henhaobio
Sign in

medRxivGenetic & Genomic MedicinePreprint

Pan-UK Biobank GWAS improves discovery, analysis of genetic architecture, and resolution into ancestry-enriched effects

Konrad J. KarczewskiiD, Rahul GuptaiD, Masahiro KanaiiD, Wenhan LuiD, Kristin TsuoiD, Ying WangiD, Raymond Kenney WaltersiD, Patrick TurleyiD, Shawneequa CallieriD, Nirav N. Shah, Nikolas A. BayaiD, Duncan S. PalmeriD, Jacqueline I. GoldsteiniD, Gopal SarmaiD, Matthew SolomonsoniD, Nathan Cheng, Sam Bryant, Claire L. ChurchhouseiD, Caroline M. CusickiD, Timothy PoterbaiD, John Compitello, Daniel King, Wei ZhouiD, Cotton Seed, Hilary K. FinucaneiD, , Alicia R. MartiniD

PDFDOIData
In-depth full-text digest available · Jump to digest ↓

Abstract

Large biobanks, such as the UK Biobank (UKB), enable massive phenome by genome-wide association studies that elucidate genetic etiology of complex traits. However, individuals from diverse genetic ancestry groups are often excluded from association analyses due to concerns about population structure introducing false positive associations. Here, we generate mixed model associations and meta-analyses across genetic ancestry groups, inclusive of a larger fraction of the UKB than previous efforts, to produce freely-available summary statistics for 7,266 traits. We build a quality control and analysis framework informed by genetic architecture. Overall, we identify 14,676 significant loci (p < 5 x 10-8) in the meta-analysis that were not found in the EUR genetic ancestry group alone, including novel associations for example between CAMK2D and triglycerides. We also highlight associations from ancestry-enriched variation, including a known pleiotropic missense variant in G6PD associated with several biomarker traits. We release these results publicly alongside FAQs that describe caveats for interpretation of results, enhancing available resources for interpretation of risk variants across diverse populations.

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 2 (1 Oct 2024) · CC BY 4.0

Meta-analyzing per-ancestry GWAS of 7,266 UK Biobank traits across six genetic ancestry groups adds 14,676 loci missed by European-only analysis, including ancestry-enriched CAMK2D and G6PD signals.

Phenome-wide UK Biobank GWAS have mostly analyzed only the European-ancestry (EUR) group, omitting more than 20,000 participants with mainly non-European ancestries over stratification concerns and losing power from ancestry-enriched variants. This resource adds open multi-ancestry summary statistics, a heritability-based quality-control framework for highly imbalanced sample sizes, and evidence that per-group meta-analysis curbs stratification better than mega-analysis.

Key results

  1. ShownAncestry-stratified SAIGE mixed-model GWAS plus inverse-variance meta-analysis produced summary statistics for 16,528 ancestry-trait pairs over 7,266 traits, with non-EUR groups still yielding independent loci.

    Up to 441,331 people in 6 groups (20,800 non-EUR; 42,535 unclassified); 10-23 million SNPs. For 922 traits run in all six groups, non-EUR averaged 2.26 independent loci; 25% of traits had >18 EUR loci vs under 10% with 3 or more across non-EUR groups. Despite smaller N, the African (AFR) group found 15% more height hits than the Central/South Asian (CSA) group (1,261 vs 1,089). · A resource of multi-ancestry association results…; Tables

  2. ShownMeta-analysis tracked EUR-only significance yet added loci missed by EUR GWAS; versus Open Targets Genetics per Experimental Factor Ontology (EFO) category, a minority of hits were unreported, mostly boosted EUR-suggestive signals, overrepresented on chromosome X and near haploinsufficient genes.

    Meta vs EUR p-value r = 0.999 (452 phenotypes); 14,676 of 237,360 (6.2%) LD-independent hits across 431 phenotypes not significant in EUR. Of 71,372 EFO-mapped regional hits, 3,112 (4%) had no known same-category association within 1 Mb (3,092 suggestive in EUR; chrX 573 of 2,448, 23%); 66% (124/187) of haploinsufficient genes vs 34% of all genes lie near one. · Genomic discoveries powered by Pan-UK Biobank…; Figure legends

  3. ShownAllele-frequency differences drive meta-vs-EUR discrepancies: variants rare in EUR but common in smaller groups became detectable, e.g., an AFR-enriched CAMK2D variant for triglycerides and the G6PD missense rs1050828, pleiotropic for red-cell traits and glycated hemoglobin (HbA1c).

    Meta-only hits were 6-fold enriched for variants commoner in AFR (4-fold, any non-EUR). CAMK2D rs193059864 (AFR frequency 1.6%, 114-fold enriched; meta N = 416,764): EUR p = 0.0017. G6PD rs1050828 (AFR frequency 16%) hit five traits in AFR; HbA1c meta p = 1.1 x 10^-299 (N = 408,539). Variants weaker in meta than EUR were 2.4-fold AFR-enriched, 2.5-fold heterogeneity-enriched. · Genomic discoveries powered by Pan-UK Biobank…; Inclusion of multiple genetic ancestries…; Refining biological signals through ancestry…; Figure legends

  4. ShownA stepwise quality-control cascade (SNP-heritability z, genomic inflation, residual-stratification ratio, multi-group support) kept a small reliable subset of GWAS; retained traits showed correlated heritability between EUR and CSA groups, higher in CSA.

