Improving polygenic risk score based drug response prediction using transfer learning.

Area of Science:

• Genomics
• Pharmacogenomics
• Precision Medicine

Background:

• Traditional polygenic risk scores (PRS-Dis) for pharmacogenomics (PGx) have limited accuracy due to incomplete heritability capture.
• Scarcity of large-scale, individual-level PGx data hinders the development of direct PGx PRS methods.

Purpose of the Study:

• To develop a novel transfer learning method (PRS-PGx-TL) for enhanced PGx polygenic risk score prediction.
• To leverage both disease and PGx data for more accurate prognostic and predictive polygenic risk scores.

Main Methods:

• Developed PRS-PGx-TL, a transfer learning approach integrating large-scale disease summary statistics with individual-level PGx data.
• Utilized a two-dimensional penalized gradient descent algorithm for optimizing weights, starting from disease data and refining with cross-validation.
• Applied the method to simulated data and the IMPROVE-IT PGx GWAS dataset.

Main Results:

• PRS-PGx-TL significantly improved prediction accuracy for drug response compared to traditional PRS-Dis methods.
• The method demonstrated enhanced patient stratification capabilities.
• Validated findings through simulations and real-world PGx GWAS data analysis.

Conclusions:

• PRS-PGx-TL offers a promising approach to overcome data limitations in PGx research.
• This method advances precision medicine by enabling more effective use of genetic information for personalized treatment decisions.