Joint Bayesian modelling of molecular QTL and GWAS effects improves polygenic prediction for complex traits

The paper introduces SBayesCO, a Bayesian framework that jointly models GWAS and molecular QTL effect sizes to significantly improve polygenic prediction accuracy and SNP prioritization for complex traits compared to existing methods.

The Gist

Imagine you are trying to predict the weather.

For years, meteorologists have looked at the big picture: the temperature, the wind speed, and the humidity. In the world of genetics, this is like looking at GWAS (Genome-Wide Association Studies). Scientists scan a person's entire DNA to find tiny markers (SNPs) that seem to link to complex traits like heart disease, height, or cholesterol levels.

But here's the problem: The weather is complicated. Just knowing the wind speed doesn't tell you why a storm is forming. Similarly, knowing a DNA marker is linked to high cholesterol doesn't tell you how it happens. Is it because the marker changes how a gene is read? Is it because it changes how much protein is made?

Currently, most prediction models are like a weather app that only looks at the wind. They are okay, but they miss the deeper mechanics.

The New Tool: SBayesCO

This paper introduces a new method called SBayesCO. Think of it as upgrading that weather app to include a "mechanics engine."

Instead of just looking at the final weather (the disease or trait), SBayesCO looks at the molecular machinery in between. It asks: "Does this DNA marker change how much of a specific protein is made? Does it change how a gene is turned on or off?"

These "molecular markers" are called molQTLs (molecular Quantitative Trait Loci). They are like the sensors on the engine that tell you exactly how much fuel is being burned or how fast the pistons are moving.

The Big Idea: It's Not Just a "Yes/No" Switch

Here is the clever part.

Previous methods treated these molecular sensors like a simple light switch: "Is the gene active? Yes or No?"

• Analogy: Imagine trying to predict a car's speed by asking, "Is the gas pedal pressed?" (Yes/No). This helps a little, but it doesn't tell you if the driver is gently tapping it or flooring it.

SBayesCO treats the sensor like a volume knob.

• Analogy: It asks, "How much is the gas pedal pressed?" It looks at the exact size of the effect. If a DNA marker increases a protein by a tiny bit, it counts that. If it increases it by a huge amount, it counts that too.

By using this "volume knob" approach, the model can distinguish between a marker that is just a bystander and one that is actually driving the change.

What Did They Find?

The researchers tested this new method on 11 different human traits, mostly related to blood and the immune system (like cholesterol levels, platelet counts, and asthma).

1. Better Predictions: When they used SBayesCO, their predictions were significantly more accurate than the old methods. It's like upgrading from a 2-star weather forecast to a 5-star one.
2. The "Protein" Advantage: They found that looking at proteins (pQTLs) worked even better than looking at gene expression (eQTLs).
   • Analogy: Gene expression is like reading the recipe. Protein levels are like the actual cake coming out of the oven. Since the disease is the "eating of the cake," knowing how much cake was actually baked (protein) is a better predictor than just knowing the recipe was followed (gene expression).
3. Finding the Culprit: In complex genetic neighborhoods where many DNA markers look similar, SBayesCO was better at pinpointing the exact one causing the problem. It's like a detective who can look at a crowd of suspects and say, "It's definitely this person," whereas older methods were just pointing at the whole group.

Why Does This Matter?

1. For Patients: Better predictions mean doctors might be able to tell you sooner if you are at high risk for a disease, allowing for earlier prevention.
2. For Science: It helps us understand how diseases work. Instead of just knowing "Gene X is linked to Disease Y," we can say "Gene X makes too much Protein Z, which causes Disease Y."
3. For the Future: The authors suggest that as AI gets better at predicting how DNA affects biology, we can plug those AI predictions into this same "volume knob" system to make predictions even sharper.

The Bottom Line

This paper is about moving from guessing based on broad patterns to understanding based on precise mechanics. By listening to the "volume" of the molecular signals rather than just checking if they are "on," SBayesCO gives us a clearer, more accurate picture of our genetic future.

Technical Summary

1. Problem Statement

Polygenic prediction for complex traits remains challenging despite the identification of thousands of trait-associated variants via Genome-Wide Association Studies (GWAS). Current limitations include:

• Limited Heritability Explanation: Variants reaching genome-wide significance often explain only a small fraction of SNP-based heritability.
• Underutilization of Functional Data: While existing methods (e.g., LDpred-funct, SBayesRC) incorporate functional annotations, they typically treat these as binary indicators (e.g., "is this SNP in a promoter?"). This approach fails to distinguish between SNPs with strong versus weak regulatory effects.
• Loss of Quantitative Information: Molecular Quantitative Trait Loci (molQTLs), such as eQTLs (expression) and pQTLs (protein), provide not just evidence of regulatory activity but also quantitative effect sizes (magnitude and direction). Current polygenic models largely ignore these quantitative effect sizes, treating molQTLs merely as presence/absence annotations or aggregating them at the gene level (e.g., TWAS), which ignores intergenic variants.

2. Methodology: SBayesCO

The authors introduce SBayesCO, a novel Bayesian framework designed to jointly model GWAS and molQTL effects at the genome-wide SNP level.

