nature genetics

Article https://doi.org/10.1038/s41588-024-02035-8

Fine-scale population structure and

widespread conservation of genetic effect
sizes between human groups across traits

Received: 16 August 2023 Sile Hu 1,2,3 , Lino A. F. Ferreira 1,2, Sinan Shi 1, Garrett Hellenthal4,5,9,
Jonathan Marchini 6,9, Daniel J. Lawson 7,8,9 & Simon R. Myers 1,2,9
Accepted: 18 November 2024

Published online: 3 February 2025

Understanding genetic differences between populations is essential for
Check for updates avoiding confounding in genome-wide association studies and improving
polygenic score (PGS) portability. We developed a statistical pipeline to
infer fine-scale Ancestry Components and applied it to UK Biobank data.
Ancestry Components identify population structure not captured by
widely used principal components, improving stratification correction for
geographically correlated traits. To estimate the similarity of genetic effect
sizes between groups, we developed ANCHOR, which estimates changes in
the predictive power of an existing PGS in distinct local ancestry segments.
ANCHOR infers highly similar (estimated correlation 0.98 ± 0.07) effect
sizes between UK Biobank participants of African and European ancestry
for 47 of 53 quantitative phenotypes, suggesting that gene–environment
and gene–gene interactions do not play major roles in poor cross-ancestry
PGS transferability for these traits in the United Kingdom, and providing
optimism that shared causal mutations operate similarly in different
populations.

Genome-wide association studies (GWAS) have uncovered numerous biases effect size estimates6,7, complicating studies of temporal trait
genetic influences on complex human traits, regulated by many loci evolution6 and comparisons between human groups. Moreover, PGS
with small effect sizes. For traits such as height, large sample sizes in accuracy drops in populations with different ancestry from the GWAS8,9
European groups have allowed PGS to explain a substantial fraction of cohort, particularly those with strong genetic differentiation. This lack
heritability by summing effects across many single nucleotide poly- of portability is partly caused by genetic drift making some causal vari-
morphisms (SNPs) genome-wide1. Because the true causal variants are ants group-specific, and population-specific linkage disequilibrium
unknown, SNPs in PGS are expected not to influence a trait directly, but (LD) patterns affecting tagging accuracy.
rather correlate with (‘tag’) a true causal mutation. Genetic stratifica- As well as such ‘local’ differences, recent studies10–14 have sug-
tion is a major confounder in GWAS, potentially causing false positives gested that causal variants have different effect sizes on traits in dif-
when phenotypes correlate with stratification. Methods including ferent groups, because of ‘global’ (nonlocal) factors. Changes in causal
mixed models2,3 and principal component analysis4,5 have proved SNP effect sizes, defined as their mean effects on a trait of interest, can
powerful in solving major issues, but subtle population structure still arise by either gene–gene interactions not captured by the additive

Department of Statistics, University of Oxford, Oxford, UK. 2Wellcome Centre for Human Genetics, University of Oxford, Oxford, UK. 3Human Genetics
1

Centre of Excellence, Novo Nordisk Research Centre Oxford, Oxford, UK. 4Department of Genetics, Evolution and Environment, University College
London, London, UK. 5UCL Genetics Institute, University College London, London, UK. 6Regeneron Genetic Center, Tarrytown, NY, USA. 7Department of
Statistical Science, School of Mathematics, University of Bristol, Bristol, UK. 8MRC Integrative Epidemiology Unit, Population Health Sciences, Bristol
Medical School, University of Bristol, Bristol, UK. 9These authors contributed equally: Garrett Hellenthal, Jonathan Marchini, Daniel J. Lawson, Simon R.
Myers. e-mail: [email protected]; [email protected]; [email protected]

Nature Genetics | Volume 57 | February 2025 | 379–389 379

Article https://doi.org/10.1038/s41588-024-02035-8

model, or gene–environment interactions, differing across popula- Our pipeline uses this panel to generate an ancestry decom-
tions. Varying effect sizes have been documented in populations9, so position for a new ‘target’ sample (Fig. 1a), extending our previous
potentially contribute to differences between populations. However, approaches22,33, by closely mirroring the steps used to generate the
other studies suggest similarities across groups in underlying effect reference panel itself. Unified panel markers are phased and imputed
sizes15–19 and, for stronger GWAS hits, in direction of effect at least20. in the target, using the phasing and imputation panel. For UKB, hap-
One recent study14 leverages individuals of mixed African and Euro- lotype data are prephased. Imputed haplotypes are then matched
pean ancestry, decomposing local ancestry to test whether causal vari- against the labeled painting reference panel using ChromoPainter
ants on African versus European chromosomes show different effect to quantify recent ancestor sharing. Finally, a non-negative least
sizes. By focusing on within-individual comparisons, this approach square (NNLS) based approach (Methods)22 is used to infer ancestry
eliminated factors including gene–environment interactions14,21 and coefficients from ChromoPainter output by fitting the observed
long-range gene–gene interactions21 that might differ across popu- haplotype-matching vector as a mixture of those from the 127 ref-
lations. It instead examines whether, for example, local gene–gene erence panel groups. This approach leverages haplotype informa-
interactions alter effect sizes on different ancestral backgrounds. tion to infer population structure33 alongside Hidden Markov Model
The study inferred strong sharing of underlying effect sizes at this approximations to the coalescent33, analyzing each sample in minutes
within-individual level, suggesting local interactions are not major (Methods).
factors for most traits. However, it also inferred strong differences We applied our pipeline to 487,409 UKB participants. Because
in causal effects across individuals possessing different continen- our pipeline uses ‘out-of-sample’ comparisons based on population
tal ancestries, possibly because of gene–environment interactions. structure, analyzing individuals independently, it captures popula-
Resolving the role of such interactions is vital to successfully apply tion structure information in the United Kingdom and Ireland but not
genetic findings across groups, design efficient studies and understand genetic relatedness between samples. Ancestry is defined through
whether evolution or environmental differences drive interpopulation subjective groupings of reference individuals using genetic clusters
trait variation. Here, we develop an approach to do this, using models identifiable via fineSTRUCTURE. Because the true genetic ancestry of
in which genetic effect sizes are shared across ancestries within indi- UKB individuals is unknown, we assessed the pipeline’s effectiveness by
viduals (as observed in ref. 14), but may vary between individuals with examining the relationship between inferred ancestry and individuals’
mainly European or African ancestry. birthplace and self-reported ethnicity, the ability of Ancestry Compo-
To analyze the impacts of population structure, we introduce a nents (ACs) to predict principal components (PCs) and the capacity of
fine-scale ancestry pipeline. We use ‘ancestry’ throughout to refer to our ACs to correct for ancestry in GWAS, compared with widely used
sharing inferred most-recent ancestors with someone from a particular approaches including PCs and BOLT-LMM4,34. We also developed an
self-reported ethnicity or geographic region. Expanding on our previous expectation-maximization algorithm (Methods and Supplementary
work22–24, we created a pipeline inferring individuals’ ancestry contribu- Note 1) to estimate allele frequencies across SNPs. This method pro-
tions from 127 geographically meaningful and genetically recoverable vides frequency estimates for each of 25,485,700 UKB imputed muta-
(Methods) regions worldwide. We applied this pipeline to all 487,409 tions25 within 127 reference groups, aiding studies of regional allele
participating UK Biobank25 (UKB) individuals. We show that using frequencies within the United Kingdom.
detailed ancestry information in GWAS better corrects for population
stratification versus other state-of-the-art methods2,4,26, reduces likely Fine-scale population structure across the United Kingdom
false positives and uncovers previously undiscovered associations, but As a first test, the mean ancestries for the 434,781 UKB participants
nonportability remains strong. We also develop ANCHOR, a statistical born in the United Kingdom or Republic of Ireland with self-reported
inference approach to estimating cross-population similarity in causal white British and/or Irish (WBI) ethnicity were: British–Irish (BI), 94.9%;
effect sizes using admixed individuals, complementing approaches Dutch, 1.35%; Swiss, 0.79%; Norwegian, 0.49%; Polish, 0.29%; and Dan-
including POPCORN10 and XPASS27 that only apply to nonadmixed ish, 0.19%. For participants whose self-reported ethnicity is ‘other white
individuals (‘Discussion’). Notably, ANCHOR requires no prior assump- background’, the BI proportion drops to 25.5%. For individuals born in
tions about the underlying effect size distribution. As in ref. 14 and other the Republic of Ireland, inferred Irish ancestry averages 74.2%, within
recent studies11–13, ANCHOR decomposes ancestry at fine scales along 98.4% BI ancestry, demonstrating the pipeline’s accuracy in capturing
the genome. It assigns mutations to specific ancestries to quantify the geographic and ethnicity information (Extended Data Fig. 1).
contributions of gene–environment interactions on differences of The average ancestry proportion from 22 British Isles regions
predictive power for individuals with different ancestries. Application (Methods) varies by birthplace (Fig. 2a). These ancestry proportions are
of ANCHOR to 8,003 UKB mixed-ancestry individuals yields different based on an out-of-sample dataset of UK individuals with grandparents
findings from recent studies, which we discuss along with implications. born within an ~80-km radius in each region23. Mean BI ancestry associ-
ated with a region decreases with geographic distance from that region,
Results indicating that DNA information is informative for birthplace (Fig. 1b
A statistical pipeline to infer precise individual ancestry and Extended Data Fig. 1). Some 41.5% of UK-born individuals have more
We developed an approach able to decompose the fine-scale ancestry of than 50% of their ancestry from a single region, matching their birth-
a genome into a mix of 127 regions, including 23 in the United Kingdom place 59.2% of the time, rising to 82.7% after expanding to neighboring
and Ireland (Supplementary Table 1). The UKB dataset comprises many regions (Supplementary Table 2a). There is a strong correspondence
individuals of mainly British ancestry, alongside a substantial fraction between self-identified ethnicity and birthplace for non-UK ancestries
with ancestry from elsewhere. Our approach identified 105 regions (Fig. 2b and Extended Data Fig. 2). However, regional variation exists:
present in at least 5 individuals for at least 10% of their ancestry. Our ancestry localization is weaker in southern and eastern England23, and
pipeline leverages data from previous studies23,28–31 of human genetic stronger in Scotland, Wales, Northern England and South West England.
variation, and uses methods25,32 to impute these data on a common set An entropy-based statistic (Methods) shows regional mixing, with
of variants for high-quality imputation. This generates a unified ‘refer- London having the highest entropy and the highest average fraction of
ence panel’ of haplotypes (Methods and Fig. 1a), phased using SHAPEIT2 ancestry outside the United Kingdom. Other major cities and South East
(ref. 32) to ensure consistent data quality. Reference haplotypes are England also show strong mixing (Extended Data Fig. 3). Non-British
labeled through semi-supervised clustering with ChromoPainter33 and ancestries also vary geographically, with higher Irish ancestry (>10%)
fineSTRUCTURE33, combined with geographic and ethnicity labels, to seen in Liverpool, Birmingham, Manchester and London; Dutch ances-
produce a painting reference panel. try found in the south of England and around Bristol (3.5%); and Polish

Nature Genetics | Volume 57 | February 2025 | 379–389 380

Article https://doi.org/10.1038/s41588-024-02035-8

a
Ancestral inference pipeline

Phasing and imputation Painting reference panel

reference panel Pop. 1 Pop. 2 Pop. 3 Pop. 4 Pop. 5

Phasing Imputation Painting Fitting

Genotype

= A1Pop. 1 + A2Pop. 2 + A3Pop. 3 + A4Pop. 4 + A5Pop. 5

b
North Yorkshire South Yorkshire South East England

0.6 0.30
0.35
0.5 0.25
0.30

0.25 0.4 0.20

0.20 0.15
0.3
0.15
0.2 0.10
0.10
0.1 0.05
0.05

Fig. 1 | Schematic diagram of the ancestral inference pipeline and the fitting stage, non-negative coefficients summing to one and representing the
performance in the UKB BI individuals. a, Main steps in the ancestral inference proportions of ancestries from the labeled groups in the reference panel are
pipeline. The pipeline accepts individual genotype data (microarray or inferred. b, Geographical average proportion of DNA in BI individuals, positioned
sequencing data) as input. The genotype data are phased and imputed against according to their birthplaces (Methods), inferred to come from three regional
a phasing and imputation reference panel in first ‘Phasing’ and ‘Imputation’ groupings: North Yorkshire, South Yorkshire and South East England (which
stages, and painted against the painting reference panel including preclustered correspond to the excess ancestry locations colored red). Pop., population; A,
groupings (5 groups in this example; 127 in our actual analysis). In a final mixture proportion of ancestries from each population.

