Skip to content

Distinct Genetic Architecture in Trait Tails

Summary

Souaiaia et al. (2026) demonstrate that common-variant polygenic risk scores (PRSs) systematically regress toward the population mean in phenotypic extremes across 74 complex traits.[1] This departure is driven by rare, large-effect alleles concentrated in phenotypic tails as a consequence of stabilizing and directional selection.[1] Incorporating rare variants from sequencing data markedly reduces these tail deviations, increasing tail disease odds ratios by 71.7% despite only modest overall variance gains.[1]

Mathematical Formulation of Tail Outlier Tests

To evaluate whether genetic architecture varies across the phenotypic continuum, Souaiaia et al. introduced two complementary statistical methods designed to detect departures from linear common-variant expectations.[1]

1. Population-Based Outlier Test (POPout)

POPout evaluates whether observed PRS values in phenotypic extremes deviate from a linear expectation fitted across the full trait distribution.[1] For an individual $i$ with residual trait value $y_i$ and common-variant PRS $g_i$, the expected PRS is modeled via linear regression across the population:[1] $$\hat{g}i = \alpha + \beta y_i$$ The POPout effect size for a designated tail quantile threshold $\tau$ (e.g., lower or upper 1%) is computed as the standardized difference between mean observed and expected PRS values:[1] $$\text{POPout Effect} = \frac{\frac{1}{N$$ Under a pure common-variant polygenic model, $\text{POPout Effect} = 0$. A positive POPout effect indicates that extreme individuals carry fewer common risk alleles than predicted by their phenotype, signifying that large-effect rare variants uncaptured by common PRSs drive the extreme phenotype.[1]}}\sum_{i \in \text{tail}} (g_i - \hat{g}_i)}{\sigma_g

2. Sibling-Based Tail Architecture Test (STANDout)

STANDout is a family-based joint test ($\chi^2$ statistic with 4 degrees of freedom) that aggregates sibling trait correlations to resolve tail architecture without susceptibility to population stratification or unmeasured environmental confounding:[1] - Polygenic Common-Variant Model: Siblings of an extreme index individual regress partway toward the population mean proportional to narrow-sense heritability ($h^2$).[1] - Segregating Rare-Allele (Mendelian-like) Model: Siblings exhibit a bimodal phenotype distribution — 50% inherit the large-effect rare allele and remain extreme, while 50% lack the variant and match the background population mean.[1] - De Novo Mutation Model: Siblings show zero trait enrichment, completely matching the general population background.[1]

Trait Architecture Models in Phenotypic Tails
├── Common Polygenic Model  ──> Linear trait-PRS fit; siblings regress to mean by h^2
├── Segregating Rare Allele ──> Non-linear PRS regression; sibling bimodal distribution (50% extreme)
└── De Novo Mutation        ──> Non-linear PRS regression; siblings resemble population background

Empirical Trait Landscape & Risk Stratification Failure

Applied to 74 quantitative traits in the UK Biobank ($N = 369,132$ European-ancestry individuals in discovery/scoring splits):[1]

  • Pervasiveness of Tail Outliers: 68 of 74 traits exhibited significant POPout departures (FDR < 5%) in at least one tail.[1] Of 148 evaluated trait tails (lower/upper 1%), 108 were significant, with 98 displaying positive POPout effects.[1]
  • Tail-Depth Gradient: Sensitivity analyses across quantile cutoffs showed increasing POPout effect magnitudes deeper into phenotypic extremes: mean Z-score deviation rose from ~0.01 at 10% tail cutoffs to ~0.18 at 0.5% cutoffs and ~0.33 at 0.1% cutoffs.[1]
  • Representative Index Traits:
  • Red Blood Cell Distribution Width (RDW): Strong upper-tail departure ($\text{POPout Z} \approx 0.50$).[1]
  • Haemoglobin Concentration: Strong lower-tail departure ($\text{POPout Z} \approx 0.37$).[1]
  • Sitting Height: Symmetric departures in both lower and upper tails (symmetric stabilizing selection).[1]
  • Phosphate: Zero deviation in either tail, adhering strictly to a common polygenic model.[1]

Clinical Impact on Disease Risk Stratification

When phenotypic extremes define clinical disease cutoffs, common-variant PRSs fail precisely where clinical risk is highest:[1] - RDW (Emulating Top 1% Disease): The linear trait-PRS model predicted an odds ratio (OR) of 4.1 for individuals in the top 5% of PRS, but the observed OR was only 1.8.[1] - Haemoglobin (Emulating Lower 1% Anemia): Expected OR for the lowest 5% PRS was 3.7, whereas the observed OR was 1.8.[1] - Clinical Implication: Common PRSs stratify moderate risk effectively but underperform in identifying individuals at extreme risk of early-onset or severe disease driven by rare monogenic or oligogenic alleles.[1]

