Abstract
Traditional methods for pharmacogenomics (PGx), like those using disease-specific polygenic risk scores (PRS-Dis), often fail to capture the full heritability of drug response, leading to poor predictions. Direct PGx PRS approaches could improve this, but the scarcity of relevant PGx datasets limits the wide application. To overcome these challenges, we introduce PRS-PGx-TL, a novel transfer learning method. It models large-scale disease summary statistics data alongside individual-level PGx data, leveraging both sources to create more accurate prognostic and predictive polygenic risk scores. In PRS-PGx-TL, we further develop a two-dimensional penalized gradient descent algorithm that starts with weights from disease data and then optimizes them using cross-validation. In simulations and an application to IMPROVE-IT (ClinicalTrials.gov, NCT00202878, September 13, 2005) PGx GWAS data, PRS-PGx-TL significantly enhances prediction accuracy and patient stratification compared to traditional PRS-Dis methods. Our approach shows great promise for advancing precision medicine by using an individual's genetic information to guide treatment decisions more effectively.