Improving Polygenic Risk Score Based Drug Response Prediction Using Transfer Learning

Pharmacogenomics (PGx) studies aim to perform drug response prediction and patient stratification using genome-wide association study (GWAS) data from randomized clinical trials. Polygenic risk scores (PRS) are useful tools for PGx. By combining information across the genome, they have shown great promise in predicting disease risk and how patients respond to a particular treatment. A common practice when developing polygenic models for drug response prediction, is to use disease GWAS summary statistics derived from large cohorts of related disease phenotypes. However, this disease PRS approach (PRS-Dis) lacks the ability to incorporate any predictive (or genotype-by-treatment interaction) effects in the PRS training stage and thus cannot fully capture the heritability of drug response, often resulting in poor predictive performance. On the other hand, a direct PGx PRS approach (PRS-PGx) requires an independent PGx GWAS dataset with the same or similar drug response phenotype, which is usually not available. To fill this gap, we propose a transfer learning (TL) based method (PRS-PGx-TL) that jointly models large-scale disease GWAS summary statistics from the base (training) cohort and individual-level PGx data from the target cohort, leveraging both for parameter optimization and prognostic and predictive PRS construction. In PRS-PGx-TL, we develop a two-dimensional penalized gradient descent algorithm, which utilizes the PRS weights from the disease GWAS as initial values and optimizes the tuning parameters using a cross-validation framework while updating both prognostic and predictive effect estimates simultaneously. Through extensive simulation studies, we show that PRS-PGx-TL improves prediction accuracy and population stratification performance compared to the traditional PRS-Dis methods (e.g., PRS-CS, Lassosum). We further demonstrate its advantages by applying it to the IMPROVE-IT PGx GWAS data for predicting treatment related LDL cholesterol reduction. Overall, our proposed TL-based PRS method shows great value in improving drug response prediction and patient stratification and can help facilitate precision medicine by using an individual’s genotype information to guide treatment.