Improving polygenic score prediction for underrepresented groups through transfer learning

The advent of large biobanks has substantially increased the accuracy of polygenic scores (PGS). However, most existing PGSs were derived from European-ancestry data and often exhibit reduced predictive performance when applied to individuals of non-European ancestries. Transfer Learning offers a promising strategy to address this limitation by leveraging information learned in one population to improve prediction in another. Here, we introduce GPTL, an R package that implements three Transfer Learning based approaches for developing PGS: (1) gradient descent with early stopping, (2) a penalized regression model that shrinks variant-effect estimates toward prior values, and (3) a Bayesian method with a finite-mixture prior that enables integration of multiple prior sources of information. Using both simulated data and real data from the UK-Biobank and All of Us, we demonstrate that PGS generated with GPTL's Transfer Learning algorithms consistently outperform single-ancestry PGS and, in many settings, match or exceed the performance of multi-ancestry ensemble-based PGS. Our software can be used with either individual genotype-phenotype data or summary statistics from genome-wide association studies.