henhaobio
Sign in

bioRxivGenomicsPreprint

Representation in genetic studies affects inference about genetic architecture

Jared M. ColeiD, Shane Rybacki, Samuel Pattillo SmithiD, Olivia S. SmithiD, Arbel HarpakiD

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

Abstract

Knowledge of a trait's "genetic architecture," namely the joint distribution of allele frequencies of causal variants and the direction and magnitude of their effects, is essential to understanding its evolution and underlying biology. Inferences about genetic architecture are based on data collected in different ways in cohorts recruited through heterogeneous mechanisms. As a result, cohorts differ in genotype, environment, and trait distributions. For example, the UK Biobank (UKB) was designed for broad population representation, whereas FinnGen drew extensively from clinical registries enriched for diagnosed health conditions. Here, we asked whether representation in genetic studies influences inferences about genetic architectures. Using GWAS data from the UKB, All of Us (AoU), and FinnGen, we compared several summaries of genetic architecture. We find systematic differences in some summaries, such as SNP heritability, which were often (in 6/13 traits) significantly lower in AoU than in the UKB (and never the reverse), even when matching samples such that they have similar genetic ancestry compositions. This result aligns with other recent evidence that biobanks enriched for diagnosed health conditions, also sometimes characterized by less-standardized phenotyping, have lower heritability than population-based biobanks. We highlight a second case, where a summary of genetic architecture varies considerably but not systematically across traits and biobanks. Such is the case for the mean direction of allelic effects ("sign bias"). For example, 72% of rare minor alleles affecting type 2 diabetes risk are inferred to be risk-increasing based on AoU data, while nearly all (>99%) are inferred to be risk-increasing based on UKB data. We hypothesize that the inferred sign bias is heavily influenced by the skewness of the trait distribution in the study and otherwise largely independent of other study or trait characteristics, including whether the trait is binary or quantitative. We provide strong support for this hypothesis through simulations and data from the three biobanks: the variation in inferred sign bias for rare minor alleles across traits and biobanks is explained remarkably well (82% and 97% of variance explained for trait-associated SNPs and randomly selected SNPs, respectively) solely by the trait's skewness in the biobank, with residual biobank-specificity explaining little. Our findings suggest that inferences about the map between genetic and trait variation can depend on study design and participation in genetic studies in surprising ways.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 3 (19 Aug 2026) · CC BY 4.0

Across UK Biobank, All of Us and FinnGen, the inferred mean direction (sign bias) of rare minor-allele effects varies and is largely explained by each cohort's trait skewness.

Summaries of genetic architecture are often assumed to be trait-intrinsic, with cross-study differences mostly blamed on ancestry, yet cohorts also differ in country, healthcare system and recruitment, and reweighting does not probe the drivers. Comparing ancestry-matched UK Biobank (UKB) and All of Us (AoU) GWAS, plus FinnGen, the work ties inferred sign bias to cohort trait skewness and, to the authors' knowledge for the first time, shows lower AoU SNP heritability under ancestry matching.

Key results

  1. ShownAcross traits and biobanks, the inferred sign bias of rare minor alleles (frequency up to 0.1%) is predicted mainly by the trait's skewness in that cohort; biobank-specific terms add little.

    Most-associated SNP per linkage disequilibrium (LD) block: pooled Spearman ρ = 0.967 (P < 0.001); quadratic logit fit adjusted R² = 0.82 excluding one high-leverage outlier (white blood cell count in AoU); biobank intercepts add +1.7% and further biobank slopes +3.3% adjusted R². Random SNPs: ρ = 0.843, adjusted R² = 97%. With the outlier kept, R² = 0.76 and biobank slopes add +10.2%. · Skewness of the trait distribution in a study…; Evaluating the relationship between trait…; Abstract

  2. ShownSign bias is negligible for common alleles (frequency ≥ 0.1) but rises for rarer ones, most for binary diseases. At the rarest frequencies, biobanks disagree sharply on it for the same trait.

    Share of most-associated minor alleles (frequency ≤ 0.1%) inferred risk-increasing: type 2 diabetes 99.2% (UKB), 72.2% (AoU), 57.3% (FinnGen); asthma 96.5%, 71.7%, 62.2%; schizophrenia and Alzheimer's disease >99% in UKB and AoU vs 85.5% and 73.7% in FinnGen. Mean corpuscular hemoglobin: 62.5% trait-increasing in UKB vs 43.6% in AoU. · Sign bias: A case study of drivers of…; Abstract

  3. SuggestedIn simulations with zero sign bias in the population, sampling right-skewed cohorts makes the estimated sign bias rise monotonically with cohort skewness, while the cohort's true sign bias changes little.

    Population of 2 × 10^6 with M = 4,000 variants, equal numbers of trait-increasing and trait-decreasing effects of equal magnitude (h² = 0.6). Cohorts of n = 10,000 at target skewness 1 to 10 came from threshold sampling of a Normal population or from mixing the central and right modes of a symmetric tri-modal population. Sign bias was estimated with adaptive shrinkage (ash) for minor alleles up to 1% frequency. · Skewness of the trait distribution in a study…; Simulating traits and associated allelic effects

  4. ShownIn ancestry-matched samples, LD Score Regression (LDSC) SNP heritability is on average lower in All of Us than in UK Biobank, whereas effective polygenicity differs only slightly for most traits.