• Core Concept: SBayesCO treats the complex trait and relevant molecular phenotypes (e.g., gene expression, protein abundance) as genetically correlated traits. It performs joint inference on SNP effects for both phenotypes simultaneously.
• Model Structure:
  • Genome Partitioning: The genome is divided into genic regions (cis-regions, ±1 Mb around genes) and intergenic regions.
  • Bivariate Mixture Prior:
    • Intergenic SNPs: Modeled with a standard univariate prior (similar to BayesC) to capture residual polygenic signals.
    • Genic (cis) SNPs: Modeled using a bivariate mixture prior ("Either In Either Out" - EIEO). A SNP can have:
      1. A null effect on both traits.
      2. A non-zero effect on the complex trait only.
      3. A non-zero effect on the molecular phenotype only.
      4. Non-zero effects on both (pleiotropy or mediation).
  • Quantitative Integration: Unlike previous methods that use binary annotations, SBayesCO directly incorporates quantitative molQTL effect sizes and their standard errors. It estimates the genetic covariance between the complex trait and the molecular phenotype to weight the contribution of molQTL effects to the trait effect estimation.
• Implementation:
  • Implemented in the BayesOmics suite (C++/R).
  • Supports both individual-level data and summary-level statistics (GWAS summary stats + molQTL summary stats).
  • Uses eigen-decomposition of LD matrices to handle summary statistics efficiently and avoid convergence issues.

3. Key Contributions

1. Novel Framework: Development of SBayesCO, the first method to explicitly model the joint distribution of GWAS and molQTL effect sizes using a bivariate Bayesian mixture model.
2. Quantitative vs. Binary Modeling: Demonstrates that modeling molQTLs as quantitative effect sizes yields superior performance compared to treating them as binary annotations (as done in SBayesCC/SDBayesRC).
3. Unified Approach: Provides a flexible framework that can integrate various molecular layers (eQTL, pQTL) and supports both individual and summary-level data, making it applicable to large-scale biobank resources.
4. Software Release: Open-source implementation (BayesOmics) facilitating reproducibility and broader adoption.

4. Key Results

A. Simulation Studies

• Performance: SBayesCO consistently outperformed the baseline SBayesC (GWAS only) and SBayesCC (binary annotation) across various scenarios.
• Sample Size Sensitivity: The improvement was most pronounced when GWAS sample sizes were modest (e.g., 5K), where integrating large-scale eQTL data (e.g., 50K samples) significantly boosted prediction accuracy.
• Architecture Robustness: Under realistic genetic architectures involving both causal mediation and pleiotropy, SBayesCO maintained higher predictive accuracy than baselines.

B. Real Data Application (UK Biobank)

• Dataset: 11 blood and immune-related traits (e.g., cholesterol, platelet count, asthma) using ~1M HapMap3 SNPs and European ancestry individuals ($N \approx 327,569$).
• Data Sources: eQTLs from eQTLGen Consortium and pQTLs from UKB Pharma Proteomics Project (UKB-PPP).
• Prediction Accuracy:
  • eQTL Integration: SBayesCO-eQTL improved prediction $R^2$ by an average of 2.9% relative to SBayesC and 3.3% relative to SBayesCC.
  • pQTL Integration: SBayesCO-pQTL showed even greater gains, improving $R^2$ by 3.7% relative to SBayesC and 3.6% relative to SBayesCC.
  • Trait Specifics: Improvements were consistent across diseases, blood biomarkers, and cell counts. For example, for Total Cholesterol, SBayesCO-eQTL increased standardized $R^2$ by 5.3%, whereas SBayesCC only gained 0.37%.
  • Binary vs. Quantitative: SBayesCC (binary) often showed negligible gains or even reduced accuracy for disease traits, highlighting the limitation of ignoring effect size magnitude.

C. Trans-Ancestry Portability

• South Asian (SAS): SBayesCO showed robust improvements in the South Asian population (genetically closer to Europeans), with relative $R^2$ gains up to 12.8% for dyslipidemia.
• African (AFR) & East Asian (EAS): Improvements were inconsistent or minimal, likely due to greater genetic distance and differences in LD structure, though SBayesCO remained more stable than binary annotation models.

D. SNP Prioritization and Fine-Mapping

• Functional Enrichment: SBayesCO concentrated Posterior Inclusion Probabilities (PIPs) significantly more into regulatory regions (genic-eQTL/pQTL) and depleted intergenic signals compared to baselines.
• Resolution: In locus-specific examples (e.g., SLC22A1 for cholesterol, AP2B1 for platelet count), SBayesCO successfully resolved ambiguous LD clusters, assigning PIPs near 1.0 to the true causal variant, whereas SBayesC and SBayesCC distributed probability mass across multiple correlated SNPs.

5. Significance and Implications

• Paradigm Shift: The study establishes that quantitative functional data (effect sizes) are far more valuable for polygenic prediction than binary functional annotations.
• Future Directions: The authors suggest that this framework is ideal for integrating emerging AI-based regulatory effect predictions (e.g., from DNA foundation models), which provide continuous scores rather than binary labels.
• Biological Insight: The superior performance of pQTL integration suggests that protein abundance is a more proximal mediator of complex traits than gene expression, capturing non-redundant biological signals.
• Clinical Utility: By improving polygenic risk scores (PRS) and refining causal variant identification, SBayesCO enhances the potential for precision medicine applications, particularly for traits where current prediction accuracy is suboptimal.

Limitations Noted:

• Assumes no sample overlap between GWAS and molQTL datasets (though pQTL data used had some overlap, which could inflate PIPs).
• Current implementation uses a BayesC baseline; future work could integrate more flexible priors (e.g., SBayesR).
• Computationally intensive due to Gibbs sampling, though faster alternatives (Variational Bayes) are proposed for future development.