ancestry peaking near Wrexham in Wales, the site of the second highest PCs from these 127 ACs, and conversely predicted 16 common BI ACs
concentration of Polish burials in the United Kingdom35. from the first 140 UKB PCs (Methods, Fig. 3a,b and Supplementary
For non-UK-born individuals, we again achieve fine-scale ancestry Figs. 1 and 2). We found a good prediction of most PCs from ACs, but
resolution (Fig. 2b, Extended Data Fig. 2 and Supplementary Table 2b). not the converse, indicating that ACs capture additional information
Although some countries (for example, Philippines and Japan) show (Supplementary Fig. 1). PC-based prediction was also poor for non-UK
genetic homogeneity, perhaps because of small sample sizes, most regions (Supplementary Fig. 2). Second, we performed GWAS on 104
exhibit diverse ancestry patterns. Because non-UK-born individuals UKB quantitative traits with more than 10,000 unrelated white British
reflect people who made their home in the United Kingdom, rather than individuals (Methods). We compared the effectiveness of ACs versus
unbiased samples from birthplace countries, we observe widespread BI PCs for correcting population stratification by using LD-score regres-
ancestry. Other patterns are present, such as the over-representation sion (LDSC; Extended Data Fig. 4)37, which measures systematic inflation
of Gujarat ancestry among UKB individuals born in Uganda and Kenya, because of uncorrected stratification. An intercept estimate close to
which might be explained by the post-colonial exodus of South Asians 1 indicates effective correction, although intercepts slightly above 1
from East Africa. can occur for highly heritable traits like height, or large sample sizes
in practice37,38.
ACs aid GWAS stratification for geographically linked traits Birth location is often used as a control phenotype to test for
To avoid false-positive GWAS associations, correcting for population stratification39 because few SNPs are likely to be causally associated
stratification is crucial. PCs are widely used for this purpose4,5,36, either with it. For latitude (Fig. 3c), PC-based correction or BOLT-LMM analysis
alone or with mixed-model analysis3,34. However, selecting the number (Fig. 3c, Supplementary Fig. 3a and Supplementary Table 3a,b) yielded
of PCs is nontrivial; too few can result in false positives, whereas too many apparent independent hits, and substantial genome-wide infla-
many can cause overcorrection because of extensive LD (for example, tion (LDSC intercept of 1.6608; Extended Data Fig. 4a,b), even with 100
beyond the first 40 publicly released UKB PCs36). We compared the PCs (Supplementary Fig. 3b). By contrast, AC-based correction reduced
use of 127 ACs with PCs in GWAS. First, we predicted the first 16 UKB the number of association signals from 470 to 7 (Fig. 3c; with five shared

Nature Genetics | Volume 57 | February 2025 | 379–389 381

Article https://doi.org/10.1038/s41588-024-02035-8

a
34,484 n = 92

4,699 2,585

2,123 2,357

83,091 42,765

2,556 12,408

1,029 50,244

39,597 25,636

34,808 3,062

5,201 6,051

619 5,751

344 58,585

6,635

East Anglia South East England Devon Lincolnshire Cyprus Switzerland Syria
South Central England Merseyside North West Wales Cumbria Germany Balkans Denmark
Central England South England South Yorkshire North West Scotland Lithuania Portugal
North Central England South East Wales Cornwall North Ireland and South Scotland Norway Netherlands
North Yorkshire North East Scotland Cornwall Tip Ireland Poland Italy
Northumberland South West Wales Orkney France Spain Tuscany

b Turkey (153) Iraq (325) Iran (520) Afghanistan (101) Pakistan (1412) Nepal (134) Japan (261)

Oth Whi Asi Oth Whi Asi Oth Whi Asi Oth Pa Asi Oth Asi Oth
Lebanon (72) China (408)

Oth Bri Whi Chi

Israel (86) Hong Kong (640)

Oth Bri Whi Chi Bri

Yemen (104) Vietnam (66)

Ind Oth Bri Asi Chi Oth

Kenya (1651) Philippines (332)

Ind Bri
India (3921) Asi Oth
Uganda (608) Malaysia (666)
Worldwide
Africa
North Europe Ind Bri
Ind Bri Afr Chi Bri
Sri Lanka (730) Myanmar (Burma)
Tanzania (421) Indonesia (62)
South Europe
West Europe
Pacific America Asi Oth Asi Oth Mix Bri
Ind Bri
Chi Oth Bri Whi
Malawi (109) Mauritius (384) Bangladesh (24) Thailand (99) Singapore (496) Australia (919) New Zealand (680)

Ind Mix Bri Afr Ind Asi Oth Ind Oth Mix Bri Asi Chi Oth Bri Chi Bri Bri Whi Bri Whi

Fig. 2 | Ancestry inference for UKB individuals born in the United Kingdom or African countries. Colors are as shown on the map, with colors for ancestry from
Ireland and worldwide. a, Ancestry inference stratified by birthplace region for additional regions given in the legend. White lines on the map delineate the
UKB WBI individuals; for each regionally labeled bar plot, each column shows borders of different countries. Self-reported ethnicity labels are shown below
ancestry decomposition for a single individual, with colors representing regions each bar plot. Color legends differ in a and b. Self-reported ethnicity: Afr, African;
shown on the map and numbers representing counts of individuals from each Asi, Asian; Bri, British; Chi, Chinese; Ind, Indian; Ire, Irish; Mix, mixed; Other,
area. b, As a, but showing decomposition for Asian, Oceanian and selected East other ethnic group; Pa, Pakistani (Asia); Whi, white (Europe).

Nature Genetics | Volume 57 | February 2025 | 379–389 382

 Article https://doi.org/10.1038/s41588-024-02035-8

a b
0.8 0.5 0.05 1.0 1.0 1.0
PC2 PC3 PC4 South 0.8 North West 0.8 Northumber- Ireland
0.6 PC1 1.0 0.8
0.3 0.1 –0.05 0.8 Yorkshire 0.6 England 0.6 land 0.6
0.4 0.6 0.4 0.4 0.4
0.1 0 0.4 0.2 0.2 0.2
0.2 –0.15 0.2
0 0 0 0
–0.1
R2 = 0.999 –0.1 0.996 0.987 0.93 R2 = 0.526 –0.2 0.438 –0.2 0.423 –0.2 0.734
0 –0.25 –0.4
0
0.2
0.4
0.6
0.8

–0.1

0.1

0.3

0.5

–0.1
0
0.1

–0.25

–0.15

–0.05

0.05

–0.2
0
0.2
0.4
0.6
0.8

–0.2
0
0.2
0.4
0.6
0.8
1.0

0
0.2
0.4
0.6
0.8
1.0
1.0

–0.4

–0.2
1.0
0.8
0.6
0.4
0.2
0
0.05 0.05 1.0 North Central 1.0 South Central 1.0 North Ireland 1.0 Nigeria
PC5 PC6 PC7 0.05 PC8
0.05 0.8 England 0.8 England 0.8 0.8
0.03 –0.10
–0.05 0.6 0.6 0.6 0.6
0.01 –0.25 –0.05 0.4 0.4 0.4 0.4
–0.01 –0.40 –0.15 0.2 0.2 0.2 0.2
–0.15 0 0 0
–0.03 0.883 –0.55 0.851 0.681 –0.25 0.747 0.197 0.228 0.303 0 0.972
–0.2 –0.2
–0.03
–0.01
0.01
0.03
0.05

–0.55
–0.40
–0.25
–0.10
0.05

–0.15

–0.05

0.05

–0.25

–0.15

–0.05
0.05

–0.2
0
0.2
0.4
0.6
0.8
1.0

–0.2
0
0.2
0.4
0.6
0.8
1.0

0
0.2
0.4
0.6
0.8
1.0

0
0.2
0.4
0.6
0.8
1.0
0.06 1.0 1.0 1.0 1.0
0.10 South East South East North West North
0.04 PC9 PC10 0.10 PC11 0.05 PC12 0.8 England 0.8 Wales 0.8 0.8
0.05 Scotland Yorkshire
0.6 0.6 0.6 0.6
0.02 0 0.05 –0.05 0.4 0.4 0.4 0.4
0 –0.05 0.2 0.2 0.2 0.2 0.062
0
0 0
PC predicted by AC

AC predicted by PC
–0.02 0.79 –0.10 0.722 0.713 –0.15 0.677 0.171 0 0.247 0.394 0

0
0.2
0.4
0.6
0.8
1.0

0
0.2
0.4
0.6
0.8

0
0.2
0.4
0.6
0.8
1.0

0
0.2
0.4
0.6
0.8
1.0
–0.02
0
0.02
0.04
0.06

–0.10
–0.05
0
0.05
0.10

0.05

0.10

–0.15

–0.05

0.05

1.0
0.2 1.0 1.0 1.0 1.0
0.02 South West
PC13 PC14 0.10 PC15 PC16 0.8
Cumbria Central
0.8 England 0.8
Anglia
0.8
0.1 0.01 0.02 Wales
0.05 0.6 0.6 0.6 0.6
0 0 0 0.4 0.4 0.4 0.4
–0.01 0.117 0.2 0.102 0.084 0.2
–0.05 –0.02 0.2 0.2
–0.1 –0.02 0
0.722 0.577 0.631 0.47 0 0 0 0.629
–0.10
–0.1

0.1
0.2

–0.02
–0.01

0.01
0.02

–0.10
–0.05
0
0.05
0.10

–0.02
0
0.02

0
0.2
0.4
0.6
0.8
1.0

0
0.2
0.4
0.6
0.8
1.0

0
0.2
0.4
0.6
0.8
1.0
0
0.2
0.4
0.6
0.8
1.0
UKB PC UKB AC

c d e
Birth place North/South Employment score England Waist circumference
25 25 25

20 20 20

160
15 15 15
120

80

10 10 10 40

40 80 120 160
–log10(PPC)

5 5 5

0 0 0
0 5 10 15 20 25 0 5 10 15 20 25 0 5 10 15 20 25
–log10(PAC)

Allergy Cognitive function Blood pressure Other medical condition Autoimmune Education attainment Anthropometric
No GWAS hits Lipids Blood counts Lifestyle Population density Socioeconomic status Not queried

Fig. 3 | Comparison between AC-corrected and PC-corrected GWAS. a,b, and waist circumference (e). All plots are colored according to the legend shown
Predictions are based on ‘linear’ combinations of ACs or PCs. a, Prediction of first at the bottom, indicating earlier evidence from GWAS for each SNP in particular
16 UKB PCs (x axes) using a linear model-based prediction from the 127 ACs (y phenotypic categories (gray: SNPs show no prior GWAS evidence, perhaps
axes) shows strong correlations (R2 values). b, As a, but now predicting 16 UK and consistent with likely false-positive associations). The horizontal and vertical
worldwide ACs from 140 PCs, often showing poor prediction. c–e, Comparison dark blue lines indicate the genome-wide P-value threshold (P = 5 × 10⁻⁸) in a −
of AC-corrected (x axis) and PC-corrected (y axis) −log10(P values) for SNPs in log10 scale, while the light blue line represents y = x. In each plot, the points show
three exemplar GWAS for labeled traits: birthplace (c), employment score (d) only independent SNPs with P < 5 × 10−8 for one or both approaches.

