henhaobio
Sign in

bioRxivGeneticsPreprint

Heterogeneity in polygenic scores for common human traits

Erin Bakshis WareiD, Lauren SchmitziD, Jessica D. FauliD, Arianna M. GardiD, Colter M. MitchelliD, Jennifer A. SmithiD, Wei ZhaoiD, David R. WeiriD, Sharon LR Kardia

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

Abstract

Introduction Introduction Results Discussion Materials and Methods Conflict of Interest References Genome-wide association studies (GWAS) on a wide range of important human traits have identified hundreds of variants in multi-cohort meta-analyses that have highly significant associations (p<5 x10-8) that replicate across studies. However, the identified genetic variants, or single nucleotide polymorphisms (SNPs), generally explain a very small fraction of variability in the trait of interest. Moreover, for most behavioral and health scientists, the use of such massive and complex genetic data is unwieldy. In response to the desire to harness genome-wide information, use of polygenic scores (PGSs), also known as \"genetic risk scores\", has grown rapidly in the social, behavioral, and health sciences rec ...

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 1 (5 Feb 2017) · CC BY-NC-ND 4.0

Polygenic scores built 434 ways in Health and Retirement Study (HRS) participants varied widely in variance explained and inter-score correlation; European-derived weights underperformed in non-Hispanic Blacks.

Polygenic scores (PGSs) are increasingly used in social and health research, but construction choices (genotyped vs imputed SNPs, genome-wide association study (GWAS) p-value thresholds, linkage disequilibrium (LD) trimming) are seldom reported and their combined effects were largely unknown. Earlier evaluations focused on predictive power while varying a subset of choices; this work varies them jointly and also examines score correlations and ancestry transferability.

Key results

  1. ShownWithin one cohort and trait, construction choices changed whether a PGS significantly predicted the phenotype and possibly what genetic signal it reflected: scores at one p-value threshold could be almost uncorrelated, and many non-Hispanic White scores were non-significant.

    Correlation between two PGSs at the same p-value threshold (pT) was as low as r=0.02; 205 out of 1088 (18.8%) non-Hispanic White PGSs were not significantly associated with their trait (α = 0.05); approaches differed significantly in semi-partial R^2 (bootstrapped 95% CIs); most major depressive disorder (MDD) scores were non-significant in both groups. · Discussion; LD trimming and LD trimming options

  2. ShownScores weighted by European-ancestry GWAS predicted traits far less often and less strongly in non-Hispanic Black than in non-Hispanic White participants, although correlation patterns among scores were similar in the two groups.

    Significant (p ≤ 0.05) in 81.2% (883 out of 1088) of non-Hispanic White vs 29.4% (320 out of 1088) of non-Hispanic Black PGSs; at α = 0.001, 79.9% (869) vs 5.2% (57); non-Hispanic Black scores explained about half as much variance; most body mass index (BMI), educational attainment (EA 2013, EA 2016) and height scores did significantly better in non-Hispanic Whites. · Cross-race/ethnicity comparison; Discussion

  3. ShownImputed SNPs never significantly outperformed directly genotyped SNPs, and the few significant differences favored genotyped scores; yet genotyped scores agreed less with one another across LD methods. HRS adopted genotyped SNPs, no LD trimming and pT=1 for its released scores.

    Per trait, 108 genotyped-imputed pairs; no imputed R^2 significantly exceeded its genotyped pair; without LD trimming, 95% CIs separated in five pairs, all EA 2016 in non-Hispanic Whites; genotyped won 58.7% (317/540) of clumping pairs, imputed 52.9% (226/480) of pruning pairs; non-Hispanic White height, pT=0.001: median inter-method r 0.83 genotyped vs 0.92 imputed. · Directly genotyped SNPs versus imputed SNPs; Discussion

  4. ShownThe best p-value threshold depended on the trait: genome-wide significant top-SNP scores explained the least or near-least variance and correlated poorly with other scores; broader thresholds often helped BMI, education and depression, but height peaked at a stringent threshold.

    Top-SNP (p<5x10^-8) scores gave the lowest or near-lowest R^2 (MDD had no such SNPs); for BMI, EA 2013, EA 2016 and MDD, genotyped scores at pT=0.001 or 0.01 often trailed pT=0.05 to 1.0; best genotyped non-Hispanic White height score explained 10.4% (95% CI 9.2, 11.6) via clumping (LD R^2 0.1, 500 kb, pT=0.001), a setting among the weakest for other traits. · P-value threshold; Directly genotyped SNPs versus imputed SNPs; Figure legends

  5. SuggestedNo LD trimming method or parameter set consistently maximized variance explained, and LD method mattered more than genotyped vs imputed input; stringent-threshold scores agreed less across methods, while broad thresholds improved consistency for imputed but not genotyped scores.

