PRS Cross-Ancestry Portability & Recalibration Benchmark

Challenge

Polygenic risk scores (PRS) trained predominantly on European-ancestry GWAS are documented to lose predictive accuracy when applied to other ancestries, but isolating how much of that loss comes from LD/allele-frequency differences versus genuine gene-by-ancestry effect heterogeneity requires a ground truth no real trait provides — real phenotypes don't come with known causal variants or heritability.

Approach

A GCTA-simulated quantitative phenotype (300 causal variants, target h²=0.5) was built on real 1000 Genomes Phase 3 genotypes across five super-populations, under two scenarios: identical effect sizes across ancestries (Scenario 1) and independently perturbed per-ancestry effect sizes (Scenario 2). A EUR-only discovery GWAS (plink2 --glm, 402 discovery / 101 held-out samples) fed three PRS construction methods — PRSice-2 (clumping+thresholding baseline), PRS-CSx (Bayesian continuous shrinkage, single-discovery-population mode), and LDpred2-auto (bigsnpr, self-tuning) — each built only from EUR discovery data, then scored via plink2 --score against held-out EUR plus the full AFR/AMR/EAS/SAS sample sets. A separate track applied a real published PGS Catalog height score (PGS000297) to the same genotypes as a descriptive, non-validating cross-check, since 1000 Genomes carries no real phenotype data.

What we found

PRSice-2's best EUR-tuned threshold beat PRS-CSx's same-ancestry R² in the EUR held-out set in both scenarios, but that same model lost to PRS-CSx in all eight non-EUR (ancestry × scenario) comparisons — the literature-documented pattern of genome-wide Bayesian shrinkage porting across ancestries better than a threshold-based method tuned to one ancestry. LDpred2-auto showed a similar but weaker pattern (beating PRSice-2's EUR-tuned threshold in six of eight non-EUR comparisons, trailing PRS-CSx in seven of eight). Separately, empirical per-ancestry recalibration (recentering/rescaling each ancestry's score distribution to the EUR reference) recovered 100% of calibration (mean/SD matched exactly) and, by mathematical necessity, 0% of discriminative accuracy — R² identical before and after recalibration to four decimal places across all 95 result rows, since a linear transform cannot change a Pearson correlation.

Limits of this demonstration

This is a chr21+chr22 smoke-test scope (~6% of the autosomal genome), a single simulation replicate, with per-ancestry sample sizes of 347–661 — confidence intervals on individual R² estimates are wide, and results should not be generalized to genome-wide behavior. LDpred2-auto's LD reference in this run was chr22-only (100kb bp-radius window, no genetic map available), narrower than PRS-CSx's official multi-population panel, which plausibly explains its weaker showing rather than reflecting a property of the method itself. PRS-CSx ran in single-discovery-population mode, not its namesake multi-ancestry coupling. The real-score cross-check against PGS000297 is descriptive/illustrative only, not a validated accuracy claim — not benchmarked against a client baseline, and 1000 Genomes has no phenotype data to validate against.