hits), implying that fine-scale ancestry information can effectively association. Because the protective allele against hay fever is more com-
remove stratification impacts. Similar results were observed for mon in southern England where hay fever is most prevalent (Extended
home location (Extended Data Fig. 4a,b and Supplementary Fig. 3c,d). Data Fig. 5), such geographic differentiation44 might reflect selective
Although any residual birthplace hits may reflect inadequate adjust- migration of people carrying this variant and/or past natural selection.
ment for stratification, the few association signals identified using ACs We also performed a GWAS for ‘employment score England’, a trait
show a modest enrichment in SNPs showing previous GWAS evidence defined based on the region in which a person lives. GWAS associations
(Supplementary Table 4; odds ratio (OR) = 5.9, P = 0.039 compared with such traits have been observed6, but such associations can be con-
with 12% PC-corrected hits with previous GWAS evidence). Specifically, founded by regional stratification. Indeed (Fig. 3d), ACs and PCs shared
the strongest remaining signal (P < 10−15) after using ACs surrounds a number of hits, but some signals were only observed in the PC-based
rs5743618 on chromosome 4, linked to toll-like-receptor genes involved analysis, and AC-based showed a stronger enrichment of previous
in innate immunity40,41, including TLR1 and TLR10, one of the strongest GWAS signals (Supplementary Table 4; OR = 11, P = 0.0013) compared
hay fever hits in European GWAS cohorts42,43, with weaker asthma risk with PCs. In total, 18% of PC-only associations overlapped with previous

Nature Genetics | Volume 57 | February 2025 | 379–389 383

Article https://doi.org/10.1038/s41588-024-02035-8

GWAS signals, similar to the null trait of latitude (OR = 1.6, P = 0.44), and European-like segments, yielding ‘African PGS’ (APGS’ and ‘Euro-
compared with 71% for AC-based associations (OR = 18, P = 1.8 × 10−10 pean PGS’ (EPGS) scores for each individual, each ‘mean-centered’
versus latitude). Moreover, many AC-based signals were significant in by ancestry-dependent mean SNP frequencies, a critical step both in
previous GWAS for educational attainment or socioeconomic status theory (Supplementary Note 2) and in simulations. The phenotype
(Fig. 3d and Supplementary Table 3c). These trait classes were fur- is regressed against APGS and EPGS, including other covariates, to
ther enriched relative to PC-only hits (OR = 16, P = 0.005), suggesting estimate coefficients βAf and βEu, which quantify ancestry-specific PGS
PC-only hits are likely false positives (Fig. 3d). These findings suggest predictive power as the respective average trait increases in admixed
that previous GWAS for ‘regionally defined’ traits can produce false individuals (Supplementary Note 2) per unit APGS or EPGS increase.
positives, and ACs help correct for this. Finally, βEu is then compared with the corresponding βObs.Eu estimate
Among the other 99 nonregionally defined traits, PC-based and from BI individuals to estimate ρ, as βEu/βObs.Eu (Supplementary Note
AC-based correction provided similar P values and LDSC intercepts, 2). Importantly, the validity of the ρ estimation requires no assump-
with subtle improvements for AC (Extended Data Fig. 4a,b). However, tions regarding underlying effect size distribution, causal mutation
differences included five hits unique to the AC-corrected analysis frequencies, LD patterns or selection. Under a local effect only (‘null’)
(Methods, Fig. 3e and Supplementary Table 3e), some in strong LD model, causal effect sizes are identical in all European-ancestry regions
with known GWAS hits for the same traits43,45–47, and thus unlikely to and ρ = 1. If nonlocal ancestry-specific (gene–gene or gene–environ-
be false positives. For traits including waist circumference (Fig. 3e and ment) interactions occur, ρ ≠ 1, and if the variance of underlying genetic
Supplementary Table 3e), AC-specific hits occurred at SNPs in regions effect sizes is the same in each group, then ρ < 1 (Supplementary Note
of strong LD with high regional SNP loadings for particular PCs (Sup- 2). Therefore, testing for ρ ≠ 1 provides a test for differing effect sizes
plementary Table 5a–e and Extended Data Fig. 6). Therefore, applying across groups. Assuming effect sizes scale linearly with genome-wide
ACs likely removed false negatives caused by PC overcorrection from ancestry, ANCHOR can predict ρ for 100% European or African ances-
PCs strongly correlated with genotypes in particular genomic regions. try without sampling (Methods and Supplementary Note 2). Pooling
ρ values among traits helps reduce overall uncertainty, and confi-
Causal effects are similar across ancestries dence intervals (CIs) for all coefficients are obtained by bootstrapping
To study PGS portability across UKB groups, we created PGS using (Methods).
343,047 white British individuals for 53 heritable (estimated herit- We verified HAPMIX’s48 ability to accurately infer ancestry and
ability >5%) UKB-measured quantitative traits, correcting for ancestry construct ancestry-specific PGS by comparing it with trios with known
using either ACs or PCs, and tested their performance in independent phase (Extended Data Fig. 7 and Supplementary Fig. 7). We then tested
samples representing different ancestries; that is, with more than 50% ANCHOR by simulating 24 quantitative traits with various settings of
inferred ancestry (by ACs) from seven respective labeled regions: South heritabilities, clustering of causal mutations and causal marker fre-
Central England, Northumberland, Republic of Ireland, Poland, India, quency spectra, using genetic data from the 8,003 African-ancestry
China and West Africa. AC-based versus PC-based correction yielded individuals (Methods and Supplementary Note 1). We performed GWAS
different group-specific PGS means, but strong within-group corre- and downstream analyses exactly as for the real phenotypes. In nonad-
lations (92.7% to 99.9%) (Supplementary Fig. 4 and Supplementary mixed populations, mean-centering of genotypes has no impact on the
Table 6) so we focus on AC-corrected PGS for the remaining analyses. PGS predictive power, but is crucial for ANCHOR’s validity in admixed
By regressing the PGS against actual traits in each group, we quanti- populations (Supplementary Note 2). Under the null (ρ = 1), across sim-
fied the increase in trait per unit PGS increase (that is, the regression ulations ρ is correctly estimated from mean-centered EPGS and APGS,
slope) and denoted it as β, and denoted the variance explained by the but estimates are strongly downward biased without mean-centering
PGS as ΔR2, after regressing out covariates (Methods). Both ΔR2 and β (Extended Data Fig. 8 and Supplementary Figs. 8 and 9), even when
decreased with increasing genetic distance from the British ancestry masking short or uncertain segments (Supplementary Figs. 8a and 9).
groups (Supplementary Figs. 5 and 6), particularly for sub-Saharan Therefore, we use mean-centering for all ANCHOR analyses. We also
African ancestry, with a >2.2-fold reduction in ΔR2. Fine-scale ances- observe significantly reduced ρ estimates for real quantitative traits
try showed this occurs even between BI regional ancestries for some without mean-centering, fully consistent with the simulation findings
traits: for standing height, forced expiration volume in 1 s (FEV1) and (Supplementary Figs. 10–12).
apolipoprotein B, ΔR2 for Northumberland differs from that for Ireland If ρ = 1, 95% bootstrapped CIs for 96% of individual traits con-
(P < 0.02, P < 0.005 and P < 0.02, respectively). tained the true value (ρ = 1), with robustness to different GWAS P-value
To partition the drop-off in PGS performance across ances- thresholds (0.05 versus 0.0001) (Supplementary Fig. 13). Averaging
tries into local effects (causal variant tagging) and nonlocal effects ρ estimated by ANCHOR across traits for groups of individuals with
(ancestry-specific gene–gene or gene–environment interactions), we similar ancestry levels also yielded ρ = 1 (Extended Data Fig. 9), indicat-
developed ANCHOR, which leverages variation in local ancestry along ing good performance under the null across both traits and ancestry.
the genome and between admixed individuals (Fig. 4a). ANCHOR takes In simulations in which ρ declines, ANCHOR estimates of ρ remain well
PGS coefficients from an independent sample as input, and analyzes calibrated (Fig. 4b, Extended Data Fig. 9 and Supplementary Fig. 14). In
quantitative phenotype and genotype data for a group of admixed indi- all cases we observe βAf/βObs.Eu < 1 as expected because of local effects
viduals. It produces an estimate of ρ, the mean trait increase in admixed (Supplementary Note 2), so African-ancestry segments are less predic-
individuals per unit increase in a perfect PGS for nonadmixed (for tive of traits, but βAf values varied similarly to βEu (Extended Data Fig. 9
example, European) individuals, constructed using their (unknown) and Supplementary Fig. 14).
true effect sizes. ρ equals the correlation in true effect sizes between We applied ANCHOR to 53 UKB quantitative phenotypes and found
populations under reasonable assumptions (Supplementary Note strikingly constant average ρ = βEu/βObs.Eu estimates across genome-wide
2). For our analysis, we generated PGS coefficients from 343,047 UKB ancestry bins (Fig. 4c and Methods), overlapping the null value ρ = 1
white British samples, and analyzed 8,003 ‘African ancestry’ individuals for all bins. This indicates that European-ancestry segments retain
with varying (mean 83.6%) inferred sub-Saharan African ancestry and predictive power similar to that in European-ancestry individuals.
BI + Europe (mean 11.5%) ancestry. Thus, a ‘true’ PGS from Europeans predicts similar trait increase, on
ANCHOR estimates local ancestry along the phased genome, for average, per unit PGS increase, in African-ancestry individuals. This
example, using HAPMIX48 here, masking uncertain and short segments implies either conserved effect sizes between groups at shared causal
to ensure ancestry segments extend further than LD (Methods and SNPs, across the broad range of human molecular and quantitative
Supplementary Note 2). It calculates separate PGSs for African-like phenotypes we examined or—less parsimoniously—systematically

Nature Genetics | Volume 57 | February 2025 | 379–389 384

 Article https://doi.org/10.1038/s41588-024-02035-8

a b βEu/βObs.Eu
All

1.2
[0.10, 0.35] n = 373
European African 1.1 [0.35, 0.55] 560
[0.55, 0.78] 475
1.0
[0.78, 0.95] 2,556
0.9

(Weighted mean) estimated ratio

0.8

0.7

0.6 n = 8,003

0.5

0.4

0.3

0.2

Tag SNP 0.1

Causal SNP 0

1.0
3

05

01
0.

0.
0.
0.

0.

00
0.
True effect size correlation

0.
=
P

=
(simulation)

P
[0.00, 0.10] n = 20,107 [0.35, 0.55] 560 [0.78, 0.95] 2,556 53 UKB traits

c 1.2
[0.10, 0.35] 373 [0.55, 0.78] 475 [0.95, 1.00] 4,039 d
Standing height
1.1
0.5 n = 8,003

Height increase per unit PGS increase

1.0 Af Eu
European βEu βEu
(Weighted mean) estimated ratio

0.9
0.4 βObs.Eu All
Impact of local LD

n = 1,491
βEu
5,355 1,502
0.8
13,261

0.7 0.3
4,438

0.6 1,266
Eu
0.5 0.2 βAf
7,104
All Af
0.4 African βAf βAf

0.3 0.1

0.2

0.1 Impact of global ancestry 0

Af Eu Af Eu Af Eu
ic Ch a
Eu a
Po nd
nd

pe
Ire nd
be nd

Local ancestry
i
a/ in

0
d

0 21 46 71 89 99
la
la

ro
rla
a

In
um gl

All Af Eu
th n

Percentage African ancestry (bin average)

or l E

Global ancestry
N tra

fr
en

tA
/C

es
th

W
u
So

Fig. 4 | Separation of local and nonlocal factors influencing portability. a, Test ANCHOR for 53 UKB traits across varying African ancestry binned as shown (x
principles: in UKB samples with European (blue) and African (red) ancestries, a axis; colored regions). For each bin, mean estimates across traits of ratios
causal variant contributing to a trait is captured by a tag SNP whose predictive βEu /βObs.Eu (blue) and βAf /βObs.Eu (red) are shown. Also shown are ratio estimates
power (pink arrow thickness) varies by ‘local’ ancestry (upper versus lower for individuals of ~100% European (leftmost point at y = 1) or ~100% African (red
chromosomes), or nonlocal factors captured by genome-wide ‘global’ ancestry horizontal bar) ancestry. CIs crossing y = 1 are consistent with identical effects to
(left versus right individuals); ANCHOR separates these contributions to PGS ~100% European-ancestry individuals, and similarly for red points or bar. d, Mean
portability. b–d, βij values refer to the mean increase in phenotype per PGS unit increase in standing height per PGS unit increase across populations (seven
increase for local ancestry j and global ancestry i (see Methods for further left-hand columns); alongside corresponding ANCHOR estimates for height
details). b, ANCHOR performance for 24 simulated traits and 53 UKB quantitative (final six columns) labeled by global or local ancestry combinations. Data are
traits with PGSs constructed using different P-value thresholds (P = 0.05 and presented as (weighted) means (b,c) or as estimated values (d) with 95% central
P = 0.0001; right). True effect size correlations ρ (x axis) between African and bootstrapped CIs. Error bars indicate 95% bootstrapped CIs, Af, African; Eu,
European ancestries are compared with the ANCHOR estimator βAll Eu
/βObs.Eu (y European.
axis). Colors denote African ancestry bins, as defined in c. c, Application of

larger average effects across all these phenotypes in African-ancestry ancestry or African ancestry genome-wide (Figs. 4d and 5). For stand-
individuals, in such a manner as to coincidentally balance the impact ing height (Fig. 4d), we show βAf and βEu estimates alongside βObs.Eu, and
of incomplete correlation. β estimates for various AC-defined UKB cohorts. β estimates decline
For individual traits, as in the simulations we combine all individu- with increasing genetic distance because of LD changes, reducing PGS
als to estimate a trait-specific ρ, and extrapolated to estimate effect predictive power. However, ANCHOR shows that European-ancestry
sizes for European segments in individuals with almost 100% European segments in African-ancestry genomes have strong predictive power,

