Multiethnic polygenic risk prediction in diverse populations through transfer learning

Area of Science:

• Genetics and Genomics
• Computational Biology
• Population Health

Background:

• Polygenic risk scores (PRS) estimate disease risk based on an individual's genotype but are less accurate in non-European populations due to smaller genome-wide association study (GWAS) sample sizes.
• Existing PRS prediction methods are primarily developed for and validated in European ancestry populations, leading to performance disparities in other ethnic groups.

Purpose of the Study:

• To introduce TL-Multi, a novel transfer learning method for constructing accurate PRS in non-European populations.
• To leverage knowledge from European GWAS data to correct biases and improve PRS prediction accuracy in underrepresented populations.

Main Methods:

• Developed TL-Multi, a transfer learning framework using non-European GWAS data as target and European GWAS data as auxiliary information.
• Applied TL-Multi to predict systemic lupus erythematosus (SLE) risk in Asians and asthma risk in Indians, utilizing European population data.

Main Results:

• TL-Multi demonstrated superior prediction accuracy compared to traditional methods like Lassosum and meta-analysis in both simulated and real-world datasets.
• The transfer learning approach effectively borrowed information from the European population to enhance PRS accuracy for non-European target populations.

Conclusions:

• TL-Multi offers a robust and accurate solution for improving PRS prediction in non-European populations, addressing a critical gap in personalized medicine.
• This method holds significant potential for more equitable and effective disease risk assessment across diverse ethnic groups.