henhaobio
Sign in

Cell GenomicsGenetic & Genomic MedicinePeer-reviewed journal article

Integrative polygenic risk score improves the prediction accuracy of complex traits and diseases

Buu Minh Thanh TruongiD, Leland E. HulliD, Yunfeng RuaniD, Qin HuangiD, Whitney HornsbyiD, Hilary C. MartiniD, David A. van HeeliD, Ying WangiD, Alicia R. MartiniD, Sang LeeiD, Pradeep NatarajaniD

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

Abstract

Polygenic risk scores (PRS) are an emerging tool to predict the clinical phenotypes and outcomes of individuals. Validation and transferability of existing PRS across independent datasets and diverse ancestries are limited, which hinders the practical utility and exacerbates health disparities. We propose PRSmix, a framework that evaluates and leverages the PRS corpus of a target trait to improve prediction accuracy, and PRSmix+, which incorporates genetically correlated traits to better capture the human genetic architecture. We applied PRSmix to 47 and 32 diseases/traits in European and South Asian ancestries, respectively. PRSmix demonstrated a mean prediction accuracy improvement of 1.20-fold (95% CI: [1.10; 1.3]; P-value = 9.17 x 10-5) and 1.19-fold (95% CI: [1.11; 1.27]; P-value = 1.92 x 10-6), and PRSmix+ improved the prediction accuracy by 1.72-fold (95% CI: [1.40; 2.04]; P-value = 7.58 x 10-6) and 1.42-fold (95% CI: [1.25; 1.59]; P-value = 8.01 x 10-7) in European and South Asian ancestries, respectively. Compared to the previously established cross-trait-combination method with scores from pre-defined correlated traits, we demonstrated that our method can improve prediction accuracy for coronary artery disease up to 3.27-fold (95% CI: [2.1; 4.44]; P-value after FDR correction = 2.6 x 10-4). Our method provides a comprehensive framework to benchmark and leverage the combined power of PRS for maximal performance in a desired target population.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 3 (23 Mar 2023) · CC BY 4.0

Cross-trait Elastic Net mixing of PGS Catalog scores (PRSmix+) lifted mean PRS accuracy 1.72-fold in Europeans and 1.42-fold in South Asians over the best single score.

Many published PRS exist for the same trait, yet many are seldom validated beyond their training cohorts and transfer poorly across ancestries, and no standard procedure existed to pool this corpus for a chosen target population. PRSmix supplies such a procedure, harmonizing PGS Catalog scores, keeping well-powered ones and learning cohort-specific mixing weights, with an option to borrow pleiotropic signal from correlated traits.

Key results

  1. ShownTrait-specific mixing (PRSmix) beat the best single PGS Catalog score in both a European and a South Asian cohort. The best single score was chosen in training and assessed on held-out test data. Gains were largest for traits whose single scores were weak.

    Mean fold-ratio of R^2 vs best catalog score: 1.20-fold (95% CI [1.10; 1.3]; P = 9.17 × 10^−5) across 47 European-ancestry traits (All of Us) and 1.19-fold (95% CI [1.11; 1.27]; P = 1.92 × 10^−6) across 32 South Asian traits (Genes & Health). Top gains: rheumatoid arthritis 3.36-fold (European), CAD 2.32-fold (South Asian). · Abstract; Overview of methods; Combining trait-specific PRS improves prediction…

  2. ShownAdding well-powered scores from other traits (PRSmix+) uses pleiotropy. In each ancestry it improved on both the best catalog score and trait-specific PRSmix.

    Vs best catalog score: 1.72-fold (95% CI [1.40; 2.04]; P = 7.58 × 10^−6) in European and 1.42-fold (95% CI [1.25; 1.59]; P = 8.01 × 10^−7) in South Asian ancestry. Vs PRSmix: 1.46-fold (95% CI [1.17; 1.75]; P = 0.002) and 1.19-fold (95% CI [1.07; 1.32]; P = 0.001). Candidate scores came from 2600 catalog PRS. · Abstract; Cross-trait combination further improved PRS…; Power and variance of PRS accuracy

  3. ShownFor coronary artery disease (CAD), adding PRSmix+ to a clinical risk-factor model gave larger net reclassification gains than PRSmix or the best catalog CAD score in both ancestries. PRSmix+ also had the largest incremental AUC and a higher top-decile odds ratio than the best catalog score.

    Continuous NRI: European 35% (95% CI [26%; 45%]) vs 30% (PRSmix) vs 28% (best catalog); South Asian 27% (95% CI [16%; 38%]) vs 15% vs 7%. Incremental AUC 0.02 vs 0.016 vs 0.012 (European) and 0.008 vs 0.006 vs 0.003 (South Asian). Top-decile OR, PRSmix+ vs best catalog: 2.53 vs 1.67 (European), 2.34 vs 1.73 (South Asian). · Clinical utility for coronary artery disease; Assessment of clinical utility

  4. ShownThe benchmark covered 5 traits, using LDpred2-auto scores from 26 public GWAS. PRSmix+ beat wMT-SBLUP and an Elastic Net of LDpred2 scores. Pleiotropy added little for height, but every cross-trait method beat trait-specific mixing for type 2 diabetes (T2D).