Nature Genetics | Volume 57 | February 2025 | 379–389 385

Article https://doi.org/10.1038/s41588-024-02035-8

Global All Af All

βLocal n = 8,003 βEu/βObs.Eu βEu/βObs.Eu βAf/βObs.Eu
Standing height
(PGS catalog)
Standing height
Sitting height
FVC
FEV1
Basal metabolic rate
Whole body water mass
Weight
Leg fat percentage (left)
Arm fat free mass (right)
Platelet count
(Weighted) mean
(Other 16 UKB non-molecular traits)
Cholesterol
HDL
LDL
Triglycerides
Apolipoprotein B
Apolipoprotein A
ALP
Total bilirubin
IGF1
Urate
(Weighted) mean
(Other 17 UKB molecular traits)
(Weighted) mean
(all 53 UKB traits, P = 0.05)
(Weighted) mean
(all 53 UKB traits, P = 0.0001)

0 0.5 1.0 1.5 2.0 0 0.5 1.0 1.5 2.0 0 0.5 1.0 1.5 2.0

Fig. 5 | ANCHOR results for 53 UKB traits. Data are presented as estimated values African ancestry and (as expected, reduced) predictive power for African-
of ratio of true effect sizes with 95% central bootstrapped CIs. βij values refer to ancestry segments. From top to bottom, the rows above and below the first
mean increase in phenotype per PGS unit increase for local ancestry j and global horizontal dashed line represent non-molecular and molecular traits and rows
ancestry i (see Methods for further details). Colors of βij : blue, European; purple, above and below the second dashed line represent individual traits and their
projected to 100% African ancestry; red, African ancestry. Black rows represent weighted average estimation. Vertical dotted lines: grey lines indicate ρ = 0.5 (left
individual UKB traits; the first standing height row uses an existing PGS16; the of the red dotted lines) and ρ = 1.5 (right of the red dotted lines); red lines indicate
dark green rows show combined estimates. Columns (left to right) estimate ρ for ρ = 1. ALP, alkaline phosphatase; FVC, forced vital capacity; HDL, high-density
‘all’ 8,003 African-ancestry individuals, ρ for individuals of 100% projected lipoprotein; IGF1, Insulin-like growth factor-1; LDL, low-density lipoprotein.

similar to wholly European-ancestry individuals (blue bars), and does Bonferroni correction. In future, larger sample sizes may better detect
not identify any significant effect sizes changes across the range of heterogeneity of effect sizes for individual traits.
genome-wide ancestry. African-ancestry segments (red) show much
lower predictive power, explaining nonportability because of local LD Discussion
and allele frequency differences, rather than gene–gene or gene–envi- We introduce new approaches to understand human ancestry and its
ronment interactions. connections to GWAS and PGS prediction. Decomposing UKB individu-
Across all 53 quantitative and molecular phenotypes (Fig. 5), ρ als’ ancestry into ACs at a subnational level, improves confounding
estimates (first column) were almost indistinguishable from 1, even for correction by capturing information missing from PCs, reducing likely
individuals with the highest African ancestry (second column), indicat- false positives and likely false negatives. ACs offer better control for
ing embedded European segments have similar predictive power to geographic effects than current methods, within the UKB at least. Many
nonadmixed Europeans, but with strongly reduced overall PGS power traits show similar P values using ACs versus other popular approaches,
(third column). The joint bootstrap yielded an overall ρ = 0.98 ± 0.07 but ACs uniquely avoid likely false positives in ‘regional traits’ defined
for these traits, extremely close to 1. Results were consistent using on groups of individuals6. Together with observations of likely false
varying P-value thresholds (P value of 0.05 and 0.0001) from the initial negatives in GWAS using PCs, and potential larger sample sizes of future
GWAS (Supplementary Fig. 15), and correction methods (ACs versus studies, incorporating ACs into GWAS may improve stratification cor-
PCs) in the initial GWAS (Supplementary Fig. 16). Results for standing rection by reducing overcorrecting linked to local genomic regions, and
height show little change using a previously published, alternative, avoids the issue of deciding the number of PCs in GWAS. The complex
PGS16 (Fig. 5). Two lung-function traits: FEV1 and forced vital capacity, patterns of ancestral mixing within UK regions or between, for exam-
and the correlated trait49,50 of sitting height, showed ρ < 1 at nominal ple, cities and rural areas, suggest particular care is needed for traits
significance (P < 0.05) (Fig. 5, Supplementary Fig. 15), with white blood showing similar regional variation such as educational attainment24. We
cell count, red blood cell count and albumin showing ρ > 1 (P < 0.05). note that the achievable granularity and interpretation of ACs depends
Only the white blood cell count remained significant (P < 0.05) with a on the size and diversity of the ancestry reference panel used, as well
GWAS P-value threshold of 0.0001, and no traits were significant after as the related number of identifiable predefined ‘ancestry groups’. A

Nature Genetics | Volume 57 | February 2025 | 379–389 386

Article https://doi.org/10.1038/s41588-024-02035-8

limitation of our approach is, therefore, that current ACs are some- Our results do not imply that gene–environment, or even gene–
what United-Kingdom-specific, but region-specific ACs for non-UK gene, interactions are absent across the traits we studied. Such interac-
GWAS could offer similar benefit. Until then, ACs can complement tions likely cause variation in effect size across individuals with African
PCs, though even the best ACs may not completely correct for subtle ancestry but must largely be shared with other ancestries to avoid
population stratification. differences in overall (mean) effect sizes across populations. Previous
To compare causal effect sizes across human populations, we work9 has shown effect sizes variation within UKB individuals of Brit-
consider the mean effect of a causal mutation on a trait12,51 and compare ish ancestry, based on age, gender and socioeconomic status. We also
this across groups. We note that the common and often convenient observe differences between males and females in mean effect sizes
scaling of genotypes by their population standard deviations also (Extended Data Fig. 10), and subtle differences among UK groups strati-
scales effect sizes differently across populations, so to avoid down- fied by ACs (Supplementary Fig. 6), likely reflecting gene–environment
ward biases in estimating causal effect correlations, we avoid such interactions. However, the strong lack of portability in African-ancestry
scaling14, and instead define identical impacts as a mutation causing individuals seems not driven by these interactions, apart from specific
the same average increase in phenotype in each group. This biologi- traits like FEV1. This result encourages joint fine-mapping56 efforts
cally natural approach provides an interpretable scale for practical and suggests that improving causal variant tagging is key to applying
applications. genetic findings across groups, simplifying the process by avoiding
ANCHOR leverages local ancestry inference to estimate the ratio the need of re-estimating effect sizes.
of underlying effect sizes between an admixed group and a reference The UKB resource contains diverse ancestries, but collects homog-
group similar to that used for the initial GWAS and PGS construction, by enous data for individuals within a single country, minimizing trait
assessing the predictive power of PGS among groups in terms of their definition differences and environmental effects to better isolate
relative ‘mean trait increase per unit PGS increase’. This ratio is 1 if effect underlying biological impacts. Our results likely rule out differential
sizes are identical, and estimates the correlation between true effect gene–gene interactions as a major driver of nonportability, even in
sizes in the two groups (Supplementary Note 2). Overall our results indi- other settings, because these would still operate strongly within the
cate clearly that—within the UK at least—using effect sizes from 100% United Kingdom. However, effect size differences might be stronger
European individuals yields near-identical performance in individuals between countries with greater environmental differences or differing
of African ancestry for various quantitative phenotypes. This suggests trait definitions. In future, ANCHOR might be applied to groups outside
PGS utility across populations, at least for causal mutations not private the United Kingdom, for example, African Americans14, to analyze
to one group. Although an underlying correlation between effect sizes various traits, or extended to analyze binary disease traits. As GWAS
ρ < 1 is possible (Supplementary Note 2), if coincidentally counterbal- sample sizes in admixed and other populations grow, methods includ-
anced by larger average effect sizes in African-ancestry individuals, ing ANCHOR will likely uncover variable effect sizes across countries
this seems unlikely across diverse quantitative traits a priori. If instead or cohorts, whereas other approaches10,52,54 enable comparisons of
ρ ≈ 1, most causal effect sizes are very similar between individuals of groups of similar ancestry.
African ancestry and European ancestry, across the range of quantita-
tive phenotypes we studied. Of interest for further study, a few traits Online content
such as FEV1 do suggest evidence of differences. Any methods, additional references, Nature Portfolio reporting sum-
Gene–gene and gene–environment interactions likely, therefore, maries, source data, extended data, supplementary information,
do not drive the lack of PGS portability in the UKB. Population structure acknowledgements, peer review information; details of author contri-
is also an unlikely cause, given consistent effect size estimation between butions and competing interests; and statements of data and code avail-
PGSs constructed by AC-corrected and PC-corrected GWAS. Instead, ability are available at https://doi.org/10.1038/s41588-024-02035-8.
local LD and SNP frequency differences appear to be the main factors.
African-ancestry segments show significantly reduced predictive References
power (about a threefold reduction) compared to European ancestry 1. Yengo, L. et al. A saturated map of common genetic variants
segments (Fig. 5). Our (ρ = 1) simulations confirm that PGS coeffi- associated with human height. Nature 610, 704–712 (2022).
cients match real data for European segments, but show only a modest 2. Yang, J., Lee, S. H., Goddard, M. E. & Visscher, P. M. GCTA: a tool
reduction (less than twofold for traits with >100 causal markers) for for genome-wide complex trait analysis. Am. J. Hum. Genet. 88,
African-ancestry segments. This suggests reduced tagging of causal 76–82 (2011).
variants by the PGS in African-ancestry regions potentially because of 3. Yang, J. et al. Common SNPs explain a large proportion of the
selection against trait-impacting variants52,53 causing larger frequency heritability for human height. Nat. Genet. 42, 565–569
differences for true causal variants between populations than the ran- (2010).
domly selected variants in our simulations, and the observed strong 4. Price, A. L. et al. Principal components analysis corrects for
drop-off between groups. Stronger stratification correction by ACs stratification in genome-wide association studies. Nat. Genet. 38,
could help investigate such selection in future. 904–909 (2006).
Our results differ from previous findings51,54,55 of often considerably 5. Novembre, J. & Stephens, M. Interpreting principal component
lower correlations between European-ancestry and African-ancestry analyses of spatial population genetic variation. Nat. Genet. 40,
samples using methods that model genetic variance components and 646–649 (2008).
local LD differences. For example, one study51 finds different effects for 6. Abdellaoui, A. et al. Genetic correlates of social stratification in
body mass index in UKB, another55 finds different effects for height, and Great Britain. Nat. Hum. Behav. 3, 1332–1342 (2019).
an estimated correlation of only ~50% across 26 traits in UKB. Because 7. Abdellaoui, A., Dolan, C. V., Verweij, K. J. & Nivard, M. G. Gene–
ANCHOR focuses on PGS prediction with minimal assumptions, it is environment correlations across geographic regions affect
challenging to attribute our near-parity estimates to confounding. genome-wide association studies. Nat. Genet. 54, 1345–1354
These differences might result from how genotypes are scaled and cen- (2022).
tered, as shown by the importance of appropriate centering in ANCHOR 8. Duncan, L. et al. Analysis of polygenic risk score usage and
(Extended Data Fig. 8 and Supplementary Figs. 8 and 9). Alternatively, performance in diverse human populations. Nat. Commun. 10,
if causal variants have unique evolutionary histories because of natural 3328 (2019).
selection, this might impact some methods more than others. Further 9. Mostafavi, H. et al. Variable prediction accuracy of polygenic
study is needed to understand these discrepancies. scores within an ancestry group. eLife 9, e48376 (2020).

