Abstract
Pharmacogenomics (PGx) utilizes genetic information to predict drug response and optimize patient treatment. Polygenic risk scores (PRS) are a promising tool in the PGx field, aggregating the effects of multiple genetic variants to predict drug response and stratify patients. Existing PRS development strategies have significant limitations. A common approach using disease-related genome-wide association study (GWAS) summary statistics (PRS-Dis) usually fails in capturing predictive genetic effects, thereby limiting its capability in stratifying patients and explaining drug response heritability. An alternative, PRS based on PGx GWAS data from treated individuals (PRS-PGx-T), directly targets drug response but is often hindered by small sample sizes and its inability to distinguish between prognostic and predictive effects. To overcome these challenges, we first propose a naïve composite method, PRS-DisPGx-C, which integrates disease and PGx-T GWAS summary statistics on either PRS level (PRS-DisPGx-C-p) or variant level (PRS-DisPGx-C-v) to separately estimate the aggregated prognostic and predictive genetic effects. Our theoretical research work further indicates that the performance of an optimal PRS approach depends on the genetic correlation between disease and drug response, the correlation between base and target cohorts (e.g., between two prognostic effects), and the correlation between prognostic and predictive effects. This insight motivates us to extend our composite method into an adaptive ensemble approach, PRS-DisPGx-E(-p/v), which dynamically combines disease and PGx-T GWAS data with optimal weights to further enhance drug response prediction. In simulation studies under various scenarios, our proposed ensemble approach significantly outperforms the conventional PRS-Dis method based on R^2 and predictive p-value metrics. Application to the IMPROVE-IT PGx GWAS data demonstrates the practical utility of our methods for improved prediction of LDL cholesterol reduction and patient stratification.