TL;DR
This paper introduces PRS-CS, a novel Bayesian regression method for polygenic risk score (PRS) prediction that uses continuous shrinkage priors. PRS-CS significantly improves prediction accuracy for complex traits and diseases compared to existing methods, especially with larger training datasets, by better modeling genetic architectures and local linkage disequilibrium.
What Problem Does This Paper Address?
Polygenic risk scores (PRS) are powerful tools for predicting an individual's genetic predisposition to complex traits and diseases. However, existing methods for calculating PRS face several challenges:
- Information Loss: Traditional methods often discard genetic variants based on linkage disequilibrium (LD) and p-value thresholds, leading to a loss of valuable information.
- Modeling Genetic Architectures: Complex traits and diseases have diverse genetic architectures, ranging from highly polygenic (many genes with small effects) to having a few genes with larger effects. Existing methods struggle to adapt to this variability.
- Computational Burden: Sophisticated Bayesian methods that can better model genetic architectures often suffer from high computational costs, limiting their scalability.
- LD Handling: Accurately accounting for local LD patterns, which describes how genetic variants are inherited together, is crucial for accurate PRS but challenging for many methods.
- Summary Statistics: Many applications require PRS calculation from publicly available genome-wide association study (GWAS) summary statistics, without access to individual-level data.
This paper aims to address these limitations by developing a more accurate, computationally efficient, and flexible PRS prediction method.
Methods
The core innovation of PRS-CS lies in its use of continuous shrinkage (CS) priors within a Bayesian high-dimensional regression framework.
- Bayesian Regression Framework: The model treats the prediction of a trait (y) as a linear combination of genetic markers (X) with associated effect sizes (β), plus residual error (ε): $y = X\beta + \epsilon$. The goal is to infer the posterior distribution of β.
- Continuous Shrinkage Priors: Unlike previous methods that used discrete priors (e.g., point-normal priors in LDpred), PRS-CS places a continuous shrinkage prior on the SNP effect sizes. This prior is a global-local scale mixture of normals, meaning:
- Global Shrinkage ($\phi$): A single parameter controls the overall degree of shrinkage across all SNPs.
- Local Shrinkage ($\psi_j$): Each SNP has its own shrinkage parameter, allowing for adaptive shrinkage based on the evidence for its effect.
- Continuous Mixing Density (g): This density is continuous, allowing for a flexible distribution of effect sizes that can be sharply peaked at zero (to shrink noise) while having heavy tails (to accommodate true signals). The paper specifically uses the Strawderman-Berger prior.
- Multivariate Block Update: A key advantage of CS priors is that they enable multivariate block updates of SNP effect sizes. Instead of updating each SNP's effect size individually, PRS-CS updates the effects of SNPs within local LD blocks jointly. This allows for more accurate modeling of local LD patterns and significantly improves computational efficiency.
- PRS-CS and PRS-CS-auto:
- PRS-CS: Requires tuning a global shrinkage parameter ($\phi$) by selecting the value that yields the best performance on a validation dataset.
- PRS-CS-auto: A fully Bayesian version that automatically learns the global shrinkage parameter from the data using a half-Cauchy prior, eliminating the need for a separate validation set.
Datasets
- UK Biobank: Used for simulation studies to evaluate the performance of PRS-CS and other methods across various genetic architectures and training sample sizes.
- Partners HealthCare Biobank: Used for real-world application of PRS-CS to predict six common complex diseases and six quantitative traits. This dataset provided individual-level data for validation and testing.
- 1000 Genomes Project (1KG) European Sample: Used as an external LD reference panel for all analyses. The paper also briefly discusses the impact of using an in-sample LD reference panel.
Key Insights and Ideas
- Continuous Shrinkage is Key: The shift from discrete to continuous shrinkage priors is the central methodological advance, enabling better modeling of diverse genetic architectures and improved computational efficiency.
- Adaptive Shrinkage: The local shrinkage parameter allows the method to adaptively shrink small, noisy effect sizes towards zero while preserving larger, data-supported signals.
- Efficient LD Modeling: Multivariate block updates within local LD regions are crucial for accurately capturing LD structure, a significant improvement over methods that update SNPs independently or use approximations.
- Scalability: The computational advantages of CS priors and block updates make PRS-CS scalable to genome-wide datasets, even with summary statistics.
- Robustness to Genetic Architectures: PRS-CS performs well across a wide range of genetic architectures, from sparse to highly polygenic, making it broadly applicable.
Key Results
- Simulation Studies: PRS-CS consistently outperformed existing methods (unadjusted PRS, P+T, LDpred, LDpred-inf) in terms of prediction accuracy (measured by R²) across various genetic architectures and training sample sizes. The improvement was particularly pronounced with larger training sample sizes.
- Real-World Application (Partners HealthCare Biobank):
- For quantitative traits, PRS-CS showed substantial improvements in prediction accuracy (e.g., 27.75% for height, 32.05% for cholesterol) compared to LDpred.
- For complex diseases, PRS-CS also demonstrated significant gains, especially for breast cancer (41.85% relative increase vs. LDpred) and rheumatoid arthritis (28.62% relative increase vs. LDpred).
- PRS-CS-auto showed promising results, particularly for quantitative traits, but its performance was sometimes lower than PRS-CS or LDpred with smaller GWAS sample sizes, suggesting that learning the global shrinkage parameter accurately requires sufficient data.
- Calibration: PRS-CS demonstrated good calibration, with a regression slope close to one, indicating that the predicted PRS accurately reflects the true phenotype.
Limitations and Future Work
- Training Sample Size: While PRS-CS generally outperforms other methods, its advantage is most pronounced with larger training sample sizes. For very small training sample sizes, LDpred could sometimes outperform PRS-CS.
- PRS-CS-auto Performance: In real-world applications with current GWAS sample sizes, PRS-CS-auto's performance was sometimes lower than PRS-CS or LDpred, suggesting that accurately learning the global shrinkage parameter from summary statistics may require larger GWAS datasets.
- LD Reference Panel: While the 1000 Genomes Project European sample served as a valid LD reference, using an in-sample LD reference panel generally provided marginal improvements in prediction accuracy.
- Future Directions:
- Exploring other CS priors: Investigating different continuous shrinkage prior distributions beyond the Strawderman-Berger prior.
- Individual-level data: Applying PRS-CS to individual-level data, where it is expected to offer significant computational advantages over existing Bayesian methods.
- Multi-trait and functional annotation: Incorporating multiple genetically correlated traits and functional annotations to further improve prediction accuracy.
- Trans-ethnic prediction: Developing methods for trans-ethnic prediction to maximize the utility of PRS in diverse populations.
- Clinical Utility: Acknowledging that current PRS prediction accuracy is still lower than ideal for widespread clinical use, and more work is needed to improve predictive performance and translational value.
Summary updated