Nature Genetics | Volume 57 | February 2025 | 379–389 387

Article https://doi.org/10.1038/s41588-024-02035-8

10. Brown, B. C., Ye, C. J., Price, A. L. & Zaitlen, N. Transethnic 33. Lawson, D. J., Hellenthal, G., Myers, S. & Falush, D. Inference of
genetic-correlation estimates from summary statistics. Am. J. population structure using dense haplotype data. PLoS Genet. 8,
Hum. Genet. 99, 76–88 (2016). e1002453 (2012).
11. Bitarello, B. D. & Mathieson, I. Polygenic scores for height in 34. Loh, P.-R. et al. Efficient Bayesian mixed-model analysis increases
admixed populations. G3 (Bethesda) 10, 4027–4036 association power in large cohorts. Nat. Genet. 47, 284–290
(2020). (2015).
12. Patel, R. A. et al. Genetic interactions drive heterogeneity in 35. Krausova, A. & Vargas-Silva, C. Wales: census profile.
causal variant effect sizes for gene expression and complex traits. Migration Observatory at the University of Oxford Briefing
Am. J. Hum. Genet. 109, 1286–1297 (2022). https://migrationobservatory.ox.ac.uk/resources/briefings/
13. Marnetto, D. et al. Ancestry deconvolution and partial polygenic wales-census-profile/ (2013).
score can improve susceptibility predictions in recently admixed 36. Privé, F., Luu, K., Blum, M. G., McGrath, J. J. & Vilhjálmsson, B.
individuals. Nat. Commun. 11, 1628 (2020). J. Efficient toolkit implementing best practices for principal
14. Hou, H. et al. Causal effects on complex traits are similar for component analysis of population genetic data. Bioinformatics
common variants across segments of different continental 36, 4449–4457 (2020).
ancestries within admixed individuals. Nat. Genet. 55, 549–558 37. Bulik-Sullivan, B. K. et al. LD Score regression distinguishes
(2023). confounding from polygenicity in genome-wide association
15. Weissbrod, O. et al. Leveraging fine-mapping and multipopulation studies. Nat. Genet. 47, 291–295 (2015).
training data to improve cross-population polygenic risk scores. 38. Loh, P.-R., Kichaev, G., Gazal, S., Schoech, A. P. & Price, A. L.
Nat. Genet. 54, 450–458 (2022). Mixed-model association for biobank-scale datasets. Nat. Genet.
16. Privé, F. et al. Portability of 245 polygenic scores when derived 50, 906–908 (2018).
from the UK Biobank and applied to 9 ancestry groups from the 39. Cook, J. P., Mahajan, A. & Morris, A. P. Fine-scale population
same cohort. Am. J. Hum. Genet. 109, 12–23 (2022). structure in the UK Biobank: implications for genome-wide
17. Márquez‐Luna, C., Loh, P. R., South Asian Type 2 Diabetes (SAT2D) association studies. Hum. Mol. Genet. 29, 2803–2811 (2020).
Consortium, SIGMA Type 2 Diabetes Consortium & Price, A. 40. Kawasaki, T. & Kawai, T. Toll-like receptor signaling pathways.
L. Multiethnic polygenic risk scores improve risk prediction in Front. Immunol. 5, 461 (2014).
diverse populations. Genet. Epidemiol. 41, 811–823 (2017). 41. Mikacenic, C., Reiner, A. P., Holden, T. D., Nickerson, D. A. & Wurfel,
18. Martin, A. R. et al. Current clinical use of polygenic scores will M. M. Variation in the TLR10/TLR1/TLR6 locus is the major genetic
risk exacerbating health disparities. Nat. Genet. 51, 584–591 determinant of interindividual difference in TLR1/2-mediated
(2019). responses. Genes Immun. 14, 52–57 (2013).
19. Wang, Y. et al. Theoretical and empirical quantification of the 42. Waage, J. et al. Genome-wide association and HLA fine-mapping
accuracy of polygenic scores in ancestry divergent populations. studies identify risk loci and genetic pathways underlying allergic
Nat. Commun. 11, 3865 (2020). rhinitis. Nat. Genet. 50, 1072–1080 (2018).
20. Carlson, C. S. et al. Generalization and dilution of association 43. Churchhouse, C. et al. Rapid GWAS of thousands of
results from European GWAS in populations of non-European phenotypes for 337,000 samples in the UK Biobank.
ancestry: the PAGE study. PLoS Biol. 11, e1001661 (2013). Neale Lab Blog www.nealelab.is/blog/2017/7/19/
21. Atkinson, E. G. Estimation of cross-ancestry genetic correlations rapid-gwas-of-thousands-of-phenotypes-for-
within ancestry tracts of admixed samples. Nat. Genet. 55, 337000-samples-in-the-uk-biobank (2019).
527–529 (2023). 44. Bycroft, C. et al. Patterns of genetic differentiation and the
22. Hellenthal, G. et al. A genetic atlas of human admixture history. footprints of historical migrations in the Iberian Peninsula. Nat.
Science 343, 747–751 (2014). Commun. 10, 551 (2019).
23. Leslie, S. et al. The fine-scale genetic structure of the British 45. Martin, S. et al. Genetic evidence for different adiposity
population. Nature 519, 309–314 (2015). phenotypes and their opposing influences on ectopic fat and risk
24. Haworth, S. et al. Apparent latent structure within the UK Biobank of cardiometabolic disease. Diabetes 70, 1843–1856 (2021).
sample has implications for epidemiological analysis. Nat. 46. Christakoudi, S., Evangelou, E., Riboli, E. & Tsilidis, K. K. GWAS
Commun. 10, 333 (2019). of allometric body-shape indices in UK Biobank identifies loci
25. Bycroft, C. et al. The UK Biobank resource with deep phenotyping suggesting associations with morphogenesis, organogenesis,
and genomic data. Nature 562, 203–209 (2018). adrenal cell renewal and cancer. Sci. Rep. 11, 10688 (2021).
26. Patterson, N., Price, A. L. & Reich, D. Population structure and 47. Graff, M. et al. Genome-wide physical activity interactions in
eigenanalysis. PLoS Genet. 2, e190 (2006). adiposity―A meta-analysis of 200,452 adults. PLoS Genet. 13,
27. Cai, M. et al. A unified framework for cross-population trait e1006528 (2017).
prediction by leveraging the genetic correlation of polygenic 48. Price, A. L. et al. Sensitive detection of chromosomal segments
traits. Am. J. Hum. Genet. 108, 632–655 (2021). of distinct ancestry in admixed populations. PLoS Genet. 5,
28. Cann, H. M. et al. A human genome diversity cell line panel. e1000519 (2009).
Science 296, 261–262 (2002). 49. Torres, L. A., Martinez, F. E. & Manço, J. C. Correlation between
29. Schlebusch, C. M. et al. Genomic variation in seven Khoe-San standing height, sitting height, and arm span as an index of
groups reveals adaptation and complex African history. Science pulmonary function in 6–10‐year‐old children. Pediatr. Pulmonol.
338, 374–379 (2012). 36, 202–208 (2003).
30. Pagani, L. et al. Ethiopian genetic diversity reveals linguistic 50. Lum, S. et al. Lung function in children in relation to ethnicity, phy­
stratification and complex influences on the Ethiopian gene pool. sique and socioeconomic factors. Eur. Respir. J. 46, 1662–1671 (2015).
Am. J. Hum. Genet. 91, 83–96 (2012). 51. Momin, M. M. et al. A method for an unbiased estimate of
31. Novembre, J. et al. Genes mirror geography within Europe. Nature cross-ancestry genetic correlation using individual-level data.
456, 98–101 (2008). Nat. Commun. 14, 722 (2023).
32. Delaneau, O., Zagury, J.-F. & Marchini, J. Improved 52. Shi, H. et al. Population-specific causal disease effect sizes in
whole-chromosome phasing for disease and population genetic functionally important regions impacted by selection. Nat.
studies. Nat. Methods 10, 5–6 (2012). Commun. 12, 1098 (2021).

Nature Genetics | Volume 57 | February 2025 | 379–389 388

Article https://doi.org/10.1038/s41588-024-02035-8

53. Yair, S. & Coop, G. Population differentiation of polygenic score Open Access This article is licensed under a Creative Commons
predictions under stabilizing selection. Philos. Trans. R. Soc. Lond. Attribution 4.0 International License, which permits use, sharing,
B Biol. Sci. 377, 20200416 (2022). adaptation, distribution and reproduction in any medium or format,
54. Galinsky, K. J. et al. Estimating cross‐population genetic as long as you give appropriate credit to the original author(s) and the
correlations of causal effect sizes. Genet. Epidemiol. 43, 180–188 source, provide a link to the Creative Commons licence, and indicate
(2019). if changes were made. The images or other third party material in this
55. Guo, J. et al. Quantifying genetic heterogeneity between article are included in the article’s Creative Commons licence, unless
continental populations for human height and body mass index. indicated otherwise in a credit line to the material. If material is not
Sci. Rep. 11, 5240 (2021). included in the article’s Creative Commons licence and your intended
56. LaPierre, N. et al. Identifying causal variants by fine mapping use is not permitted by statutory regulation or exceeds the permitted
across multiple studies. PLoS Genet. 17, e1009733 (2021). use, you will need to obtain permission directly from the copyright
holder. To view a copy of this licence, visit http://creativecommons.
Publisher’s note Springer Nature remains neutral with regard org/licenses/by/4.0/.
to jurisdictional claims in published maps and institutional
affiliations. © The Author(s) 2025

