Do European Polygenic Scores Predict Differences Within China? New Evidence From 28 Provinces
What happens when you take genetic scores built in Europe and ask them to predict variation inside a non-European country? In our analysis of 28 Chinese provinces, we put this question to the test. The answer, while not definitive, is surprisingly consistent.
Polygenic scores (PGS) are genome wide estimates of genetic predispositions and are mostly trained on European datasets. This is a well known limitation. Because genetic architecture varies slightly across populations, PGS generally lose accuracy as the target population becomes genetically distant from the discovery population.
Africa and Europe, for example, diverge deeply in population history and PGS portability is particularly weak there. East Asians, however, are genetically closer to Europeans, and recent work suggests that for many complex traits the underlying architecture is quite similar.
That creates a natural scientific question:
If a PGS trained on millions of Europeans truly captures trait relevant genetic signal, should it generalize, at least partly, to a different continental population such as the Han Chinese?
Our paper (Piffer & Kirkegaard, 2024) tested this.
The Dataset: 28 Chinese Provinces and More Than One Hundred Thousand Genomes
The study uses allele frequencies from PGG.Han 2.0, a dataset of about 137,000 Han Chinese individuals, covering 28 provinces with adequate sample size. For each province we computed polygenic scores based on genome wide significant SNPs from major studies of:
Educational attainment (EA)
European GWAS: Lee et al. (2018) and Okbay et al. (2022)
East Asian GWAS: Feng et al. (2022) and Kim et al. (2017)Intelligence (IQ)
Two European MTAG based GWAS (Hill et al. 2019; Savage et al. 2018)Height
A multi ancestry GWAS of more than 5 million samples (Yengo et al. 2022)
A within family GWAS (Howe et al. 2022)Schizophrenia
Multi ancestry GWAS (Trubetskoy et al. 2022)
Provincial phenotypes were drawn from published surveys: averages for height, IQ (adults and children), HDI, and infant mortality.
European EA Polygenic Scores Predict Cognitive Differences Across China
One of the central results:
European derived EA PGS predict provincial IQ differences in China quite well (r = 0.52).
An EA PGS trained on East Asian samples correlates at r = 0.21, and not significantly.
This is almost certainly due to the extremely large sample size of the European GWAS, which reaches nearly 3 million individuals. Effect size estimates are simply more precise in those studies.
Still, the result is notable.
A score trained entirely outside Asia captures meaningful regional variation inside China.
This suggests that:
The genetic architecture of educational attainment is broadly shared between Europeans and East Asians, consistent with high cross population genetic correlations reported elsewhere.
Population stratification cannot explain the results, because Chinese provinces are genetically very similar and the PGS was trained in a different continent.
Height PGS Show Very Strong Portability
Height is the trait for which polygenic scores are most reliable worldwide.
Across Chinese provinces:
Multi ancestry height PGS: r = 0.71 with measured height
Within family European height PGS: r = 0.82 with measured height and r = 0.82 with latitude
This mirrors the well known pattern that northern Chinese are taller, consistent with Bergmann’s rule.
That a European within family PGS, which is resistant to confounding by parental environment, tracks a north–south gradient in China so closely is difficult to explain with any purely environmental model. It strongly suggests a real genetic gradient correlating with ecological selection pressures.
Schizophrenia PGS and Cognitive PGS Are Negatively Correlated
Across provinces, the schizophrenia PGS correlates:
−0.44 with EA PGS
−0.50 with IQ PGS
−0.67 with measured IQ
This replicates the negative genetic correlation observed in Western datasets. The same pattern appears inside China despite the GWAS being trained mainly on Europeans.
Environmental Factors Matter but Genetics Has the Stronger Signal
A regression of height on:
height PGS
HDI or infant mortality
finds:
Genetic effect: β ≈ 0.74 to 0.81
Environmental effect: β ≈ 0.22 to 0.45
Both matter, but genetics is the stronger predictor.
This aligns with global findings: environmental improvements raise average height, but they do so on top of a persistent genetic scaffold that varies across regions.
Jensen’s Method Shows Small but Significant Evidence of Polygenic Selection
The study applied Jensen’s method of correlated vectors to evaluate whether SNPs with stronger statistical support (lower p values and higher reliability) show stronger province level associations.
They do:
EA PGS: r = 0.27
Height PGS: r = 0.25
These modest but significant effects suggest that variants most likely to be real contributors to the trait also show stronger geographic structure. This is consistent with some directional selection within China, although this evidence is not definitive.
Implications for Polygenic Score Research
1. It challenges overly pessimistic claims about PGS portability.
Some argue that European derived PGS do not work in non European populations. This study shows that for traits such as height and educational attainment they can predict meaningful differences inside a relatively homogeneous East Asian population.
2. It suggests some role for evolution in shaping regional differences.
Bergmann’s rule in China, detected genetically and phenotypically, is a clear example.
3. It shows that between family confounding is not driving the geographic results.
The within family height PGS performs as well or better, so the signal cannot be an artifact.
4. It highlights the need for larger East Asian GWAS.
The East Asian EA PGS underperformed because its sample size was about 180,000, much smaller than European datasets.
A More Connected Picture of Human Genetic Architecture
At a high level, this study adds to accumulating evidence that:
PGS predict variation among world populations
among U.S. ethnic groups
and now among Chinese provinces
The predictions are not perfect, but they are consistent and often align with longstanding anthropological and ecological patterns.
No single dataset settles the question of how group differences emerge. Yet each dataset reinforces the idea that:
Genetic architecture for major complex traits is widely shared, moderately portable, and partially reflected in geographic gradients.
And the accuracy of PGS will continue to improve as global GWAS sample sizes grow.
References
Piffer, D., & Kirkegaard, E. O. W. (2024). Predictive Accuracy of Polygenic Scores from European GWAS among Chinese Provinces. Mankind Quarterly, 65(1), 58–71. https://doi.org/10.46469/mq.2024.65.1.6
Interesting findings, thanks.
"Africa and Europe, for example, diverge deeply in population history and PGS portability is particularly weak there. East Asians, however, are genetically closer to Europeans, and recent work suggests that for many complex traits the underlying architecture is quite similar."
PGS portability appears to be an excellent tool for assessing immigration policy.