Robustness, Multi-Ancestry & Sibling Validation

POPout effect sizes replicated across three independent validation frameworks:[1] 1. UK Biobank Repeated Measures ($N = 12,007$ per trait): Pearson $R = 0.69$ ($P = 1.26 \times 10^{-15}$), confirming that tail departures are not driven by transient environmental exposures or measurement error.[1] 2. UK Biobank Multi-Ancestry Cohort ($N = 17,407$ per trait): Pearson $R = 0.62$ ($P = 3.63 \times 10^{-13}$), establishing that tail deviations are a universal feature across ancestries despite common PRS transferability decay.[1] 3. All of Us Research Program ($N = 53,196$ European-ancestry individuals per trait): Pearson $R = 0.61$ ($P = 1.07 \times 10^{-6}$), demonstrating cross-continent stability in distinct healthcare settings.[1] 4. Sibling STANDout Validation: STANDout statistics correlated strongly with POPout ($R = 0.60, P = 1.0 \times 10^{-15}$), and sibling-estimated heritability matched SNP heritability ($R = 0.80, P = 7.0 \times 10^{-18}$).[1]

Rare Coding & WGS Variant Integration

Integrating rare single-variant associations ($0.1\% < \text{MAF} < 1\%$ and $0.01\% < \text{MAF} < 0.1\%$) from WGS and gene-burden signals from WES:[1]

  • Reduction of POPout Outliers: Adding rare-variant PRSs reduced POPout departures in 56 of 66 testable tails.[1] 11 traits achieved statistically significant POPout reductions averaging ~50%, including complete elimination of POPout for cholesterol (100%), a 90% reduction for albumin, a 60% reduction for plateletcrit, and a 38% reduction for RDW.[1]
  • Disproportionate Risk Stratification Gain: Incorporating rare variants increased population-wide $R^2$ by a modest 11.6%, but increased tail-specific disease odds ratios by 71.7% on average.[1]
  • Pathogenic Gene Overlaps: Significant rare single-variant and burden signals overlapped established ClinVar pathogenic disease genes, including ACAN, HBB, JAK2, LDLR, MC4R, TFR2, and CHEK2.[1]

SLiM-4 Forward Simulations & Selection Dynamics

Forward-in-time population genetic simulations using SLiM-4 under a gamma distribution of mutational effect sizes confirmed that stabilizing selection concentrates large-effect rare variants in phenotypic extremes:[1] - Under neutrality, ~1% of extreme-tail individuals carried large-effect rare variants.[1] - Under strong stabilizing selection, 28% of extreme-tail individuals carried large-effect rare alleles, reproducing empirical non-linear POPout curves.[1] - Simulations with Gaussian mutational distributions failed to produce substantial POPout effects, demonstrating that heavy-tailed mutational effect distributions (gamma) are necessary to explain complex trait tail architecture.[1]

Contemporary Fitness & Fecundity Modeling

Extending regression models of lifetime reproductive success (fecundity) using linear ($\beta$) and quadratic ($\gamma$) selection terms classified trait architectures:[1] - Positive Directional Selection ($\beta > 0$): Displayed significantly larger lower-tail POPout effects ($P = 2.9 \times 10^{-6}$).[1] - Negative Directional Selection ($\beta < 0$): Displayed significantly larger upper-tail POPout effects ($P = 0.014$).[1] - Non-Directional Stabilizing Selection ($\gamma < 0$): Displayed symmetric POPout departures across lower and upper tails ($P = 0.52$).[1] - Fitness Correlations: POPout magnitudes correlated negatively with number of children fathered ($P = 5.9 \times 10^{-4}$) and maternal live births ($P = 0.048$), and positively with paternal age ($P = 0.018$).[1]

Synthesis: This work complements Burden Heritability Regression (BHR) and Rare-Variant Genetic Architecture of Depression. While BHR shows that rare coding variant heritability is concentrated in constrained genes population-wide, Souaiaia et al. prove that rare variants disproportionately drive phenotypic extremes, explaining why rare-variant integration frameworks like RICE achieve their largest risk stratification gains in extreme disease tails.

See Also

Citations

[1] Souaiaia, T., Wu, H.M., Ori, A.P.S., Choi, S.W., Hoggart, C.J., & O'Reilly, P.F. (2026). Distinct genetic architecture in the tails of complex traits. Nature, 655, 676–684. Supports: POPout and STANDout formulation; empirical results across 74 traits; RDW and Hb disease emulation failure; multi-ancestry and sibling replications; WGS/WES rare variant integration and 71.7% OR increase; SLiM-4 simulations; selection and fecundity modeling. Location: Main text (Figs. 1–5); Extended Data Figs. 1–7; Supplementary Tables 1–3. Source paper: s41586-026-10516-5.md