    For BMI, including more SNPs tended to raise R^2 (height differed); in non-Hispanic Whites, clumping at LD R^2=0.1 dipped for BMI, EA 2013 and EA 2016 but not height; pT=0.001 and 0.01 scores had lower median Pearson r than pT=0.05 to 1.0, and a 109-score BMI correlation map showed pT=0.001 scores weakly correlated with other thresholds (checkerboard pattern). · LD trimming and LD trimming options; P-value threshold; Figure legends

Methods, data and reproducibility

Polygenic scores for height, BMI, educational attainment (2013/2016 GWAS) and depression were built in non-Hispanic White and Black HRS samples via a modified PRSice/PLINK R wrapper, crossing seven p-value thresholds, genotyped vs imputed SNPs and no LD trimming, clumping or pruning; bootstrapped semi-partial R^2 and inter-score correlations were compared.

  • Health and Retirement Study (HRS) non-Hispanic White participants with genetic samples collected in 2006, 2008 and 2010 · human · n = 9,991 non-Hispanic Whites
  • HRS non-Hispanic Black participants from the same genetic sample collection years · human · n = 2,279 non-Hispanic Blacks
  • Published European-ancestry-based GWAS meta-analysis results for height and BMI used as SNP weights; for all traits HRS was absent or mathematically removed · human
  • Educational attainment GWAS meta-analyses published in 2013 (EA 2013) and 2016 (EA 2016), used as two alternative SNP weight sets · human
  • Psychiatric Genomics Consortium MDD GWAS results, already heavily LD-trimmed and containing no genome-wide significant SNPs · human
Methods & tools
PRSice (basis of the heavily modified R wrapper), PLINK, R, LD clumping (LD R^2 thresholds by region size in kb), LD pruning (LD R^2, window size in kb, sliding increment in SNPs), Genotype imputation (imputed vs directly genotyped SNP sets), Semi-partial R^2 with bootstrapped 95% CIs (1,000 repetitions), Pearson correlation among participant-level PGSs, Illumina HumanOmni2.5-4v1 and -8v1 BeadChips
Reagents & environment
Heavily modified R wrapper built on PRSice and PLINK; no software versions, containers or package lists are reported.
Reproducibility gaps
No code or data links are given (including for the modified PRSice/PLINK R wrapper or HRS genotype access); detailed methods are deferred to supplementary material; full consortium GWAS results needed by LDPred were unavailable.
Limitations (6)
  • The explored decision space was necessarily incomplete: PGS packages needing detailed consortium GWAS meta-analysis results, notably LDPred, could not be fully evaluated because such results were not available. · stated by the authors
  • The reported advantage of genotyped over imputed SNPs may depend on the HRS arrays (Illumina HumanOmni2.5-4v1 and -8v1 BeadChips) and might not generalize to studies genotyped on other platforms. · stated by the authors
  • Scores weighted by non-Hispanic Black GWAS could not be tested in either HRS sample because large, replicated GWAS of these traits in that group were lacking, so the effect of applying European-ancestry effect sizes to other groups remains incompletely examined. · stated by the authors
  • Most MDD scores were non-significant predictors in both ancestry groups, so the depression analyses could not inform how LD trimming affects score performance. · stated by the authors
  • Genotyped-vs-imputed differences were judged by non-overlap of bootstrapped 95% CIs, a conservative informal test run over many comparisons with no multiplicity correction described; with only five pairs separated, the Discussion's claim that genotyped data work better is weakly supported. · noted in this summary
  • All analyses come from one cohort, and top-performing approaches (e.g., the 10.4% height score) were identified in the same sample that produced their R^2, with no independent replication, so best-case R^2 values and approach rankings may be optimistic. · noted in this summary

Read this if you construct or analyze polygenic scores in cohort data and want evidence on how SNP source, p-value threshold, LD trimming and ancestry shift score performance and agreement.

Citations
123

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
THISBennett '21Lambert '21Chen '25Carvalho '22Zhao '19Socrates '17
66% semantic matchGenomics
Declan Bennett, Dónal O’Shea, John P. Ferguson+2 more·Scientific Reports·Oct 1, 2021· 17 citations