Multivariate adaptive shrinkage improves cross-population transcriptome prediction for transcriptome-wide association...

Area of Science:

• Genomics
• Population Genetics
• Bioinformatics

Background:

• Transcriptome prediction models trained on European data exhibit reduced accuracy in other populations due to variations in linkage disequilibrium and allele frequencies.
• Improving cross-population transcriptome prediction is crucial for equitable genomic research and understanding genetic associations in diverse ancestries.

Approach:

• Developed and evaluated multiple transcriptome prediction models (Elastic Net, JTI, Matrix eQTL, MASHR, TIGAR) for transcriptome-wide association studies (TWAS).
• Assessed out-of-sample prediction accuracy in both population-matched and cross-population settings.
• Integrated models with multi-ethnic genome-wide association study (GWAS) summary statistics from PAGE and Pan-UK Biobank to evaluate TWAS performance.

Key Points:

• MASHR models demonstrated superior or equivalent performance in both population-matched and cross-population transcriptome prediction accuracy compared to other methods.
• MASHR-based TWAS analyses identified more reproducible discoveries across diverse datasets (PAGE, PanUKBB), including novel loci not previously found in GWAS.
• The study highlights the benefit of methods that incorporate effect size estimates from multiple populations.

Conclusions:

• Methods leveraging shared regulatory effects across populations enhance transcriptome prediction accuracy.
• MASHR is a robust method for improving multi-ethnic TWAS, leading to increased discovery power in underrepresented populations.
• Adopting cross-population aware methods is essential for advancing precision medicine and genetic discovery in global populations.