TL;DR
This paper introduces mvSuSiE, a novel computational method for "fine-mapping" genetic associations across multiple traits simultaneously. By learning complex patterns of how genetic variants affect various traits, mvSuSiE significantly boosts the power to identify causal genetic variants and improves the precision of these findings compared to analyzing each trait independently.
What Problem Does This Paper Address?
Genome-wide association studies (GWAS) have identified many genetic regions linked to complex traits and diseases. However, pinpointing the exact genetic variants (SNPs) responsible for these associations, especially when multiple traits are involved, is challenging. Traditional methods often fine-map each trait separately, which is inefficient and misses opportunities to leverage shared genetic signals. Integrating results post-hoc is also difficult. Existing multi-trait methods often struggle with computational demands, make restrictive assumptions about how genetic effects are shared across traits, or cannot handle correlated traits. This paper addresses the need for a fast, flexible, and powerful method that can jointly fine-map many correlated traits, accurately modeling complex patterns of genetic effect sharing.
Methods
The core of this work is the mvSuSiE (multivariate Sum of Single Effects) model. It extends the single-trait SuSiE model by introducing a flexible, data-driven prior distribution that models the patterns of shared genetic effects across multiple traits. This prior is a mixture of multivariate normal distributions, allowing mvSuSiE to adapt to how genetic variants influence different traits, whether effects are specific to one trait, shared across a subset, or affect many traits. Key methodological contributions include:
- Flexible Effect Sharing: mvSuSiE learns and models heterogeneous patterns of genetic effect sharing across traits, unlike methods that assume independent effects.
- Data-Driven Prior: The parameters of the effect-sharing prior are estimated from the data, making the method adaptive.
- Handling Correlated Traits: mvSuSiE explicitly models correlations among traits, which is crucial for many biological datasets.
- Scalability: The method is designed to be computationally practical for large-scale datasets, including "biobank scale" data.
- Summary Data Analysis: mvSuSiE can also analyze summary statistics (e.g., z-scores and LD matrices), making it applicable to a wider range of studies.
- Novel Inference Summaries: The paper introduces methods to summarize multi-trait fine-mapping results, including cross-trait PIPs and credible sets, and trait-wise significance measures (local false sign rates).
Datasets
The primary dataset used for demonstrating mvSuSiE's application is the UK Biobank. Specifically, the authors analyzed:
- Genotype data: From approximately 257,605 "White British" individuals.
- Phenotype data: Detailed measurements for 16 blood cell traits from these individuals.
Simulations were also conducted using UK Biobank genotypes to evaluate the performance of mvSuSiE against other methods under various scenarios of trait correlation and effect sharing.
Key Insights and Ideas
- Shared Genetic Effects are the Norm: The analysis of blood cell traits revealed that a significant proportion of identified causal SNPs affect more than one trait (68% in the UK Biobank data), highlighting the importance of multi-trait analysis.
- Flexibility is Key: mvSuSiE's ability to learn data-driven effect-sharing patterns, rather than relying on pre-defined assumptions, is crucial for accurately fine-mapping complex biological systems like blood cell traits.
- Improved Discovery and Resolution: Jointly analyzing traits with mvSuSiE not only discovers more causal SNPs but also provides more precise localization (narrower credible sets) compared to single-trait analyses.
- Heterogeneity in Genetic Determination: Even within a single gene locus, different genetic variants can have distinct patterns of association with various blood cell traits, underscoring the complexity of genetic regulation.
Key Results
- Increased Discovery: In the UK Biobank blood cell trait analysis, mvSuSiE identified over 3,000 causal SNPs, a substantial increase compared to single-trait fine-mapping.
- Improved Precision: Credible sets (CSs) generated by mvSuSiE were narrower on average, leading to more precise localization of causal variants.
- Enhanced Power and Coverage in Simulations: mvSuSiE consistently outperformed single-trait SuSiE and competing multi-trait methods like CAFEH in simulations, showing higher power and better coverage for identifying causal SNPs, especially under complex shared effect scenarios.
- Computational Efficiency: mvSuSiE is computationally practical for large-scale datasets, handling "biobank scale" data efficiently.
- Broader Effect Detection: 68% of the identified causal SNPs in the UK Biobank data showed significant effects in more than one blood cell trait, demonstrating mvSuSiE's ability to capture pleiotropic effects.
Limitations and Future Work
- Model Assumptions: mvSuSiE assumes a standard multivariate Gaussian linear model, making it primarily applicable to quantitative traits. While it can be applied to binary traits with small effect sizes, this application is less theoretically grounded.
- Missing Data: The current model assumes complete data. For datasets with missing values, imputation is suggested, which might lead to a loss of power. More rigorous approaches for handling missing data in multivariate models are needed.
- Trait Specificity: While mvSuSiE models shared effects well, explicitly modeling trait-specific sparsity of causal SNPs could be beneficial in settings with a very large number of less related traits.
- Future Directions: The authors suggest that extensions to handle generalized linear models (for binary traits) and linear mixed models would be valuable. Further work on handling missing data and exploring more complex sparsity patterns could also enhance the method's applicability.
Summary updated