Skip to content

Genome-Wide Fine-Mapping (GWFM)

Summary

Genome-wide fine-mapping (GWFM) jointly fits all SNPs across the genome in a single Bayesian mixture model (a "genome-wide BMM," or GBMM), rather than fine-mapping one GWAS locus at a time as SuSiE and FINEMAP do.[1] Wu et al. (2026) show that this approach, implemented via SBayesRC, out-calibrates and outperforms region-specific methods on error control, power, resolution, precision, replication rate, and trans-ancestry prediction.[1] A substantial share of the causal signal it recovers (30% of credible sets and 15.3% of explained heritability) lies outside genome-wide-significant loci entirely.[1]

Why Genome-Wide Rather Than Locus-by-Locus

Region-specific fine-mapping restricts analysis to GWAS-significant loci (typically 1–2 Mb windows around lead SNPs) and estimates the prior probability of association within each region in isolation — often conservatively, as the inverse of the number of SNPs in the region.[1] This omits signal from loci that have not reached genome-wide significance, and prevents functional annotations from being jointly learned across the whole genome.[1] GBMMs address this by fitting genome-wide SNPs simultaneously, letting SNPs that share a functional annotation "borrow" evidence from one another to jointly estimate both the genetic architecture and the annotation weights through Markov chain Monte Carlo (MCMC) sampling.[1]

Mathematical Model and Methodological Extensions

GWFM here is implemented with SBayesRC, a multicomponent Bayesian mixture model that fits all common SNPs jointly across ~2,800 approximately independent LD blocks.[1]

1. Bayesian Mixture Model Specification

The summary-statistic regression model relates the vector of marginal SNP effect estimates $\mathbf{b}$ to joint effects $\boldsymbol{\beta}$ via the LD correlation matrix $\mathbf{R}$:[1] $$\mathbf{b} = \mathbf{R}\boldsymbol{\beta} + \mathbf{e}, \quad \mathbf{e} \sim \mathcal{N}\left(0, \frac{1}{N}\mathbf{R}\right)$$

For each SNP $j$, the joint effect $\beta_j$ is drawn from a mixture distribution with variance components $\boldsymbol{\gamma} = [0, 10^{-5}, 10^{-4}, 10^{-3}, 10^{-2}]'$ scaling the total SNP-based heritability $h^2_{\text{SNP}}$:[1] $$\beta_j | \gamma_k \sim \mathcal{N}\left(0, \gamma_k h^2_{\text{SNP}}\right)$$

The mixing probability $\pi_{jk}$ for SNP $j$ belonging to component $k$ is linked to $L$ functional annotations $A_{jl}$ via a probit link function:[1] $$f(\pi_{jk}) = \sum_{l=1}^{L} A_{jl} \alpha_{kl}, \quad \alpha_{kl} \sim \mathcal{N}\left(0, \sigma_\alpha^2\right)$$ where $\alpha_{kl}$ represents the annotation effect size learned jointly across the entire genome during MCMC iterations.[1]

3. Methodological Extensions

The authors contribute three formal extensions for GBMM-based fine-mapping: - Local Credible Sets ($\alpha$-LCS): Constructed by identifying a candidate causal SNP with top Posterior Inclusion Probability ($\text{PIP}j = 1 - \pi$ than a random SNP set of identical size.[1] - }$) and grouping it with SNPs in LD ($r^2 > 0.5$) until their cumulative PIP sums to $\alpha$. Each candidate set is then filtered by its Posterior Heritability-Enrichment Probability ($\text{PEP} > 0.7$), ensuring the set explains significantly more $h^2_{\text{SNP}Global Credible Sets ($\alpha$-GCS): A single genome-wide index capturing $\alpha\%$ of all causal variants for a trait, establishing an aggregate measure of fine-mapping uncertainty at the trait level.[1] - Analytical Power Prediction: Leveraging the estimated polygenicity and effect-size variance distribution, the authors derive an analytical formula predicting the sample size $N$ required to map a target fraction of causal variants or $h^2_{\text{SNP}}$ in prospective studies.[1]

Simulation Benchmarking

Across three simulated genetic architectures in 100,000 UK Biobank individuals (~1 million HapMap3 SNPs) — sparse (1% causal, 50% $h^2$), large-effects (10 variants explain 10%, rest explain 40%), and LD-and-MAF-stratified (LDMS):[1]

  • PIP Calibration: SBayesRC's PIPs tracked the True Discovery Rate (TDR) closely across 100 PIP bins.[1] In contrast, SuSiE and FINEMAP showed severe False Discovery Rate (FDR) inflation at high PIP, while SuSiE-inf and FINEMAP-inf degraded under the non-random LDMS architecture.[1]
  • Mapping Power & Resolution: At $\alpha=0.9$, SBayesRC outperformed SuSiE-inf by up to a 194% increase in power and a 21% reduction in average LCS size.[1] Under the sparse architecture, 98% of SBayesRC PIP>0.9 SNPs were truly causal, and 99% were within 18 kb of a causal variant (a 68% reduction in distance vs. alternative methods).[1]
  • GWFM vs. GWAS-Restricted Fine-Mapping: Restricting fine-mapping to 2-Mb windows around GWAS lead SNPs ($P < 5 \times 10^{-8}$) lost 46–61% of total causal variant mapping power relative to full GWFM.[1]

