How GigaPRS works

Also built in

A heritability-informed prior. Each SNP's prior variance is proportional to its expected heritability under the BLD model (86 functional annotations, with the MAF dependence estimated from the data), rather than assuming every SNP contributes equally. Quality control and imputation of summary statistics are done automatically. Fine-mapping uses the same machinery, giving posterior inclusion probabilities and credible sets.

Pipeline

Inputs: summary statistics from one or more GWAS (each labelled focal, primary or secondary and matched to an LD matrix), LD matrices for one or more ancestries, and genotypes of the target individuals. Step 1 runs the six operations below once per focal ancestry; Step 2 runs once.

Step 1 · estimate SNP effect sizes (per focal ancestry)

1. O1a Quality control and imputation. Checks that each GWAS's allele frequencies match its assigned LD matrix best (catching mis-assigned ancestries), drops SNPs whose frequencies differ by more than 0.2, and imputes missing summary statistics within LD blocks.
2. O1b Scale the summary statistics. Estimates the inflation \gamma of each GWAS with SumHer and sets \lambda = 1/\sqrt{\hat\gamma}.
3. O1c Per-SNP heritabilities. Fits the BLD heritability model to the focal GWAS to obtain \hat h^2_j for every SNP.
4. O1d Combine the GWAS. Selects informative GWAS by stepwise selection, then builds the combined summary statistics and combined LD matrix.
5. O1e Prior parameters. Chooses the mixture proportions by pseudo cross-validation (no individual-level data needed).
6. O1f Effect sizes. Variational Bayes, one LD block at a time.

Step 2 · compute scores

1. O2a Ancestry-specific scores. One PRS per ancestry, P^d = \sum_j X_j \hat\beta^d_j.
2. O2b Ancestry weighting. Infers each individual's ancestry proportions and combines the scores.

Operations in red are new in GigaPRS

Statistical model

Prior. By default each SNP effect follows a BayesR-type four-component mixture whose scale is set by the SNP's expected heritability:

The components hold SNPs with zero, small, medium and large effects. Setting \pi_1=\pi_2=\pi_3=0 gives ridge regression; \pi_2=\pi_3=0 gives an LDpred-type prior.

Heritability model. Expected per-SNP heritability is

where B_k are SNP annotations, f_j is allele frequency, and \tau_k, \alpha are estimated with SumHer. The default BLD model uses the 86 non-MAF annotations of Baseline LD v2.2 and estimates \alpha; most other PRS tools assume a uniform model, \mathbb E[h^2_j]=\tau_1.

Likelihood and inference. GigaPRS works as if Y\sim\mathcal N(X\beta, I), but everything depends on X and Y only through X_j^TY\approx n_j\rho_j and X_j^TX_k\approx n_jR_{jk}, so only summary statistics and an LD matrix are needed. Variational Bayes approximates the posterior by independent per-SNP factors, updating each SNP from its conditional posterior given the current means of the others. Because the LD matrix is block-diagonal (591 blocks), each block is solved separately, which keeps memory low. Iteration stops when the explained sum of squares changes by less than 10^{-5}\,\bar n_B. An MCMC solver (Gibbs sampling, 4 chains × 250 iterations) is available and can help with in-sample LD.

Combining GWAS of different traits and ancestries

For each SNP, GigaPRS forms a weighted average of the standardised effect estimates \rho^c_j = Z^c_j/\sqrt{n^c_j} from the C GWAS:

As in MTAG, the weights come from two C\times C matrices: \Omega (covariance of the genetic contributions, built from SNP heritabilities and coheritabilities) and \Sigma (covariance of the estimation errors, built from sample sizes and bivariate intercepts). Both are estimated with SumHer. GigaPRS extends MTAG in two ways:

• Multiple ancestries. GWAS of different ancestries are allowed; their error covariance is set to zero (no sample overlap), and each GWAS is used with its own ancestry's LD.
• Stepwise selection. Starting from the primary GWAS, GigaPRS repeatedly adds the secondary GWAS that most increases the effective sample size, considering only traits with |genetic correlation| ≥ 0.5 and stopping when the gain falls below 5% of the focal sample size.

A combined LD matrix, the matching weighted average of the ancestry-specific LD matrices, describes the covariance of the combined statistics. Effect sizes are then estimated exactly as for a single GWAS, which is why a 55-GWAS, 3-ancestry analysis costs little more than three single-GWAS runs.

Example (manuscript Fig. 1b): for African type 2 diabetes effect sizes, the four primary GWAS raise the effective sample size from 116k to 277k; adding glycated haemoglobin and HDL cholesterol brings it to 333k, after which no secondary trait adds 5%.

Ancestry-aware scores

Step 1 produces one set of effect sizes per focal ancestry, stored together with each ancestry's allele frequencies in a single .combined.effects file. In Step 2, GigaPRS regresses each target individual's genotypes on the ancestry-specific allele frequencies to estimate ancestry proportions, then combines the scores. For example, an individual inferred to be 50% African and 50% European receives

This uses global rather than local ancestry, so it needs no reference panel or phasing and runs in seconds. Local-ancestry methods can, in principle, do better for admixed individuals, but cost on the order of 100 CPU hours per 100 individuals.

Performance

Speed and memory (single CPU, 1.2M HapMap3 SNPs, one GWAS)

With three GWAS of different ancestries, GigaPRS takes about 20 minutes; PRS-CSx and JointPRS each take over 25 hours.

Accuracy from a single GWAS

Mean PRS accuracy (correlation with residual phenotype) relative to GigaPRS, across 18 quantitative and 12 binary traits.

Across all 90 trait × scenario analyses, GigaPRS was most accurate 47 times and SBayesRC 34 times; LDpred, PRS-CSx and JointPRS 6, 2 and 1 times.

Combining GWAS

• Multiple ancestries (European + African + East Asian GWAS, 18 traits): GigaPRS was best in 67 of 90 analyses, with PRS 10% and 12% more accurate than PRS-CSx for Caribbean and admixed individuals, and 8–14% more accurate than single-ancestry PRS for Caribbean, Indian and admixed individuals.
• Multiple traits (51 UK Biobank secondary traits): average AUC rose from 62.6% to 63.7% (P < 10⁻⁶) for the traits where secondary GWAS were selected, close to a tuned linear combination (63.9%) but with no tuning data. The largest gains were for obesity (via BMI) and depression (via neuroticism and age at first birth).

Fine-mapping

Run in fine-mapping mode, GigaPRS fits a SuSiE-type model (up to ten causal SNPs per window) to the combined summary statistics, with each SNP's prior inclusion probability proportional to \hat h^2_j. It reports posterior inclusion probabilities and 95% credible sets. In simulations its calibration and power were comparable to SuSiE, GWFM and SuSiEx, and multi-ancestry fine-mapping gave higher power and smaller credible sets when architectures were shared.

Feeding fine-mapping results back into the heritability model closes the loop: adding 18 PIP vectors to the BLD model (the BLD-MVP model) significantly improved model fit (P < 10⁻¹⁶) and raised PRS accuracy by 3.9% for FinnGen and 19% for Taiwan PMI GWAS.

Further detail