    16,528 ancestry-trait pairs pruned to 1,091 passing all filters across 452 traits (147 phenotypes in three or more groups; 151 independent at pairwise r < 0.1). For 113 phenotypes passing in both EUR and CSA, York regression slope 0.49, intercept 0.21. Mean EUR SNP-heritability: biomarkers 0.19, continuous 0.16, diseases 0.021, prescriptions 0.016. · A framework for identifying high-quality…; Figure legends

  5. SuggestedAncestry-stratified meta-analysis curbed residual stratification (lower genomic inflation, λGC) relative to a single pooled mixed model without an apparent loss of loci, and a Tractor pilot modeling ancestral haplotypes in AFR further improved false-positive control.

    For five phenotypes under EUR-only, mega- and meta-analysis (full and EUR-matched sample sizes), meta-analysis gave systematically lower λGC than mega-analysis with similar independent-locus counts; sample-size-scaled λ1000 was near 1 throughout. The Tractor pilot (AFR, subset of traits) improved false-positive control and group-level effect estimates relative to SAIGE. · A resource of multi-ancestry association results…; Refining biological signals through ancestry…; Discussion; Tables

Methods, data and reproducibility

Participants were assigned to six genetic ancestry groups via random-forest classification on Human Genome Diversity Panel plus 1000 Genomes principal components; per-group SAIGE mixed-model GWAS were combined by fixed-effect inverse-variance meta-analysis, with heritability-guided trait quality control (S-LDSC, RHE-mc), LD clumping and fine-mapping.

  • UK Biobank participants in six genetic ancestry groups (EUR, CSA, AFR, East Asian EAS, Middle Eastern MID, Admixed American AMR), after outlier removal · human · n = up to 441,331; 20,800 in non-EUR groups · UK Biobank return 2442 (Pan-UKB sample metadata)
  • Quantitative and binary UK Biobank phenotypes, including biomarkers, phecode diseases, prescriptions and COVID-19 traits · human · n = 7,266 traits; 16,528 ancestry-trait pairs (922 in all six groups)
  • Human Genome Diversity Panel and 1000 Genomes Project genotypes as PCA reference for ancestry assignment (plus AGVP for AFR PC space) · human
  • LD reference panels matched by genetic ancestry proportions, used to define independent meta-analysis loci · human · n = 5,000 individuals
  • Open Targets Genetics and GWAS Catalog associations, linked through Experimental Factor Ontology terms, for novelty assessment · human · n = 3,047 traits mapped to EFO; 2,566 matched a GWAS Catalog study
  • EUR pilot phenotypes for comparing S-LDSC and RHE-mc heritability estimates · human · n = 64 phenotypes
  • Phenotypes used to compare EUR-only, mega-analysis and meta-analysis inflation and locus counts · human · n = five phenotypes
  • Haploinsufficient genes versus all genes, tested for nearby previously unreported associations · human · n = 187 haploinsufficient genes; 17,428 genes
Methods & tools
SAIGE generalized mixed-model association testing, Fixed-effect inverse-variance weighted meta-analysis, PCA projection with random-forest ancestry classifier (probability > 0.5), S-LDSC (heritability and ratio statistic), RHE-mc (randomized multi-component Haseman-Elston regression), PLINK clumping, Tractor (ancestry-specific haplotype GWAS, AFR pilot), Cochran's Q heterogeneity test, York regression, Open Targets Genetics and GWAS Catalog, Experimental Factor Ontology (EFO), gnomAD, ClinVar and OMIM annotations
Reagents & environment
SAIGE, PLINK, S-LDSC, RHE-mc and Tractor are named, but no software versions, containers or environment specifications are reported.
Reproducibility gaps
Release covers summary statistics for 16,528 GWAS, LD scores/matrices and returned sample metadata; pipelines are described as released but no code URL is printed; compute resources are not reported.
Limitations (6)
  • Discrete genetic ancestry groups simplify continuous ancestry, risk reifying race or ethnicity, and methods assuming homogeneity can fail in recently admixed genomes; loci seen in one group mainly reflect power, not biological differences. · stated by the authors
  • Fine-mapping meta-analyzed summary statistics proved unstable under highly imbalanced sample sizes, as at G6PD, where the small AFR group supplies an outsized share of variance explained but little of the LD panel; the authors call for more balanced cohorts. · stated by the authors
  • Trait-level quality control is coarse: failing it does not exclude true SNP associations (some G6PD-associated traits failed in AFR); heritability in smaller groups is likely inflated by winner's curse and residual stratification, and winner's curse may inflate some effect sizes. · stated by the authors
  • Novelty calls depend on semi-manual EFO mapping, completed for 3,047 (42%) traits with trait-type-dependent success, and on 1 Mb distance windows, since LD-based independence yields spurious signals beside extremely strong associations. · stated by the authors
  • The meta- versus mega-analysis comparison spans only five phenotypes, so its advantage across thousands of traits with differing stratification is extrapolated rather than tested. · noted in this summary
  • Newly reported associations such as CAMK2D-triglycerides and PITX2-keratometry show no independent-cohort replication or functional follow-up in the text, and the Tractor pilot is described without quantitative metrics. · noted in this summary

Read this if you need phenome-wide UK Biobank GWAS summary statistics for non-EUR ancestry groups, design multi-ancestry GWAS and quality control under imbalanced sample sizes, or study ancestry-enriched variants like G6PD rs1050828.

Citations
166

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISSun '21Li '23Smith '22Constantine '22Smeriglio '26Pankratov '24
71% semantic matchGenetics
Quan Sun, Mariaelisa Graff, Bryce Rowland+16 more·Journal of Human Genetics·Aug 10, 2021· 53 citations