    PRSmix+ (pre-defined traits plus catalog) vs wMT-SBLUP: from 1.12-fold (95% CI [1.02; 1.21]; P = 0.016) for T2D to 3.27-fold (95% CI [2.1; 4.44]; FDR-corrected P = 2.6 × 10^−4) for CAD. PRSmix+ (LDpred2-auto plus catalog) vs Elastic Net of LDpred2 scores: best 1.6-fold (95% CI [1.31; 1.89]; P = 1.1 × 10^−4) for depression. · Abstract; Comparison with previous combination methods; wMT-SBLUP; Elastic Net for linear combination

  5. SuggestedSimulations on UK Biobank genotypes estimated how much training data the mixing weights need. Low-heritability traits need more samples but gain more. Gains level off at larger training sizes, which matches a plateau seen in real data for CAD.

    PRSmix beat the best single simulated PRS from N training = 200 (h^2 = 0.4) to 5000 (h^2 = 0.05), mostly showed no gain below 500, and saturated from 10,000. PRSmix+ gained negligibly beyond 30,000. In real cohorts, CAD PRSmix plateaued from a training size of 5000 in both ancestries. · Simulations were used to evaluate the…; Simulations; Clinical utility for coronary artery disease

Methods, data and reproducibility

PGS Catalog SNP weights are harmonized and scored in a target biobank split 80% training/20% testing. Well-powered scores (power > 0.95) enter a 5-fold cross-validated Elastic Net with age, sex and 10 PCs, and the learned mixing weights are pushed back into per-SNP effects. PRSmix uses same-trait scores; PRSmix+ adds scores from all other traits.

  • All of Us v6 whole-genome sequencing, European ancestry (predicted ancestry probability > 90%), with EHR, survey and physical-measure phenotypes for 47 traits · human · n = 48,112 individuals (from 98,600 with WGS)
  • Genes & Health British South Asian (Bangladeshi and Pakistani) imputed genotypes with EHR-derived phenotypes for 32 traits · human · n = 44,396 individuals; 8,935,207 SNPs
  • PGS Catalog (version July 18, 2022) published PRS weights used as input scores · human · n = over 2,600 PRS for 538 traits
  • UK Biobank European-ancestry genotypes used to simulate 7 traits for evaluating combination frameworks and training sample sizes · human (UK Biobank) · n = 200,000 GWAS and 7,000 test individuals; 200 replications
  • Publicly available European-ancestry GWAS summary statistics for benchmarking on 5 outcome traits (CAD, T2D, depression, height, BMI) · human · n = 26 GWAS
  • PheWAS of CAD PRS in All of Us European-ancestry participants · human · n = 1815 phecodes
  • European LD reference panel from the LDpred2 tutorial, used for LDpred2-auto with HapMap3 variants that overlap All of Us WGS SNPs · human · n = 1,138,726 variants
Methods & tools
PRSmix / PRSmix+ R package, glmnet R package v4.1 (Elastic Net), LDpred2-auto, wMT-SBLUP, LD score regression, PGS Catalog, Continuous (category-free) net reclassification improvement, Partial R^2 and liability R^2 with analytical power and variance, PheWAS with phecodes, All of Us Researcher Workbench (Terra), R 4.0.0, HapMap3 variant set
Software & environment
R version 4.0.0 in Jupyter Notebook on the All of Us Researcher Workbench (built on Terra); glmnet R package version 4.1; PGS Catalog version July 18, 2022; All of Us v6 Curated Data Repository.
Reproducibility gaps
All of Us and Genes & Health individual-level data need controlled-access approval. The new score weights are said to be in the PGS Catalog, but the text gives no PGS IDs. Access to the UK Biobank data used in simulations is not described.
Limitations (6)
  • Most PGS Catalog input scores were derived in European-ancestry samples. This limits single-score power, and so combined accuracy, in South Asian and other non-European groups. · stated by the authors
  • Elastic Net assumes the outcome depends linearly on the component PRS and covariates; non-linear combinations were not explored. · stated by the authors
  • Mixing weights were learned and tested only within internal splits of each cohort, never in an independent cohort. As biobanks grow, the samples behind catalog PRS may also overlap with the test cohorts. · stated by the authors
  • Only continuous traits and binary traits with prevalence > 2% were modeled, so rare conditions are not covered; the authors point to emerging cross-trait models aimed at low-prevalence diseases. · stated by the authors
  • Gains are mean fold-ratios of R^2 across traits, which weight low-baseline traits heavily (e.g., PRSmix+ for other conditions 2.66-fold, 95% CI [1.30; 4.01], European). Traits whose best single score had power < 0.95 were dropped, so very weak-signal traits are untested. · noted in this summary
  • Reported details are inconsistent. All of Us is cited as both version 5 and v6, and the simulated heritabilities (0.05, 0.1, 0.2, 0.5) omit the h^2 = 0.4 discussed in the results. Also, four per-SD odds ratios are listed for three South Asian CAD scores. · noted in this summary

Read this if you build or deploy polygenic risk scores and want a reproducible way to combine many published PGS Catalog scores, including cross-trait ones, for a specific biobank or non-European ancestry.

Citations
97
FWCI
25.48
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
THISLerga-Jaso '25Weissbrod '22Hujoel '22Kelemen '24Kulm '20Park '22
77% semantic matchGenetic & Genomic Medicine
Jon Lerga-Jaso, Andrew Terpolovsky, Biljana Novković+6 more·Scientific Reports·May 20, 2025· 18 citations