Nature Genetics | Volume 57 | February 2025 | 379–389 389

Article https://doi.org/10.1038/s41588-024-02035-8

Methods each sample’s donor-vector as a mixture of the surrogate vectors in the

Our research complies with all relevant ethical regulations. Collection surrogate panel using NNLS as described in ref. 22, merging or remov-
of the UKB data was approved by the Research Ethics Committee of ing any groups not >50% recovered.
the UKB and this research has been conducted using the UK Biobank
Resource under application number 27960. UK Biobank data included in the analysis
The UKB study includes more than 500,000 UK residents. We analyzed
Statistics and reproducibility genotypic and phenotypic data under application 27960. In total,
Sample size is clearly disclosed in the paper for each different dataset or 487,409 UKB participants with available autosomal haplotype and
sub-dataset. The sample size of ancestry-specific analysis in the section genotype imputation data (field IDs 22438 and 22828) were used in
‘Polygenic score calculations’ was determined by the corresponding our ancestry inference. Quality control, phasing and imputation for
genetic ancestry coefficients inferred using our method. Trio individu- the UKB genetic data have been described previously (http://biobank.
als with African-background ancestry were detected using the open ctsu.ox.ac.uk/). Demographic and phenotypic data are listed in Sup-
source software king (v.2.2). No data were excluded unless the partici- plementary Table 8. For PGS analysis, we selected 53 quantitative traits
pants withdrew their participation from the UKB. No randomization on which to run GWAS using the following criteria: trait measured on
was used; because there are no experimental groups it is not relevant at least 400,000 individuals; LDSC-estimated trait heritability at least
to our study. No blinding is used because it is not relevant to this study; 5%; and the trait must be noncategorical.
however, group allocation is objective using inferred genetic ancestry.
Running the ancestry pipeline on UK Biobank
Ancestry pipeline The ancestry pipeline accepts genotype data in various formats as
To construct an ancestry ‘painting’ reference panel, we merged data input (Supplementary Note 1). For UKB data, we used available phased
from Supplementary Table 7, resulting in 9,129 samples and 2,011,414 haplotypes. We performed initial imputation with IMPUTE4 in batches
distinct mutations. After relatedness pruning in plink1.9 (ref. 57) using of 1,000 UKB individuals, then ran the remaining jobs individually: first,
the IBS/IBD computation with the ‘--genome’ option and setting a PI_ remove close relatives and one random individual per donor group
HAT threshold of 0.25 to exclude related samples exceeding this value, from the donor panel; second, paint the target using parameters Ne
we retained 7,775 individuals. Phasing (using SHAPEIT2)32 and impu- and m estimated from the panel to obtain a donor-vector; and third,
tation (using IMPUTE4)25 were performed using the UK10K58 + 1000 infer global ancestry using NNLS.
Genomes data as a reference panel, retaining 851,948 sites with a mean
IMPUTE4 information score above 0.9, across all the different geno- Assign UK counties with predefined UK regions
typing platforms simultaneously, as well imputed sites, in the ‘draft Within the United Kingdom and Ireland, 23 distinct groupings were
painting panel’. identified. Each UK county was assigned to one group, to refine
The remainder of the panel construction process, detailed in geographic boundaries. We downloaded county-level UK map data
Supplementary Note 1, involved a nested sample hierarchy, chromo- (https://gadm.org), mapped the birthplaces of 426,879 UK-born
some painting using ChromoPainterv2 (ref. 33), quality control for UKB individuals to a county using the R package ‘sp’, and filtered
relatedness, ‘surrogate’ and ‘donor’ group formation based on sample individuals with >50% ancestry from one group. We then assigned
labels and admixture estimation by NNLS22, and additional SNP qual- a county to the group with the most remaining individuals born in
ity control, resulting in 677,173 SNPs in the ‘final painting panel’. We the county. ‘Irish’ ancestry was assigned to the Republic of Ireland.
also painted UKB individuals who were born in world regions that we Within Cornwall, because both ‘Cornwall’ and ‘Cornwall Tip’ localized
believed our reference individuals sampled from poorly, after exclud- to this county, we used the R package ‘raster’ to define ancestry on
ing samples with, for example, mainly BI ancestry, and used their profile finer-scale pixels, and assigned locations whose mean ‘Cornwall Tip’
as a surrogate group reference vector. ancestry was at least 0.2 greater than their mean ‘Cornwall’ ancestry
Annotating population labels into genetic groupings is an art, not to the ‘Cornwall Tip’ group.
a science. Our choices led to 206 surrogate groups that were geneti-
cally distinct, which we sum into the 127 interpretable ancestry labels Visualization of ancestry based on UKB birthplace
reported in the ‘Results’ section. For replicability we include the SNP We used Gaussian kernel smoothing to generate spatial smoothed
lists, and the final donor and surrogate group annotations (Supple- plots showing average ancestry. For each pixel (p) in the rasterized UK
mentary Note 1). map with birthplace coordinates p = (xp,yp), we calculated the mean of
a quantity of interest O (for example, AC) for each of the N = 426,879
Painting panel processing UK-born or Irish-born UKB samples, smoothed by the Gaussian kernel:
To obtain calibrated local ancestry estimates, individuals in the 2 2
N
panel must be exchangeable with those we wish to compare. We first e−q((xp −xi ) +(yp −yi ) )
Ō p = ∑ 2 2
Oi (1)
construct a SNP and sample list using the process above, excluding i=1
N
∑i=1 e−q((xp −xi ) +(yp −yi ) )

UK10K or 1000 Genomes samples; second, phase each sample inde-

pendently against the UK10K + 1000 Genomes dataset using SHAPEIT2;
third, impute each sample independently against the UK10K + 1000 where Oi is the ancestral object and (xi,yi) is the birthplace coordinate
Genomes dataset using IMPUTE4; and fourth, infer best-fit param- for individual i. Ō p is the mean of O at pixel p = (xp,yp). We used adaptive
eters Ne (which controls recombination rate) and m (which controls bandwidth smoothing for q (Supplementary Note 1).
mutation rate) using ChromoPainterv2 in expectation-maximization To quantify ancestry mixing across the United Kingdom, we cal-
mode for 10% of samples randomly chosen for each chromosome, then culated ancestry entropy for each individual:
average these for final parameters. First, to perform a ‘leave-one-out’ nr
aj aj
procedure to create reference groups with one fewer sample, we E = −∑ log (2)
n n
paint with ChromoPainterv2 using inferred parameters and repeat j=1 ∑k=1
r
ak ∑k=1
r
ak
the ‘leave-one-out’ procedure above. Second, given each sample a
‘donor-vector’ indicating genome shared (in centi-Morgans) with each where nr = 23 is the number of predefined BI regions, and aj(1 ≤ j ≤ nr)
donor group, we create a surrogate panel by averaging donor-vector are the ACs for that individual. Adaptive bandwidth Gaussian kernel
within each surrogate group. Third, we estimate admixture by treating smoothing was used to plot the UK-wide entropy profile.

Nature Genetics
Article https://doi.org/10.1038/s41588-024-02035-8

Genome-wide association study For ‘Waist circumference’, we identified 627 independent

Forty PCs (field ID 22009) were downloaded from the UK Biobank. The genome-wide significant hits for either AC- or PC-corrected GWAS.
P
first 20 PCs were used in GWAS analysis. We additionally calculated Only five hits were AC-specific with a − log10 AC of >1, where PAC and PPC
PPC
the first 200 UKB PCs using the R packages ‘pcapred.largedata’ and are P values for AC- and PC-based GWAS for ‘Waist circumference’. We
‘pcapred’ (https://github.com/danjlawson/pcapred.largedata, https:// queried four of those five variants (Fig. 3e), using ‘mrcieu/ieugwasr’,
github.com/danjlawson/pcapred), which closely matches UKB PCs while the remaining variant (15:84311431:TA:T) lacked an RS catalogue
(>99% correlation for the first 40 PCs) (Supplementary Fig. 17). identifier, so was manually queried using opentargets 60 (http://genet-
Some 127 ACs from the ancestry pipeline were used to regress ics.opentargets.org/). It was strongly associated with many different
each of 140 UKB PCs (chosen to exceed the number of ACs) on all the ‘Anthropometric’ related traits such as height, trunk fat mass and
487,409 UKB samples and predict each PC. Similarly, 140 UKB PCs were body fat percentage, and so was labeled as ‘Anthropometric’. The
used to predict the 127 ACs on the same UKB samples. R2 between the other 622 variants possessed shared signals and were therefore not
true and predicted PC or AC was then used to evaluate the prediction. queried.
In our GWAS study, we analyzed 104 continuous UKB phenotypes In assigning categories to SNPs overlapping multiple association
with a sample size >10,000 (Supplementary Table 8). In total, 343,047 types, we ranked as follows to prioritize relevance to the traits studied
unrelated white British individuals25 were included in our GWAS. Asso- in a particular GWAS in Fig. 3c–e. For ‘Birthplace (North/South)’, we
ciation testing was performed for UKB imputed SNPs G (minor allele assigned the category ‘Allergy’ to the queried SNP if it is associated
frequency (MAF) >0.001 and information scores >0.3). Covariates with allergic phenotypes such as ‘Hay fever’; for ‘Employment Score
in each association test were ‘genotype measurement batch’, ‘age at England’, we assigned category ‘Educational attainment’ to the queried
recruitment’ and ‘sex’ (field IDs 22000, 21022, and 31), as well as non- SNP if it is associated with ‘Educational attainment’ or related pheno-
linear terms ‘age2’, ‘age × sex’ and ‘age2 × sex’. The full model using ACs types, including ‘Age completed full-time education’, and then category
to correct stratification is: ‘Cognitive function’ if associated with such traits as, for example, ‘Mood
swing’; for ‘Waist circumference’, we assigned category ‘Anthropomet-
Phenotype ∼ G + age + sex + batch + age2 ric’ to queried SNPs associated with, for example, ‘ Leg fat percentage’.
(3)
+age × sex + age2 × sex + 127ACs + error Remaining uncategorized SNPs were categorized by their top-ranked
associated trait (Supplementary Table 3a–e).
The model instead using 20 PCs (and similarly when using 100 PCs) is:
Heritability estimation and GWAS inflation assessment
Phenotype ∼ G + age + sex + batch + age2 We ran LD-score regression37 on both AC-corrected and PC-corrected
(4)
2
+age × sex + age × sex + 20PCs + error GWAS summary statistics for 99 UKB traits, using precalculated LD
scores from the 1000 Genomes phase 3 Utah residents with Northern
Association testing was performed by ‘BGENIE’ v.1.2 (ref. 25) using and Western European ancestry (CEU) panel provided alongside the
unnormalized phenotypes, to keep estimated effect sizes remain in LDSC software. From LD-score regression output of the intercept and
units of the original phenotypes. χ2 values, we bounded these at below 1 to ensure non-negativity, and
We also performed GWAS using BOLT-LMM34 with 20 PCs, to test calculated a trait-wise LDSC attenuation ratio, defined as follows38:
performance of a linear mixed model for the place of birth, north coor-
dinates (field ID 129) phenotype. We performed quality control and LD Intercept − 1
Attenuation ratio = (5)
pruning on the UKB chip genotype SNPs using plink1.9 (ref. 57) (www. χ2 − 1
cog-genomics.org/plink/1.9/; Supplementary Note 1) to generate 142,182
independent SNPs (r2 < 0.1) for null model building in BOLT-LMM. The Estimation of allele frequencies for UKB SNPs
participants and imputed SNPs for association testing remained as for Highly differentiated alleles across different regions reflect strong
the other GWAS. The 20 PCs included in BOLT-LMM as covariates were local genetic drift and can reveal natural selection. We derived an
calculated by FlashPCA2 (ref. 59) from the same genetic relationship expectation-maximization algorithm to estimate maximum-likelihood
matrix constructed by the 142,182 independent SNPs and 343,047 UKB regional allele frequencies based on our individual ACs (Supplemen-
white British individuals. Other covariates used in BOLT-LMM were the tary Note 1). We applied this to 12,977,776 imputed UKB SNPs with
same as those used in the linear regression based GWAS. minimum MAF at least 0.01% and imputation information score above
To identify independent genome-wide significant signals, we used 0.9 in 339,304 white British samples with more than 95% UK + Ireland
a threshold P < 5 × 10−8 and for each SNP more significant than this ancestry. Using ‘qctool2’ (https://www.well.ox.ac.uk/~gav/qctool_v2/
threshold for either AC or PC, we used LD pruning with a threshold of index.html), ‘bgenix’61 and plink2.0 (ref. 57)), we conducted genotype
r2 < 0.01 and window size of 100 kb to keep only the most significant data preprocessing, quality control and estimated their allele frequen-
SNP among groups of variants in LD. cies across the 23 British–Irish regions.
We queried AC-significant or PC-significant SNPs in a GWAS data-
base using the R package ‘mrcieu/ieugwasr’ (https://mrcieu.github. Polygenic score calculations
io/ieugwasr/). For most traits we applied a genome-wide threshold of For each of 53 traits, we constructed two PGSs using either ACs or PCs
5 × 10−8 for query output P value, relaxed to 1 × 10−6 if no SNP met the for population stratification correction in a GWAS of 343,047 unre-
initial threshold, and returning ‘No GWAS hits’ if the relaxed threshold lated white British individuals. PGS construction used the Hapmap3
was also not met. For traits shown in Fig. 3c–e, queried results contain- SNPs62 with a minimum 1% MAF. For each phenotype, we performed
ing the same or directly related traits were excluded from categoriza- LD-clumping by ‘plink1.9’57 using 1000 Genomes phase 3 CEU63 as the
tion. Query results for input SNPs were ranked by ascending GWAS P reference panel with following parameters: significant threshold for
value across different traits such that the most significant trait was index SNPs is 0.05, secondary significance threshold for clumped SNPs
top-ranked (Supplementary Table 3a–e). SNPs without a standard is 1, LD threshold for clumping is 0.1 and physical distance threshold
RS catalogue identifier were categorized as ‘Not queried’. Queried for clumping is 500 kb, with respect to the plink command line as fol-
traits were assigned into the following categories: ‘Anthropometric’, lows: ‘--clump-p1 0.05 --clump-p2 1 --clump-r1 0.1 --clump-kb 500’. We
‘Cognitive function’, ’Lipids’, ‘Education attainment’, ‘Autoimmune’, also tested a stricter P value threshold of 0.0001 by setting ‘—clump-p1
‘Population density’, ‘Allergy’, ‘Socioeconomics status’, ‘Blood pressure’, 0.0001’ in the above plink command. The PGS was evaluated on the
‘Blood counts’ and ‘Other medical condition’ (Supplementary Table 9). remaining 144,362 UKB individuals.

