Cross-Ancestry PRS

The foundational requirement for cross-ancestry PRS is the use of GWAS summary statistics derived from multiple continental populations. Unlike standard PRS trained exclusively on European cohorts, these models incorporate data from African, East Asian, South Asian, and admixed populations. This diversity directly addresses the portability problem, where effect sizes and linkage disequilibrium patterns differ across ancestries. Key data sources include the Global Biobank Meta-analysis Initiative and All of Us Research Program.

Advanced Bayesian frameworks explicitly model ancestry-specific effect size distributions while leveraging shared genetic architecture:

• PRS-CSx: Extends PRS-CS with a shared continuous shrinkage prior across populations, coupling effect sizes while allowing ancestry-specific shrinkage
• XP-BLUP: Uses a multivariate normal prior to jointly model effect sizes from multiple GWAS
• PolyPred: Combines a standard PRS with a fine-mapping component that prioritizes likely causal variants shared across ancestries

These methods outperform simple meta-analysis by borrowing statistical strength across populations.

For admixed individuals such as African Americans or Latinos, the genome is a mosaic of ancestry segments. Cross-ancestry PRS methods incorporate local ancestry inference to assign ancestry-specific effect weights to each genomic segment. Tools like RFMix and ELAI first deconvolve the genome into haplotypes from each ancestral population, then apply the appropriate population-specific effect size. This approach dramatically improves prediction in admixed cohorts compared to treating them as a homogeneous group.

Cross-ancestry PRS evaluation requires population-stratified performance metrics to ensure no group is systematically disadvantaged:

• Ancestry-stratified AUC-ROC: Discriminative ability calculated separately for each population
• Calibration-in-the-large: Whether predicted risk matches observed prevalence across groups
• Net Reclassification Improvement by ancestry: Quantifies whether adding PRS improves risk stratification equally

A well-calibrated cross-ancestry PRS should show consistent calibration slopes across populations, avoiding systematic overestimation or underestimation of risk in any group.

When GWAS data is sparse for a target population, transfer learning frameworks adapt models trained on well-powered European studies. Techniques include importance weighting of training samples based on ancestry similarity and domain adaptation methods that learn invariant feature representations. The TL-PRS framework, for example, uses a multi-task learning architecture to jointly optimize prediction across source and target populations, significantly outperforming direct European PRS application in underrepresented groups.

Systematic differences in allele frequencies and disease prevalence between subpopulations due to ancestry, not biology. In cross-ancestry PRS, unmodeled stratification is the primary confound that causes scores trained in European cohorts to fail in African or Asian populations. Principal component analysis (PCA) and linear mixed models are standard corrections, but residual stratification persists when reference panels underrepresent target populations.

The non-random association of alleles at different loci that varies dramatically across ancestries. A causal variant tagged by a GWAS SNP in Europeans may have entirely different LD patterns in Africans due to shorter haplotype blocks. Cross-ancestry methods must account for this by using population-specific LD reference panels or by performing trans-ethnic fine-mapping to identify the true functional variants rather than relying on proxy SNPs.

A foundational step for cross-ancestry PRS that combines GWAS summary statistics from cohorts of diverse ancestries using fixed-effects or random-effects models. Key approaches include:

• MANTRA: Bayesian meta-analysis allowing heterogeneity in effect sizes across ancestries
• MR-MEGA: Models allelic effects as a function of genetic distance along principal components
• Trans-ethnic fixed-effects: Assumes shared causal variants with identical effects, which is often violated

The phenomenon where the number, frequency, and effect sizes of causal variants differ between ancestries. Cross-ancestry PRS must contend with:

• Allelic heterogeneity: Different causal variants at the same locus across populations
• Effect size correlation: Imperfect correlation of variant effects (typically 0.7-0.9 between continental groups)
• Minor allele frequency (MAF) divergence: Variants common in one population may be rare in another, reducing tagging efficiency