Real-Data Validation and 599-Trait Application

Real-Data Replication & Trans-Ancestry Portability

  • Replication Rate: In 100,000 UK Biobank discovery samples across height, red blood cell count, and HDL cholesterol, SBayesRC achieved an 8% (vs. SuSiE-inf) to 31% (vs. FINEMAP) higher replication rate at PIP>0.9 in an independent 100,000-person sample, rising to 14–36% in a 200,000-person replication sample.[1]
  • Unbiased Effect Estimation: Regression slope of discovery vs. replication joint effect sizes for height was 0.98 for SBayesRC (closest to the unbiased value of 1.0).[1]
  • Trans-Ancestry Prediction: Using European-derived posterior effect sizes to predict phenotypes in African, East Asian, and South Asian populations across six traits, SBayesRC achieved a ~10-fold higher mean relative prediction $R^2$ compared to SuSiE-inf.[1]

Comprehensive Analysis of 599 Traits

SBayesRC was applied to 597 UK Biobank traits, schizophrenia (SCZ), and Crohn's disease (CD) using 13 million imputed SNPs and BaselineLD v2.2 annotations, with deep focus on 48 well-powered traits:[1]

  • SNP-Level Findings: 1,820 SNPs reached PIP>0.9; 1,158 were not GWAS lead SNPs, and 14.9% fell outside GWAS-significant loci.[1] 469 (25.8%) were pleiotropic (associated with multiple traits), with minor allele frequency decreasing as pleiotropy count increased — consistent with negative selection acting on pleiotropic variants.[1]
  • LCS Findings: 19,863 0.9-LCSs were identified (median size 5 SNPs); 29.8% lay outside genome-wide-significant loci.[1] These LCSs covered 0.9% of total SNPs but explained 17.7% of $h^2_{\text{SNP}}$ (2.7 percentage points, or 15.3% of the explained total, coming from sub-threshold GWAS loci).[1] Cognitive traits showed the highest heritability proportion outside GWAS loci due to extreme polygenicity.[1]
  • Sample Size Projections: Predicting required sample sizes for future GWAS/fine-mapping campaigns:
  • Blood-cell traits: 1 million participants needed to map 50% of $h^2_{\text{SNP}}$, 3 million for 80%.[1]
  • Cognitive traits: 4 million participants needed for 50% of $h^2_{\text{SNP}}$, 8 million for 80%.[1]
  • Schizophrenia: 180,000 cases needed for 20% of $h^2_{\text{SNP}}$ (~1,000 causal variants), 550,000 cases for 50%, and 1.4 million cases for 80%.[1]

Case Studies: Annotation-Driven Prioritization

Three case studies highlight how genome-wide annotation learning resolves causal variants missed by locus-specific methods:[1]

  1. FTO / Body Mass Index: SBayesRC identified a 5-SNP 0.9-LCS containing the known causal variant rs1421085 ($r^2 \ge 0.997$). SBayesRC assigned rs1421085 the highest PIP (0.47), driven by cross-species conservation annotations, whereas SuSiE-inf prioritized the non-causal GWAS lead SNP.[1] SBayesRC also identified a secondary LD-masked signal (rs76488452, PIP=0.41, primate-conserved) that was only nominally significant in standard GWAS ($P=3.6 \times 10^{-4}$) but reached $P=1.8 \times 10^{-17}$ in conditional analysis.[1]
  2. SLC39A8 / Schizophrenia: The missense variant rs13107325 (regulating dendritic spine density) was prioritized via aggregated functional annotations, along with a secondary primate-conserved signal at rs34333163 (PIP=0.82).[1]
  3. ACTR1B / Schizophrenia: SBayesRC assigned PIP=0.96 to the nonsynonymous variant rs11692435 (vs. PIP=0.19 in SuSiE-inf), which was missed by original FINEMAP analysis ($P=3 \times 10^{-8}$).[1] rs11692435 is a single-cell eQTL for ACTR1B specifically in excitatory neurons and is supported by brain PWAS and pQTL colocalization.[1]

Limitations

As stated by the authors, simulations covered three effect-size scenarios; SBayesRC assumes independence between (not within) LD blocks and requires a well-matched LD reference panel; and real-data applications were limited to European-ancestry UK Biobank data with BaselineLD v2.2 annotations.[1]

See Also

Citations

[1] Wu, Y., Zheng, Z., Thibaut, L., Lin, T., Feng, Q., Cheng, H., Yengo, L., Goddard, M.E., Wray, N.R., Visscher, P.M., Zeng, J. (2026). Genome-wide fine-mapping improves identification of causal variants. Nature Genetics, 58, 940–951. Supports: SBayesRC BMM formulation; annotation probit link equation; $\alpha$-LCS and $\alpha$-GCS definitions; power equation; PIP calibration and mapping power in simulations; 1,820 PIP>0.9 variants across 599 traits; 15.3% heritability from sub-threshold loci; FTO, SLC39A8, and ACTR1B case studies. Location: Abstract; Results ("Genome-wide Bayesian mixture model fine-mapping", "Simulation benchmarking", "Replication and trans-ancestry prediction", "Application to 599 complex traits", "Case studies"); Figures 1--6; Supplementary Notes 1--8.