henhaobio
Sign in

BioinformaticsGeneticsPeer-reviewed journal article

LDpred2: better, faster, stronger

Florian PrivéiD, Julyan ArbeliD, Bjarni Jóhann VilhjálmssoniD

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

Abstract

Polygenic scores have become a central tool in human genetics research. LDpred is a popular method for deriving polygenic scores based on summary statistics and a matrix of correlation between genetic variants. However, LDpred has limitations that may reduce its predictive performance. Here we present LDpred2, a new version of LDpred that addresses these issues. We also provide two new options in LDpred2: a "sparse" option that can learn effects that are exactly 0, and an "auto" option that directly learns the two LDpred parameters from data. We benchmark predictive performance of LDpred2 against the previous version on simulated and real data, demonstrating substantial improvements in robustness and predictive accuracy compared to LDpred1. We then show that LDpred2 also outperforms other polygenic score methods recently developed, with a mean AUC over the 8 real traits analyzed here of 65.1%, compared to 63.8% for lassosum, 62.9% for PRS-CS and 61.5% for SBayesR. Note that, in contrast to what was recommended in the first version of this paper, we now recommend to run LDpred2 genome-wide instead of per chromosome. LDpred2 is implemented in R package bigsnpr.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 3 (1 Oct 2020) · CC BY 4.0

LDpred2 reworks LDpred with a 3 cM LD window plus sparse and auto options, reaching mean AUC 65.1% over 8 real traits versus 63.8% for lassosum.

The non-infinitesimal Gibbs sampler in LDpred1 breaks down under model misspecification when GWAS samples are large, and it is unstable in long-range LD regions such as HLA, so analysts often dropped regions rich in disease variants. LDpred2 fixes these instabilities, searches a much larger hyper-parameter grid in similar time, and adds a sparse mode and an auto mode that needs no validation set.

Key results

  1. ShownOn eight real case-control traits, LDpred2 had the highest AUC of all the summary-statistic methods compared. SBayesR did especially badly on the autoimmune diseases.

    Mean AUC over 8 traits: 65.1% for LDpred2 vs 63.8% for lassosum, 62.9% for PRS-CS and 61.5% for SBayesR. For T1D, SBayesR AUC was 58.1% vs 78.4% for LDpred2-grid. SBayesR matched LDpred2 on BRCA, MDD, PRCA and T2D. · Abstract; Real data

  2. ShownRun genome-wide on real traits, LDpred2-inf and LDpred2-grid predicted clearly better than their LDpred1 counterparts for BRCA, CAD, MDD, PRCA, T1D and T2D. Both versions performed about the same for Asthma and RA.

    AUC of LDpred2-grid vs LDpred1-grid: BRCA 65.5% vs 58.9%; T1D 78.4% vs 57.4%. Asthma and RA showed similar AUC across versions. · Real data

  3. ShownIn simulations using UK Biobank genotypes, LDpred2 beat LDpred1 when there were few causal variants, when causal variants sat in the HLA region, and when the GWAS was large. In the large-GWAS case, LDpred1 accuracy levelled off and then dropped.

    With 300 causal variants, mean AUC was 81.7% for LDpred2 vs 73.6% for LDpred1. In scenario all_40_3000 at GWAS N 120,000: 74.6% vs 70.7%. At 300,000, LDpred1 fell to 70.4% while LDpred2 reached 79.1% (maximum achievable 82.5%). LDpred1 performed poorly in the HLA scenarios. · Simulations

  4. ShownBoth new options keep accuracy. The sparse grid model sets many effects exactly to zero and still matches the non-sparse model. LDpred2-auto, which needs no validation set, comes close to LDpred2-grid on most traits.

    Sparse and non-sparse grid models performed equally in simulations and real data; effect sparsity ranged from 19.3% (RA) to 54.4% (Asthma). AUC of LDpred2-auto vs LDpred2-grid: CAD 61.8% vs 63.6%; Asthma 58.4% vs 59.3%. Auto was slightly worse in the HLA simulations. · Simulations; Real data

  5. ShownFitting LDpred2 genome-wide usually did at least as well as fitting each chromosome separately, with the biggest gains when the GWAS was small. This reverses the advice in the earlier preprint version.

    Simulation at GWAS N 20,000, genome-wide vs per chromosome: inf 57.7% vs 57.2%, grid 59.6% vs 58.6%, auto 59.3% vs 57.3%. Real data, LDpred2-grid: CAD 63.6% vs 61.5%; PRCA 70.2% vs 68.2%. The one exception was LDpred2-inf on T1D. · Running LDpred2 per chromosome or genome-wide; Abstract

Methods, data and reproducibility

LDpred2 turns GWAS marginal effects and an LD matrix (3 cM genetic-distance window) into joint effects using a Gibbs sampler, under a model in which a fraction p of variants are causal. Grid mode tunes p, h2 and sparsity (126 combinations) on validation data. Auto mode samples p from a Beta posterior and estimates h2 inside the sampler, over 30 chains.