Nature Genetics
Article https://doi.org/10.1038/s41588-024-02035-8

PGS was calculated as a linear sum over SNPs and genotypes: is each of the four possibilities, ordering the two chromosomes arbi-
n
trarily (for example, ‘EE’ refers to both chromosomes possessing Euro-
PGSAC = ∑ γˆ
AC
ji
Gj (6) pean ancestry). Given genotype Gij at site j for individual i, we estimate
j=1 allele frequencies f Ej , f Aj for European and African background, respec-
tively, by fitting the model
n
PGSPC = ∑ γˆ
PC EE EA AA EA
ji
Gj (7) Gij = f Ej (2Pij +Pij + PAE
ij
) + fAj (2Pij +Pij + PAE
ij
) + εi (9)
j=1

where εi is the mean-zero noise, by least squares. In practice we fit the

where for SNP j, Gj is the genotype for the target individual, γˆ AC
ji
is the equivalent model, noting PEA
ij
= PAE
ij
and that the ancestry probabilities
estimated effect size from the AC-corrected GWAS and γˆ PC
ji
is the effect sum to 1:
size from the PC-corrected GWAS. Mean-centering genotypes within
each population would yield equivalent results, so it was not done here Gij = I j + S j × (2PEE
ij
+ 2PEA
ij
) + εi (10)
(Supplementary Note 2).
To evaluate PGSAC performance across different ancestries, we where after model fitting we obtain estimates f Ej = S j + I j /2 and
identified those individuals among the 144,362 UKB ‘test’ individu- f Aj = I j /2. Given the large sample size, these estimates closely match
als with at least 50% inferred ancestry from each of seven separate true SNP frequencies for the specific admixing groups (they also cor-
AC-labeled groups: ‘South Central England’ (1,491 individuals), ‘North- relate strongly with the frequencies of the same variants in relevant
umberland’ (5,355), ‘Republic of Ireland’ (13,265), ‘Poland’ (1,503), ‘India’ 1000 Genomes cohorts) (Supplementary Fig. 18).
(4,438), ‘China’ (1,266) and ‘West Africa’ (7,108). HAPMIX also estimates expected allele counts GEij , GAij at this site
For each group and phenotype, we evaluated PGSAC for individual for European and African ancestry backgrounds (obtained via summa-
i, label as PGSi, and fit the model: tion; Supplementary Note 1), which are transformed to their
mean-centered version (Supplementary Note 2), conditional on local
Yi = I + βPGSi + covariates + εi (8) ancestry probabilities:
E
where Yi is the phenotype for individual i, εi is the error, and the model Ḡ ij = GEij − f Ej (2PEE
ij
+ 2PEA
ij
) (11)
was fit via standard linear regression, with 1,000 bootstrapped sample
replicates for evaluating uncertainty. As previously, the covariates A
A
included are batch, age, sex, age2, age × sex and age2 × sex, as well as Ḡ ij = Gij − f Aj (2PAA
ij
+ 2PEA
ij
) (12)
the 127 estimated ACs. The parameter β represents an estimator of
the increase in the mean phenotype per unit increase in the PGS. We
compared β values among groups to quantify PGS applicability in We apply genomic masking to remove uncertain or short segments
distinct groups, and used generalized values in our ANCHOR analysis. (<5 megabases) inferred by HAPMIX. However, simulations show con-
For each group, we defined the ‘residual r2’, ΔR2, a scale-independent sistent results regardless of masking, provided mean-centering is
measure of PGS performance, as 1 minus the ratio of the phenotypic correctly conducted.
variance remaining after fitting the above model to the (larger) variance The overall PGS can be decomposed into the European PGS (EPGS)
when β = 0. This represents the fraction of variance explained by the and African PGS (APGS) for African-ancestry individuals i = 1, 2, … n
PGS after accounting for confounding and covariates. as follows:
We also estimated an overall value βObs.Eu for 19,596 individuals of
J
BI ancestry by fitting the same model. The BI individuals were selected E
EPGSi = ∑ γ̂j Ḡ ij (13)
from three BI groups—South Central England, Northumberland and j=1

Republic of Ireland—filtering out individuals with the sum of these

three ancestries less than 0.9. J
A
APGSi = ∑ γ̂j Ḡ ij (14)
Ancestry-aware PGS and correlation in effect sizes j=1

To understand the behavior of the PGS in African-ancestry individu-

als, we identified 8,003 UKB European–African admixed individuals where γ̂j is the estimated per-copy effect size of SNP j on the phenotype.
whose sum of 27 UK or European ancestries and 4 sub-Saharan African To investigate local (for example, LD) and nonlocal factors (that is,
ancestries is larger than 90% and sum of 4 sub-Saharan ancestries larger interactions) attenuating the PGS performance in African-ancestry
than 10%. We further binned these individuals into five ancestry bins individuals, we fitted the following model to real and simulated data
based on their inferred African ancestry: [0.1, 0.35], 373 individuals, (Supplementary Note 2):
mean ancestry 0.207; [0.35, 0.55], 560 individuals, mean ancestry 0.46;
[0.55, 0.78], 475 individuals, mean ancestry 0.707; [0.78, 0.95], 2,556 Yi = I + (βEu EPGSi + βAf APGSi ) × (1 + ωθi ) + covariates + εi (15)
individuals, mean ancestry 0.89; and [0.95, 1], 4,039 individuals, mean
ancestry 0.986. Individuals in the bin [0.95, 1] with near 100% African where for individuals i = 1, 2, … n, Yi is the phenotype, εi is the zero-mean
ancestry were only given an ‘APGS’ PGS in our analysis (see below), the noise and the parameters to be estimated are the intercept I, and βEu, βAf,
results of which are shown in Fig. 4b,c and Extended Data Fig. 9 and ω, EPGSi and PPGSi are the centralized EPGS and APGS after regressing
Supplementary Figs. 9 and 12. out the covariates, and θi is the mean genome-wide European ancestry
In the ANCHOR approach, we generate and analyze ancestry- proportion for individual i. Covariates are as described in the section
specific PGS for a trait. We applied HAPMIX48 to estimate local ancestry ‘Polygenic score calculations’.
genome-wide for each individual, using 1000 Genomes phase 3 (ref. We fit this model across various individual subsets, combinations
63) CEU and Yoruba in Ibadan, Nigeria groups, respectively, as Euro- of phenotypes and parameter constraints (described below), obtaining
pean and African ancestry reference panels, with default parameters. parameter estimates via least squares (Supplementary Note 2), and
For individuals labeled i = 1, 2, … n, the output of HAPMIX at each locus uncertainty estimates by 1,000 bootstrap resamples of individuals and
j = 1, 2, … J provides probabilities PEE
ij
, PEA
ij
, PAE
ij
, PAA
ij
that the local ancestry model refitting. The model is linear so trivial to fit unless allowing ω ≠ 0;

Nature Genetics
Article https://doi.org/10.1038/s41588-024-02035-8

in this case, conditional on the ratio βEu /(βEu + βAf ) the model is again distributed along the genome. Effects sizes of the causal variants are
linear, so we use a grid search (1,000 values) over this ratio from 0 to 1, drawn independently either from a N(0,1) distribution, or following a
fitting the linear model for each possible value and then minimizing LDAK model64 where the effect size of variant j is a draw from a standard
the achieved sum of squares. normal multiplied by a factor [2pj (1 − pj)−0.25], where pj is that variant’s
Parameters βEu and βAf measure the increase in the phenotype per frequency, resulting in larger effects for rarer variants. This generates
unit increase in the respective PGS, indicating the scores’ predictive 3 × 2 × 2 = 12 scenarios; we simulated two different heritabilities 0.3,
power. Local factors mean that we expect βEu > βAf (Supplementary 0.6 resulting in 24 simulations in total.
Note 2; Extended Data Fig. 9 shows this via simulation). If no addi- Finally, to generate simulated phenotypes Yi for individual i = 1,
tional nonlocal factors contribute to nonportability, then βEu and βAf 2, … N, we apply the following additive model which leverages the
should remain constant across ancestry bins, that is varying θi, with βEu actual UKB genotypes and so matches properties of these data, includ-
also shared between African-ancestry individuals and 100% European ing in particular population stratification:
ancestry individuals (Supplementary Note 2). With nonlocal factors
J
acting, this no longer holds. To investigate this setting, we allow the Yi = ∑ γj gij + εi (16)
predictive power of the two scores varies with genome-wide European j=1

or African ancestry that is θi, captured by ω. Because local factors

still operate, the model covaries the predictive power of the βEu, βAf Here ɣj and gij are the effect size and genotype of individual i at site j = 1,
parameters together with ω; we lack power with current sample sizes 2, … J. Noise terms εi are drawn from a normal distribution with mean
to fit separate effects, and so we simply use ω to capture linear effects14 0 and variance σ2e . To obtain heritability h2 of 0.3 or 0.6, we set the vari-
1−h2
(Supplementary Note 2). ance of εi in equation (16), σ2e is equal to 2 σ2g , where σ2g is the observed
h
We fit this model to analyze both real and simulated datasets in an variance of the first term.
identical manner. When jointly analyzing phenotypes and averaging We also repeated these simulations, now allowing effect sizes to
results, we first binned the African-ancestry individuals by their differ in the 8,003 African-ancestry individuals (Supplementary Note
genome-wide ancestry θi (bin boundaries shown on Fig. 4c). Within 2). For a correlation in effect sizes of ρ ∈ {0.1, 0.3, 0.5, 0.7, 0.9}, we simu-
each bin ancestry varies little, so we set ω = 0 and fit the model inde- lated from the following model:
pendently for each bin for each phenotype. This provides estimates of
βEu, βAf for each phenotype-bin combination, averaged within bins for γEj 0 σ2j ρσ2j
( ) ∼ N( ( ), ( ) (17)
Fig. 4c, allowing comparisons across bins, and with the effect size βˆ Obs.Eu γAj 0 ρσ2j σ2j
obtained for individuals of mainly European ancestry. By holding ω = 0
fixed but analyzing all individuals jointly for a phenotype produces
estimates β ˆAll
and β ˆAll
summarizing the predictive power of the PGS This can be done efficiently without modifying effect sizes for non
Eu Af
across the full African-ancestry sample set, shown in, for example, African-ancestry individuals, by noting that conditional on the Euro-
Figs. 4d and 5. The ratio of the means of the estimates over phenotypes, pean effect size γEj at site j = 1, 2, … J, γAj has the following distribution:
ˆ
β All ˆ
/βObs.Eu , estimates the correlation ρ in effect sizes between
Eu
Africa-ancestry and European-ancestry groups, averaged over pheno- γAj |γEj ∼ N (ργEj , (1 − ρ2 ) σj 2 ) (18)
types (Supplementary Note 2); estimates of this are shown in, for exam-
ple, Figs. 4b and 5 for simulated and real data. For an individual or equivalently,
phenotype, if the 95% bootstrapped confidence interval of the ratio
ˆ
β All ˆ
/βObs.Eu contains 1, we accept the null hypothesis of shared effect γAj = ργEj + σj √1 − ρ2 Z (19)
Eu
sizes. The ratio β ˆ All ˆ
/βObs.Eu captures local (for example, LD) effects on
Af
prediction, and so as expected is <1 for all simulated and real traits. where Z is a standard normal random variable. We use this to generate
Finally, we fit the full model allowing ω to vary. In individuals with 100% γAj for each setting and calculate PGS for African-ancestry individuals.
African [respectively European] ancestry, this fits the unit increase in We adjust the phenotypic variance for African-ancestry individuals so
the phenotype per unit increase in EPGS as βˆ Af
Eu
= βˆ ̂ ; [ βˆ
Eu (1 + ω)
Eu
Eu
= βˆ
Eu], as to maintain their heritability at the same value as the UKB samples
ˆAf ˆ
Af
and analogous coefficients βAf , βAf correspond to the APGS. These as a whole. Based on the above scenarios, a simulated phenotype named
coefficients are shown in, for example, Figs. 4d and 5. Ratios of these as, for example, ‘#causal:1K(uniform)S:0 h2:0.6’ means the underlying
quantities to βˆ Obs.Eu, and their bootstrapped CIs, are interpreted exactly simulation scenario for this phenotype is: a phenotype with heritability
as those for ‘All’ individuals but now projected to estimate properties 0.6, in total 1,000 causal variants, uniformly distributed along the
of individuals—for example, ratios of true effect sizes—whose ancestry genome, and S:0 (versus S:0.5) means there is no effect size scaling
is 100% African or even 100% European. These quantities are defined used.
consistently across analyses and subsets of individuals.
Reporting summary
Ancestry-aware simulation Further information on research design is available in the Nature
To explore GWAS portability and evaluate performance of ANCHOR, we Portfolio Reporting Summary linked to this article.
simulated phenotypes across the entire UKB cohort, and analyzed as
for the real data. The UKB imputed genotype data includes 12,690,793 Data availability
variants after applying the following filters: MAF ≥ 0.001, minor allele UK and world map data can be accessed through https://gadm.org. UK
count ≥25, genotype missingness ≤5%, Hardy–Weinberg equilibrium Biobank data can be downloaded by approved researchers through
P ≥ 1 × 10−10 and imputation INFO score ≥0.8. From these, we chose a https://www.ukbiobank.ac.uk. Phased haplotype data from 1000
set of J causal variants (J = 100, 1,000 or 10,000) either at randomly Genomes used as reference panel for HAPMIX can be accessed through
selected genomic positions, or with clustering. For the clustered set- https://www.internationalgenome.org/category/data-access/. POBI
ting we selected J/ 10 nonoverlapping 10 KB regions, each containing was accessed using accession no. EGAS00001000672. GWAS sum-
an average of five causal variants. The number of variants placed in mary level data used in this paper can be queried using the interface
each region is drawn from a multinomial distribution with param- implemented by ‘mrcieu/ieugwasr’: https://gwas.mrcieu.ac.uk and
eters J/2 (number of clustered variants) and pk, k = 1, 2, … C/10 where through Open Target at https://www.opentargets.org. HapMap3 vari-
p1 = p2 = … = pk. The remaining 50% of causal variants were uniformly ants list can be accessed at https://ftp.ncbi.nlm.nih.gov/hapmap/. All

