henhaobio
Sign in

ScienceGenetic & Genomic MedicinePeer-reviewed journal article

Diversity and Scale: Genetic Architecture of 2,068 Traits in the VA Million Veteran Program

Anurag VermaiD, Jennifer E. HuffmaniD, Alex A RodrigueziD, Mitchell ConeryiD, Molei LiuiD, Yuk-Lam Anne HoiD, Youngdae KimiD, David HeiseiD, Lindsay A. GuareiD, Vidul Ayakulangara PanickaniD, Helene Garcon, Franciel LinaresiD, Lauren CostaiD, Ian GoethertiD, Ryan Tipton, Jacqueline P. HonerlawiD, Laura R. DaviesiD, Stacey B. WhitbourneiD, Jeremy CoheniD, Daniel C. PosneriD, Rahul Sangar, Michael Murray, Xuan WangiD, Daniel R. DochtermanniD, Poornima Devineni, , Katherine P. LiaoiD

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

Abstract

Genome-wide association studies (GWAS) have underrepresented individuals from non-European populations, impeding progress in characterizing the genetic architecture and consequences of health and disease traits. To address this, we present a population-stratified phenome-wide GWAS followed by a multi-population meta-analysis for 2,068 traits derived from electronic health records of 635,969 participants in the Million Veteran Program (MVP), a longitudinal cohort study of diverse U.S. Veterans genetically similar to the respective African (121,177), Admixed American (59,048), East Asian (6,702), and European (449,042) superpopulations defined by the 1000 Genomes Project. We identified 38,270 independent variants associating with one or more traits at experiment-wide (P < 4.6x10-11) significance; fine-mapping 6,318 signals identified from 613 traits to single-variant resolution. Among these, a third (2,069) of the associations were found only among participants genetically similar to non-European reference populations, demonstrating the importance of expanding diversity in genetic studies. Our work provides a comprehensive atlas of phenome-wide genetic associations for future studies dissecting the architecture of complex traits in diverse populations. One Sentence SummaryTo address the underrepresentation of non-European individuals in genome-wide association studies (GWAS), we conducted a population-stratified phenome-wide GWAS across 2,068 traits in 635,969 participants from the diverse U.S. Department of Veterans Affairs Million Veteran Program, with results expanding our knowledge of variant-trait associations and highlighting the importance of genetic diversity in understanding the architecture of complex health and disease traits.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (29 Jun 2023) · CC BY 4.0

Population-stratified GWAS of 2,068 traits in 635,969 Million Veteran Program participants, 29% genetically non-European, yields 98,485 associations and 57,601 fine-mapped signals with largely shared cross-population architecture.

About 95% of published GWAS participants are genetically similar to European reference populations, which limits how far findings generalize. This phenome-wide GWAS of an electronic health record (EHR)-linked biobank with 121,177 African-similar (AFR) and 59,048 Admixed American-similar (AMR) Veterans finds loci missed by European-only (EUR) analysis and fine-maps more traits and signals than earlier large efforts, especially in AFR and AMR.

Key results

  1. ShownMeta-analysis across four population groups gave 98,485 study-wide significant associations for 1,274 traits with well-controlled inflation; the paper reports high novelty rates against GWAS catalogs, though the denominators behind those rates are unclear.

    98,485 associations (38,269 lead variants, 23,584 loci) at P < 4.6 × 10^-11; genomic inflation 1.01 (0.85-1.19). Not previously identified as lead or proxy (r^2 > 0.6, 500 kb) in GWAS Catalog or Open Targets Genetics: 85.6% (158) of variant-binary and 63.3% (496) of variant-quantitative trait associations (counts as reported; denominator unstated); about 57% of novel variants had minor allele frequency (MAF) < 1%. · Biobank-scale genomic analysis across…; Determination of the number of independent…

  2. ShownAdding AFR, AMR and East Asian-similar (EAS) participants found loci that EUR-only GWAS missed, hundreds of them not even nominally significant in EUR, while recovering known African-enriched variants.

    1,608 loci not significant in EUR-only GWAS, giving 3,477 variant-trait associations across 893 traits (76% binary); 834 had EUR P > 0.05, e.g. a non-coding FAM234A variant with iron deficiency anemias (P_AFR = 2.37 x 10^-37, P_EUR = 0.42). Recovered known signals include rs72725854 near PCAT2 (prostate cancer), APOL1 rs73885319 (kidney disease), ACKR1 (neutropenia) and rs334 in HBB. · Multi-population meta-analysis improves the…

  3. ShownSuSiE fine-mapping with in-sample, population-matched linkage disequilibrium (LD) matrices gave 57,601 merged signals, a quarter resolved to one variant; AFR credible sets were slightly smaller than EUR and AMR sets larger, yet about 90% of single-variant signals were non-coding.

    57,601 signals across 936 phenotypes; 14,405 (25%) single-variant and >54% with ≤5 variants; 15,045 variant-trait pairs at posterior inclusion probability (PIP) > 0.95, 2,069 only in non-EUR groups. Paired Wilcoxon: AFR sets smaller than EUR (P = 2.26 × 10^-10) though median difference was zero; AMR larger than AFR (P = 1.30 × 10^-84) and EUR (P = 7.36 × 10^-162). · Fine mapping of multi-population associations…; Discussion

  4. ShownSNP heritability and genetic correlations indicate mostly shared architecture across groups: no trait showed opposite-sign genetic correlation between populations, but magnitudes varied and several traits (e.g. skin cancer, anemia of chronic disease) correlated only weakly across populations.