    AoU SNP heritability approximately 17% ± 4% lower on average across traits (Deming regression). The largest AoU/UKB effective polygenicity (S-LD4M) ratios were type 2 diabetes 4.65 ± 3.09 and mean corpuscular hemoglobin 2.81 ± 1.18, both statistically insignificant. · Headline summaries of genetic architecture vary…; Estimates of genetic architecture summaries

  5. ShownDespite ancestry matching, genetic correlations between UKB and AoU allelic effects fall below 1 for several traits. Random subsampling within UKB does not reproduce this.

    Bivariate LDSC rg significantly different from 1 for 6 of 13 traits, e.g. basophil percentage 0.45 ± 0.07 and neutrophil percentage 0.88 ± 0.04. Genetic correlations between random UKB subsets never deviated significantly from 1 (Text S1, Fig. S9). · Headline summaries of genetic architecture vary…

Methods, data and reproducibility

All of Us GWAS on raw traits (PLINK 2.0), ancestry-matched to UK Biobank, were compared with Neale Lab UKB summary statistics via LD Score Regression (heritability, genetic correlation) and S-LD4M polygenicity; with FinnGen added, sign bias was estimated by ash using one SNP per LD block. Sign bias was regressed on cohort trait skewness (quadratic logit) and probed by simulation.

  • UK Biobank GWAS summary statistics from the Neale Lab (array and imputed SNPs; raw, untransformed quantitative traits) for 14 traits · human · n = 361,194 unrelated individuals; 13.7 million SNPs
  • All of Us v8 Curated Data Repository: whole-genome sequence ACAF subset plus measurement and ICD-derived traits; new GWAS in a subsample ancestry-matched to the UKB set · human · n = 221,105 matched; 68,492 to 199,529 per trait; 10,102,149 variants
  • FinnGen public GWAS summary statistics (DF12 release) for 5 disease endpoints · human · n = 500,348 samples; 21,311,644 SNPs
  • Simulated populations and cohorts under two sampling schemes, with no true sign bias · n = 2 × 10^6 individuals, 4,000 variants; cohorts of 10,000
  • Approximately independent LD blocks: 1000 Genomes Phase 3 EUR (Berisa and Pickrell) for UKB; hg38 EUR map (MacDonald et al.) for AoU · human (1000 Genomes, hg38 build) · n = 1,703 and 1,361 blocks
  • Pre-computed European-ancestry LD scores for HapMap3 variants (Bulik-Sullivan et al.) used with LDSC · human (HapMap3)
Methods & tools
PLINK 2.0, LD Score Regression (LDSC), S-LD4M, adaptive shrinkage (ash), random forest classifier, Global Biobank Meta-analysis Initiative PC loadings, Phecode map X, LOINC, dbSNP, Deming regression
Software & environment
PLINK 2.0 is the only software version given. Versions for LDSC, ash and S-LD4M are not stated, and no environment or requirement files are mentioned.
Reproducibility gaps
FinnGen enters only through public summary statistics and reported case counts, because the authors had no individual-level access. The text does not describe access conditions for the individual-level All of Us data behind the new GWAS.
Limitations (6)
  • Other researchers ran the FinnGen GWAS, with no ancestry matching to the UKB and AoU subsets. Without individual-level data, only five disease endpoints (skewness taken from reported case counts) could be used, and only in the sign bias analysis. · stated by the authors
  • The skewness model is sensitive to one trait: keeping white blood cell count in AoU lowers adjusted R² to 0.76 and raises the gain from biobank-specific slopes to +10.2%. Biobank terms also matter more at less stringent allele frequency cutoffs. · stated by the authors
  • Asymmetric mutational effects and purifying selection could also shape sign bias. The authors argue these forces are largely cohort-invariant and so cannot account for differences between biobanks. · stated by the authors
  • Alzheimer's disease was left out of the heritability-ratio, polygenicity and genetic correlation comparisons because its UKB SNP heritability estimate was negative. · stated by the authors
  • The findings concern untransformed traits. The authors show that inverse-rank normalization removes skewness and leaves little to no sign bias (Fig. S8), so the effect mainly matters for raw-scale analyses and binary traits. · noted in this summary
  • The skewness and sign bias regression rests on a modest number of trait-biobank points (14 traits in two biobanks plus 5 FinnGen diseases), with an influential point removed. In the real data, skewness is not varied independently of other cohort features. · noted in this summary

Read this if you compare heritability, genetic correlations or effect-direction summaries across biobanks or ancestry groups, or if you run GWAS on skewed raw-scale traits or diseases with few cases.

Citations
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
THISSchoeler '23Hou '19Haworth '19Souaiaia '25Ge '16Rawlik '20
66% semantic matchGenetics
Tabea Schoeler, Doug Speed, Eleonora Porcu+3 more·Nature Human Behaviour·Apr 27, 2023· 406 citations