Improving polygenic score prediction for underrepresented groups through transfer learning

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Large biobanks have improved polygenic score (PGS) accuracy.
• Existing PGS often show reduced performance in non-European ancestries due to derivation from European-ancestry data.
• Transfer learning presents a method to enhance PGS prediction across diverse populations.

Purpose of the Study:

• To introduce GPTL, an R package implementing transfer learning for developing polygenic scores.
• To address the ancestry-related disparities in PGS predictive performance.
• To provide a flexible software tool for PGS development using various data types.

Main Methods:

• Implementation of three transfer learning approaches within the GPTL R package: gradient descent with early stopping, penalized regression, and a Bayesian method with finite-mixture priors.
• Utilizing simulated data and real-world data from UK-Biobank and All of Us.
• Comparison of transfer learning-based PGS with single-ancestry and multi-ancestry ensemble-based PGS.

Main Results:

• PGS developed using GPTL's transfer learning algorithms consistently outperformed single-ancestry PGS.
• In many scenarios, GPTL-based PGS achieved performance comparable to or better than multi-ancestry ensemble-based PGS.
• The developed methods demonstrated effectiveness across both simulated and real genomic datasets.

Conclusions:

• GPTL offers a robust solution for developing more accurate and equitable polygenic scores across diverse ancestries.
• Transfer learning is a powerful strategy to mitigate ancestry-related biases in genomic prediction.
• The GPTL software package facilitates the application of advanced transfer learning techniques for personalized genomics and genetic research.