    LD score regression (LDSC) heritability significant (P < 9 × 10^-6) for 233 AFR, 199 AMR, 3 EAS and 816 EUR traits; height most heritable in every group. BMI and type 2 diabetes genetic correlation 0.47 (0.41-0.52) in EUR vs 0.28 (0.18-0.39) in AFR. Popcorn EUR-AFR correlation significant for 212 traits; height 0.66, type 2 diabetes 0.65, skin cancer 0.05, white blood cell count 0.20. · Population-specific heritability and genetic…

  5. ShownEffect-size heterogeneity between populations was rare among common fine-mapped variants; the clearest case was the APOE-e4 tag rs429358, whose dementia risk was lower in AFR than EUR with the same direction of effect.

    Broad screen of common signals (MAF > 0.05) at 1,888 loci: 16 heterogeneous variant-trait associations AFR vs EUR, 11 AMR vs EUR. Coding variants with PIP > 0.95 in both groups: 6 AFR vs EUR, 2 AMR vs EUR, all same direction. rs429358 dementia risk 30% lower in AFR than EUR; no AMR-EUR difference; EAS underpowered. rs429358 is also linked to 24 traits, newly chronic liver disease and cirrhosis. · Fine-mapped variant-trait associations specific…; Cross-trait genetic architecture identifies…; Estimating heterogeneous effects

Methods, data and reproducibility

Per-population GWAS with a GPU-optimized SAIGE generalized linear mixed model on the Department of Energy Summit supercomputer (covariates: age, sex, ten genetic principal components), fixed-effect GWAMA meta-analysis, PLINK clumping, per-population SuSiE fine-mapping merged by PIP-weighted Jaccard similarity, and LDSC/Popcorn heritability and genetic correlation.

  • VA Million Veteran Program Genomics Release 4: Axiom MVP 1.0 genotypes linked to VA EHR (followed to September 2020) · human · n = 635,969 participants
  • Population groups assigned by random forest on PCs projected from 1000 Genomes superpopulations · human · n = AFR 121,177; AMR 59,048; EAS 6,702; EUR 449,042
  • Traits from ICD-9/ICD-10 Phecodes, laboratory measures, vital signs and enrollment survey responses · human · n = 1,854 binary and 214 quantitative traits (2,068)
  • Imputed variants after QC (imputation quality > 0.3, minor allele count > 20) · human · n = > 44.3M variants
  • Imputation reference: African Genome Resources panel plus 1000 Genomes Project p3v5 · human
  • Full GWAS summary statistics for all traits and population groups · human · n = 4,045 population-specific GWAS plus meta-analyses · dbGaP phs002453
Methods & tools
ThermoFisher Axiom MVP 1.0 array, SAIGE (GPU-optimized), GWAMA, PLINK, SuSiE (via coloc R package), LDSC, Popcorn, EIGENSOFT smartpca, Variant Effect Predictor (VEP), RegulomeDB v2.2, EasyQC, NHGRI-EBI GWAS Catalog; Open Targets Genetics
Reagents & environment
Partial only: RegulomeDB v2.2 and 1000 Genomes Project p3v5 reference named; SuSiE run through the coloc R package; no versions for SAIGE, GWAMA, PLINK, LDSC or Popcorn, and no container or requirements file.
Assay conditions & compute
Oak Ridge Leadership Computing Facility Summit (GWAS) and Andes (QC, meta-analysis); 14,286 GPU hours over 14 days wall time, reported as a 160-fold reduction in core hours (existing SAIGE estimated at ~251 compute years); LD-matrix and fine-mapping jobs capped at 48 h and 1 TB RAM each.
Reproducibility gaps
No code link for the GPU-optimized SAIGE or the fine-mapping pipeline. Summary statistics are listed under dbGaP phs002453 (on request), though another section says they were still being processed for transfer.
Limitations (6)
  • Phenotypes were defined automatically from Phecodes (efficient but imprecise for most traits), and models were standardized: covariates limited to age, sex and principal components, with quantitative traits inverse-normal transformed. · stated by the authors
  • LD matrices came from hard-called genotypes without covariates, so they did not fully match the SAIGE analysis; the authors note the resulting LD mismatch could yield inaccurate or spurious fine-mapping results at sensitive loci. · stated by the authors
  • Loci were defined from meta-analysis results, so some regions without a population-specific signal were fine-mapped (especially in EAS, the smallest and underpowered group), population-specific peaks may have been masked, very large loci were not fully mapped, and 115 signals were double-counted. · stated by the authors
  • Choosing precision over recall (significance and purity thresholds, at most five signals per locus) may miss true signals and undercount signals at strongly associated loci. · stated by the authors
  • Known/novel labels rest on catalog lookup after partly fuzzy text mapping of traits to EFO terms, and the proxy threshold differs between Results (r^2 > 0.6) and Methods (r^2 > 0.1); high-confidence pair counts are also inconsistent (25,953 in S9 vs 15,045 in S6). · noted in this summary
  • All discoveries come from one biobank; no independent cohort is used to replicate the novel or non-EUR-specific associations, which rest on MVP statistics plus known-signal positive controls. · noted in this summary

Read this if you need multi-population, phenome-wide GWAS summary statistics or fine-mapped credible sets from an EHR biobank with large African- and Admixed American-similar groups, or want to see how SAIGE was scaled on GPUs.

Citations
462
FWCI
121.59
Field percentile
top 1%

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
THISWendt '22Hunter-Zinc '20Troubat '24Wojcik '18Levin '25Jo '26
72% semantic matchGenomics
Frank R. Wendt, Gita A. Pathak, Jacqueline M. Vahey+20 more·bioRxiv·Jan 24, 2022· 4 citations