS) have been successfully developed for the prediction of human diseases and complex traits in the past years. For drug response prediction in randomized clinical trials, a common practice is to apply PRS built from a disease genome-wide association study (GWAS) directly to a corresponding pharmacogenomics (PGx) setting. Here, we show that such an approach relies on stringent assumptions about the prognostic and predictive effects of the selected genetic variants. We propose a shift from disease PRS to PGx PRS approaches by simultaneously modeling both the prognostic and predictive effects and further make this shift possible by developing a series of PRS-PGx methods, including a novel Bayesian regression approach (PRS-PGx-Bayes). Simulation studies show that PRS-PGx methods generally outperform the disease PRS methods and PRS-PGx-Bayes is superior to all other PRS-PGx methods. We further apply the PRS-PGx methods to PGx GWAS data from a large cardiovascular randomized clinical trial (IMPROVE-IT) to predict treatment related LDL cholesterol reduction. The results demonstrate substantial improvement of PRS-PGx-Bayes in both prediction accuracy and the capability of capturing the treatment-specific predictive effects while compared with the disease PRS approaches.

Fig. 1. Distributions of the prognostic to predictive effect size ratios calculated from the IMPROVE-IT PGx GWAS summary statistics data with n = 5661 unrelated European samples.

The left boxplot shows the distribution of whole genome SNPs (after clumping, m = 8,551,930). The right three boxplots show the distribution of top SNPs (after clumping) with their 2df (G + G × T) two-sided test p-values less than the three p-value thresholds 1e−06 (m = 16), 1e−05 (m = 81), and 1e−04 (m = 472), respectively. In each boxplot, the band indicates the median, the box indicates the first and third quartiles, and the whiskers indicate ± 1.5 × interquartile range. Effect size ratios of m SNPs are overlaid on the corresponding boxplot as dot points.

Fig. 2. Predictive performance of five polygenic prediction methods in the simulation studies, where heritability was fixed at 0.3 and ψ/ξ = 1.

The numbers of the causal variants for P(causal) = 0.001, 0.01, and 0.1 were 5, 50, and 500, respectively. The training sample size for PRS-PGx approaches was either 1000 or 3000; for PRS-Dis-LDpred2 approach was 20,000. The tuning parameters were selected via cross-validation in the training data. The performance was assessed in terms of a prediction accuracy R² of S_PGx in two arms, b predictive p-value for the two-sided S_pred × T interaction test, c R² of S_PGx under treatment arm, and d R² of S_PGx under control arm. Data are presented as mean values +/− standard deviations (error bars) with 10,000 replications, where results were calculated from the testing sets.

Fig. 3. The drug response prediction performance comparison among five methods based on the simulated data with completely separate prognostic and predictive SNPs and heritability fixed at 0.3.

The training sample size for PGx PRS approaches was fixed to be 3000. Numbers of the causal variants for P(causal) = 0.001, 0.01, and 0.1 are 5, 50, and 500, respectively. The performance was assessed in terms of a prediction accuracy R² of S_PGx in two arms, b predictive p-value for the two-sided S_pred × T interaction test. Data are presented as mean values +/− standard deviations (error bars) with 10,000 replications, where results were calculated from the testing sets.

Fig. 4. Patient stratification performance of six polygenic prediction methods in the IMPROVE-IT PGx real data analysis with n = 5661 unrelated European samples.

a Quantile plot of treatment effect using four fixed quantiles (0–25%, 25–50%, 50–75%, and 75–100%). Each dot stands for the observed Treatment Effect (TE), and each bar denotes the 95% Confidence Interval (CI). b Differential treatment effect when patients were stratified into top 10%, 20%, ⋯ , 90% percentile of the predictive score vs. the rest, respectively.