TL;DR
This paper introduces SuSiE 2.0, a significant upgrade to a popular genetic fine-mapping tool, offering improved speed, extensibility, and a new method (SuSiE-ash) that better handles complex genetic signals. SuSiE 2.0 demonstrates enhanced calibration and performance across various genetic architectures, making it a more robust and versatile tool for genetic research.
What Problem Does This Paper Address?
Genetic fine-mapping aims to pinpoint the specific genetic variants responsible for observed traits or diseases. The Sum of Single Effects (SuSiE) regression method has been a go-to tool for this, but its original implementation had limitations in terms of speed and flexibility, making it difficult to extend or integrate into complex analysis pipelines. Furthermore, in scenarios with mixed strong and moderate genetic signals, existing methods could struggle with calibration (i.e., accurately estimating the probability of a variant being causal). This paper addresses these issues by presenting a redesigned SuSiE (SuSiE 2.0) that is faster, more modular, and introduces SuSiE-ash to improve performance under complex genetic architectures.
Methods
SuSiE 2.0 features a modular redesign that separates different stages of the analysis (interface, constructor, workhorse, refinement). This makes it easier to add new features and improves code reusability. Key improvements include:
- Up to 5x speedup for summary statistics analysis, especially when using regularized LD matrices.
- SuSiE-ash: A novel method that introduces an adaptive shrinkage prior for moderate to weak genetic effects. This helps to better model complex genetic backgrounds where strong and weaker signals coexist, improving calibration and reducing false discovery rates.
- Enhanced model refinement and flexible convergence criteria for greater robustness.
- Incorporation of published extensions like SuSiE-inf for modeling infinitesimal backgrounds.
The core of the method still relies on Iterative Bayesian Stepwise Selection (IBSS), but the modular architecture allows for easier customization and integration of new components.
Datasets
The study used genotype data from the UK Biobank for their simulations. They also mention using real-data analysis scripts for TWAS (Transcriptome-Wide Association Studies) prediction. A new R package, simxQTL, was developed to implement diverse genetic architectures for benchmarking.
Key Insights and Ideas
- Modularity is key for extensibility: The redesigned architecture of SuSiE 2.0 is a significant step forward, allowing for easier integration of future methods and improvements.
- Adaptive shrinkage for complex signals: SuSiE-ash's adaptive shrinkage prior is crucial for improving calibration in scenarios with mixed strong and moderate genetic effects, a common challenge in real genetic data.
- Balancing power and FDR: SuSiE-ash achieves a better trade-off between statistical power and controlling the false discovery rate (FDR) compared to previous methods, especially under complex genetic architectures.
- SuSiE's underappreciated role in TWAS: The paper highlights that SuSiE-based methods are effective tools for TWAS prediction, a valuable application beyond just fine-mapping.
Key Results
- Speed Improvements: SuSiE 2.0 offers up to a 5x speed improvement for summary statistics analysis.
- FDR Reduction: SuSiE-ash demonstrated a 1.5-3x FDR reduction in complex polygenic backgrounds while maintaining statistical power.
- Calibration: SuSiE-ash showed improved calibration, particularly in scenarios with mixed effect sizes.
- Prediction Accuracy: Prediction accuracy for TWAS was nearly identical across methods, with SuSiE and SuSiE-ash showing a slight advantage.
- ROC Performance: SuSiE-ash achieved the best variant-level ROC performance at low false positive rates in many scenarios.
Limitations and Future Work
- Method Choice Depends on Application: The authors note that the best method choice (SuSiE, SuSiE-ash, or SuSiE-inf) depends on the specific application and the user's tolerance for false discoveries. For exploratory analysis, standard SuSiE might be preferred for sensitivity, while SuSiE-ash or SuSiE-inf are better for improved calibration.
- Future Extensions: The modular architecture of SuSiE 2.0 is designed to readily accommodate future extensions, such as generalized linear models or integration into a generalized IBSS framework. Developers can build directly on the SuSiE 2.0 codebase for compatibility.
Summary updated