Nature Genetics
Article https://doi.org/10.1038/s41588-024-02035-8

genetic data used in constructing the ancestry pipeline is provided 66. Hu, S., Lawson, D. & Myers, S. fuopen/UKB_anc: Source code of
by third parties and is available for use by others. Variant Frequency analysis for reproducing the main results/figures in ANCHOR
information for every SNP in each genetic grouping is available at paper. v0.2 edn. Zenodo https://doi.org/10.5281/zenodo.14026928
the University of Bristol data repository, data.bris, at https://doi. (2024).
org/10.5523/bris.3g5oatl682kz82as80jakjrq91. All other resources
were downloaded from their respective websites without registration Acknowledgements
requirements. The following files have been returned to UK Biobank This research has been conducted using the UK Biobank Resource
so that they might be made available to other researchers: (1) ACs under application number 27960. We thank all the participants of
on all UK Biobank participants; (2) group-specific allele frequency the UK Biobank project. POPRES was accessed under application
estimates for 25 M variants; (3) (Mean-centered) European/African number 16508: ‘Admixture and Selection for Fine-Scale Population
genotypes and local ancestry of 8,003 UK Biobank African ancestry Genetics of Europe’. This work was supported by the Wellcome Trust
individuals (including the variants annotation information); (4) Euro- grant no. 200186/Z/15/Z awarded to J.M., S.R.M. and G.H., grant no.
pean/African PGS for 8,003 African ancestry individuals across 53 UK 224575/Z/21/Z awarded to G.H. and grant no. 212284/Z/18/Z awarded
Biobank phenotypes. GWAS summary statistics files with population to S.R.M. L.A.F.F. is supported by the Wellcome Trust grant no.
structure corrected by ACs and PCs are available in GWAS catalog 222334/Z/21/Z. Computations used the high-performance computing
(https://www.ebi.ac.uk/gwas/, GCST90310137–GCST90310200 and facilities at the Department of Statistics, University of Oxford and
GCST90429571–GCST90429610). Oxford Biomedical Research Computing facility, a joint development
between the Wellcome Centre for Human Genetics and the Big Data
Code availability Institute supported by Health Data Research UK and the NIHR Oxford
The ANCHOR software package is available via GitHub at https://github. Biomedical Research Centre. This work was carried out using the
com/MyersGroup/ANCHOR and Zenodo at https://doi.org/10.5281/ computational facilities of the Advanced Computing Research Centre,
zenodo.13847648 (ref. 65). Analysis scripts can be downloaded via University of Bristol—http://www.bris.ac.uk/acrc.
GitHub at http://github.com/fuopen/UKB_anc and Zenodo at https://
doi.org/10.5281/zenodo.14026928 (ref. 66). External software/pack- Author contributions
ages used in this study are available via GitHub at https://github.com/ S.R.M., D.J.L., J.M. and G.H. designed the study. S.H., G.H., J.M., D.J.L.
danjlawson/pcapred.largedata (‘pcapred.largedata’) and https:// and S.R.M. developed the methods. S.H., G.H., J.M., D.J.L. and S.R.M.
github.com/danjlawson/pcapred (‘pcapred’). performed the main analyses with contributions for specific analyses
from L.A.F.F. and S.S. S.H., G.H., J.M., D.J.L. and S.R.M. wrote the
References manuscript.
57. Chang, C. C. et al. Second-generation PLINK: rising to the
challenge of larger and richer datasets. Gigascience 4, s13742- Competing interests
015-0047-8 (2015). S.H. became a full-time employee of Novo Nordisk Ltd during the
58. Walter, K. et al. The UK10K project identifies rare variants in health drafting of this manuscript. J.M. is a current employee and stockholder
and disease. Nature 526, 82–90 (2015). of Regeneron Pharmaceuticals. The other authors declare no
59. Abraham, G., Qiu, Y. & Inouye, M. FlashPCA2: principal competing interests.
component analysis of Biobank-scale genotype datasets.
Bioinformatics 33, 2776–2778 (2017). Additional information
60. Mountjoy, E. et al. An open approach to systematically prioritize Extended data is available for this paper at https://doi.org/10.1038/
causal variants and genes at all published human GWAS s41588-024-02035-8.
trait-associated loci. Nat. Genet. 53, 1527–1533 (2021).
61. Band, G. & Marchini, J. BGEN: a binary file format for imputed Supplementary information The online version contains
genotype and haplotype data. Preprint at BioRxiv https://doi. supplementary material available at https://doi.org/10.1038/s41588-
org/10.1101/308296 (2018). 024-02035-8.
62. International HeatMap 3 Consortium. Integrating common and
rare genetic variation in diverse human populations. Nature 467, Correspondence and requests for materials should be addressed to
52–58 (2010). Sile Hu, Daniel J. Lawson or Simon R. Myers.
63. 1000 Genomes Project Consortium. A global reference for human
genetic variation. Nature 526, 68–74 (2015). Peer review information Nature Genetics thanks Hakhamanesh
64. Speed, D. et al. Reevaluation of SNP heritability in complex Mostafavi and Benjamin Neale for their contribution to the peer review
human traits. Nat. Genet. 49, 986–992 (2017). of this work. Peer reviewer reports are available.
65. Hu, S. & Myers, S. MyersGroup/ANCHOR: ANCHOR v0.1-beta.1.
v0.1 edn. Zenodo https://doi.org/10.5281/zenodo.13847648 Reprints and permissions information is available at
(2024). www.nature.com/reprints.

Nature Genetics
Article https://doi.org/10.1038/s41588-024-02035-8

Extended Data Fig. 1 | Heatmap of regional mean ancestry proportions of 30 UK+ EU ancestries based on birthplace among UK-born and Ireland-born UKB
participants. One panel per ancestry, brightest red colour on maps indicates highest ancestry, darkest blue lowest ancestry.

Nature Genetics
Article https://doi.org/10.1038/s41588-024-02035-8

a. Central Asia b. Europe

c. Africa d. Americas

Self-reported Ethnicity Key: Europe Asia African Oth

Bri=British, Whi=White, Ire=Irish Chi:Chinese, Asi:Asian, Pa:Pakistani, Ind:Indian,Mix:Mixed African, Car:Caribbean Other Ethnic Group

Extended Data Fig. 2 | World ancestries in the UKB inferred by the ancestry Whi = White, Ire = Irish, Chi = Chinese, Asi = Asian, Pa = Pakistani, Ind = Indian,
pipeline. UKB participants are mapped to their self-reported country of birth. Mix = Mixed, Car = Caribbean, and Oth = Other Ethnic Group. a). Ancestries in
World countries are coloured by different colours if they are present in the Central Asia; b). Ancestries in Europe; c). Ancestries in Africa; d). Ancestries in the
pipeline. Ethnic groups are abbreviated as follows in the bar plots: Bri = British, Americas. White lines on the map delineate the borders of different countries.

Nature Genetics
Article https://doi.org/10.1038/s41588-024-02035-8

1.8

Aberdeen (n=833)

Fife
1.6
Angus
Perthshire and Kinross
Dundee (2,649)
Glasgow(12,132)
Edinburgh (8,022)

Cumbria (2,123) 1.4

Northumberland (42,765)

Hartlepool
Stockon-on-Tees
West Yorkshire(30,120) Middlesbrough (1,894)
Liverpool (14,683) East Riding of Yorkshire 1.2
Manchester(31,359) North East Lincolnshire
(1,439)
Birmingham (12,434)

Norfolk (1,326) 1.0

Anglese
Gwynedd
(588)
Suffolk (1,099)

Ceredigion
Carmarthenshire 0.8
London (41,325)
(935)
Cardiff (5,578)
Oxfordshire (7,366)

Plymouth (944)
Isle of Wight (258)
Bristol (7,487)

East Anglia North West Wales Lithuania

South Central England South Yorkshire Norway
Central England Cornwall Poland
North Central England Cornwall Tip Spain
North Yorkshire Orkney Switzerland
Northumberland Lincolnshire Balkans
South East England Cumbria Portugal
Merseyside North West Scotland Netherlands
South England North Ireland and South Scotland Italy
South East Wales Ireland Tuscany
North East Scotland France Syria
South West Wales Cyprus Denmark
Devon Germany

Extended Data Fig. 3 | Heatmap of regional mean entropy statistic across and areas with low entropy are coloured by blue. Selected regions of high or low
the UK and Republic of Ireland. Regional mean entropy indicates varying entropy are highlighted by the boundaries, with the barplots detailing ancestry
degrees of admixture in the UK. Areas with high entropy are coloured by red decomposition for each such region.

Nature Genetics
Article https://doi.org/10.1038/s41588-024-02035-8

a b

1.0
Place of birth (North/South)
Place of birth (East/West)
1.6 Place of birthe (North/South)
Place of birth (East/West)
Home location(East/West)

0.8
Home location (North/South)

LDSC attenuation ratio(PC)

LDSC Intercept(PC)
1.4

0.6
Home location (North/South)

Housing score (England)

Home location (East/West)

0.4
1.2

Employment score (England)

Education score (England)

0.2
Employment score (England)
Education score (England)
Housing score (England)
1.0

0.0
1.0 1.2 1.4 1.6 0.0 0.2 0.4 0.6 0.8 1.0
LDSC Intercept(AC) LDSC attenuation ratio(AC)

Extended Data Fig. 4 | Comparison of inflation between ACs-corrected corrected GWAS >0.015; b). Comparison of the LDSC attenuation ratio between
GWAS and PC-corrected GWAS across 104 quantitative traits by running LD AC-corrected GWAS (x-axis) and PC-correct GWAS (y-axis), where red dots
score regression (LDSC). a). Comparison of the LDSC intercept between AC- indicate phenotypes with difference of attenuation ratio between PC-corrected
corrected GWAS (x-axis) and PC-corrected GWAS (y-axis) where red dots indicate GWAS and AC-corrected GWAS >0.02.
phenotypes with intercept difference between PC-corrected GWAS and AC-

Nature Genetics