Bayesian polygenic scoring
How PRS-CS, LDpred2, and SBayesR turn GWAS summary statistics and an LD reference panel into a calibrated polygenic score, with convergence diagnostics that flag unreliable scores.
Classical polygenic scoring takes the marginal effect-size estimates from a GWAS and adds them up across the SNPs that pass a p-value threshold. The result is biased because GWAS effect sizes are inflated for SNPs that happened to be tagged by linkage disequilibrium with a true causal variant elsewhere. Bayesian polygenic scoring fixes this by modeling the joint distribution of true effect sizes given the observed marginal estimates and a linkage-disequilibrium reference panel.
PRS-CS, briefly
PRS-CS places a continuous shrinkage prior on the true effect sizes and uses a Gibbs sampler to draw posterior samples. The shrinkage parameter, called phi, controls how aggressively small estimated effects are pulled towards zero. PRS-CS-auto learns phi from the data using a half-Cauchy hyperprior, while a fixed-phi mode is available for cohort-specific calibration. Each Gibbs iteration also samples the residual variance, allowing the prior to adapt to the trait's underlying polygenicity.
LDpred2, briefly
LDpred2 places a spike-and-slab prior on each SNP's effect size: with probability pi the SNP is causal and drawn from a Gaussian distribution, and with probability 1-pi the effect is exactly zero. Pi is a hyperparameter that LDpred2 either fixes (infinitesimal mode), tunes through a grid search, or learns automatically (LDpred2-auto). The model is solved through Gibbs sampling against the LD reference panel.
SBayesR, briefly
SBayesR generalises the spike-and-slab to a four-component Gaussian mixture, allowing the model to capture a mix of small, medium, and large effects within the same trait. The four mixture proportions follow a Dirichlet hyperprior, and the residual variance follows an inverse-Gamma. Both are sampled jointly with the SNP effects through a Gibbs procedure.
Why we run multiple methods
No single method dominates across all traits. PRS-CS tends to perform best on highly polygenic traits with thousands of small-effect contributors. LDpred2 performs better on traits with a small number of large-effect variants. SBayesR is the most flexible but the slowest to converge and the most sensitive to LD reference quality. Lassosum is fast and reasonable but less calibrated than the Bayesian methods. Running all of them in parallel and reporting both the per-method results and the ensemble lets the user see when methods agree (high confidence) and when they disagree (lower confidence).
LD reference panels
All Bayesian methods need a linkage-disequilibrium reference panel that matches the user's ancestry. Haeckel ships LD panels for all five 1000 Genomes superpopulations (EUR, AFR, EAS, SAS, AMR), computed from the 1000 Genomes Phase 3 VCF using a shrinkage approach with a conservative mixing parameter chosen to avoid LD-mismatch amplification when applied to admixed users. For users whose ancestry is not represented cleanly in any single panel, the pipeline performs cross-ancestry calibration on the resulting PRS rather than picking a single panel and pretending the rest does not exist.
- Ge T et al. (2019). Polygenic prediction via Bayesian regression and continuous shrinkage priors. Nature Communications.
- Privé F et al. (2020). LDpred2: better, faster, stronger. Bioinformatics.
- Lloyd-Jones LR et al. (2019). Improved polygenic prediction by Bayesian multiple regression on summary statistics. Nature Communications.
- Gelman A, Rubin DB (1992). Inference from iterative simulation using multiple sequences. Statistical Science.
Explain this article in the context of my own genome and tell me what is most relevant for me.