  • UK Biobank imputed genotype dosages, restricted to unrelated, quality-controlled, genetically homogeneous individuals and the HapMap3 variants used by PRS-CS · human (UK Biobank) · n = 362,320 individuals; 1,117,493 variants
  • Simulated binary phenotypes (liability threshold, prevalence 15%) on UK Biobank genotypes; causal variants genome-wide, in HLA or mixed; GWAS N 10,000 to 300,000 tested only with 3000 causal · human (simulated on UK Biobank genotypes) · n = seven scenarios, 10 replicates each; 52,320 test individuals
  • Published external GWAS summary statistics for eight case-control traits (BRCA, CAD, MDD, PRCA, T1D, T2D, Asthma, RA), with scores evaluated in UK Biobank · human · n = 10,000 validation; 352,320 test individuals
  • European samples from 1000 Genomes, used as the PRS-CS LD reference and for allele-frequency QC of the released LD reference · human (1000 Genomes) · n = 503 individuals
  • Released European-ancestry LD reference computed from UK Biobank after allele-frequency and MAF filtering · human (UK Biobank) · n = 1,054,330 variants; 362,320 individuals · 10.6084/m9.figshare.13034123
Methods & tools
LDpred2 (R package bigsnpr), LDpred1 (LDpred v1.0.0), bigsparser (on-disk sparse matrix format), bigstatsr, Clumping and Thresholding (C+T), Stacked C+T (SCT), lassosum / lassosum-auto, PRS-CS / PRS-CS-auto, SBayesR, constrained LD score regression, future framework (R parallelization), tidyverse
Software & environment
R packages bigsnpr, bigstatsr and bigsparser (parallel C++ backend with a memory-mapped on-disk sparse LD matrix); LDpred v1.0.0 as the baseline; future framework for scheduling on an HPC cluster. No pinned bigsnpr release or container is given.
Assay conditions & compute
Under three hours for all LDpred2 models on one 100,000-variant chromosome (8 cores); under 36 hours genome-wide on 1.1 million HapMap3 variants; 11 minutes to compute LD for 10,000 individuals and 100,000 variants (16 cores); a 1M-variant LD matrix needs over 32 GB of RAM.
Reproducibility gaps
The LD reference is released for European ancestry only. Individual-level UK Biobank genotypes and phenotypes are not shared. Users are pointed to the newest bigsnpr on GitHub, with no pinned release.
Limitations (6)
  • LDpred2-auto needs quality control of the summary statistics and still slightly underperforms for CAD and Asthma. The authors suggest that meta-analyses of 48 and 66 small studies, some non-European or multiancestry, may break its assumptions. · stated by the authors
  • No Bayesian estimator of h2 could be derived, so auto mode uses a plug-in estimate. For CAD, its h2 disagreed with LD score regression and its p (about 0.001) differed from the best grid p (about 0.01), so how meaningful its estimates are still needs study. · stated by the authors
  • Genome-wide fitting is memory-heavy (over 32 GB for a 1M-variant sparse LD matrix, 128 GB for 2M). It also assumes one architecture for all chromosomes, which may misspecify traits whose heritability concentrates on chromosome 6, such as autoimmune diseases. · stated by the authors
  • The work uses and recommends HapMap3 variants only. Whether other variant sets, for example ones enriched for significant variants, would help LDpred2 and other PGS methods is left for future research. · stated by the authors
  • Every method except PRS-CS used the 10,000-person UK Biobank validation set as its LD reference, while PRS-CS used its own 503-sample 1000 Genomes reference. Part of the gap to PRS-CS may therefore come from the choice of LD reference rather than the model. · noted in this summary
  • The real-data benchmark covers only eight binary traits, scored by AUC, in one genetically homogeneous UK Biobank subset. Continuous traits and transfer to other ancestries were not tested. · noted in this summary

Read this if you build or apply polygenic scores from summary statistics and want practical guidance on LD window size, hyper-parameter grids, summary-statistic QC, or choosing between LDpred2-grid and LDpred2-auto.

Citations
726
FWCI
39.40
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
THISVilhjalmsso '15Privé '23Pain '21Mak '17Weissbrod '22Chen '24
77% semantic matchBioinformatics
Bjarni Vilhjalmsson, Jian Yang, Hilary Kiyo Finucane+30 more·The American Journal of Human Genetics·Oct 1, 2015