European-derived coronary artery disease polygenic scores over-flag genetic risk in Vietnamese and Southeast Asian populations: a multi-score analysis in 1000 Genomes

BackgroundPolygenic scores (PRS) for coronary artery disease (CAD) are derived almost entirely from European-ancestry data. Their portability to Southeast Asian populations, including the Vietnamese, is largely uncharacterised and clinically consequential when scores are used with risk thresholds. MethodsWe evaluated four independent European-derived CAD scores from the PGS Catalog (PGS000058, PGS000349, PGS002809, PGS004198; 70-5,723 variants) in 2,504 individuals from the 1000 Genomes Project, focusing on the Vietnamese Kinh (KHV) and Dai (CDX) samples. Per-individual scores were computed with PLINK2 and standardised. We assessed (i) the cross-ancestry distribution (calibration) and (ii) a clinically-relevant consequence: the proportion of each population flagged "high genetic risk" when the European top-20% threshold is applied (20% if perfectly calibrated). ResultsFor the primary score (PGS000058) the standardised PRS differed across super-populations (ANOVA F(4, 2499) = 121.1, p < 0.001); the Vietnamese Kinh mean was +0.47 SD above the European mean (Welch t = 7.77, p = 2.0x10-14). Applying the European top-20% high-risk threshold, the fraction of Vietnamese Kinh flagged ranged from 22.2% to 57.6% across the four scores, and of Dai from 21.5% to 43.0%, versus the intended 20%. Three of the four scores over-flagged Vietnamese (25-58%); the largest score (PGS004198) was approximately calibrated for East/Southeast Asians ([~]22%) but markedly over-flagged Africans (69.3%). ConclusionsEuropean-derived CAD polygenic scores are inconsistently calibrated in Vietnamese and other Southeast Asian samples, and most substantially over-flag high genetic risk when a European threshold is applied. The magnitude and even the direction of miscalibration depend on the specific score, so no such score can be assumed transferable without local validation and recalibration. Distribution shift bounds, but does not by itself quantify, loss of predictive accuracy, which requires phenotyped data.