Analysis of Genome-Wide Association Studies with Multiple Outcomes Using Penalization

Jin Liu, Jian Huang, Shuangge MaPLoS 2012 · DOI

Abstract

Genome-wide association studies (GWAS) are essential for identifying genetic markers linked to complex phenotypes. However, traditional univariate methods often fail to account for the correlation between multiple outcomes and the joint effects of high-dimensional SNP data. This paper introduces a joint modeling framework that employs a Group Lasso penalty to analyze multiple correlated phenotypes simultaneously. By grouping the effects of individual SNPs across all outcomes, the method enhances marker identification accuracy and predictive performance. Results from simulations and an analysis of a heterogeneous stock mice dataset demonstrate that this approach significantly outperforms standard single-outcome penalization methods.

Introduction: The Challenge of High-Dimensional GWAS

The primary goal of GWAS is to identify single nucleotide polymorphisms (SNPs) associated with biologically meaningful outcomes. Most current methodologies utilize a "one-at-a-time" strategy, testing associations between individual markers and single traits. This approach has significant drawbacks: it ignores the aggregate effects of multiple SNPs and fails to leverage the shared genetic architecture often found in correlated phenotypes. While penalization techniques like the Lasso have been adapted for high-dimensional genomic data, they typically lack the capacity to model multiple response variables jointly. This study proposes a multi-outcome penalization framework to bridge this gap.

Methodology: A Joint Modeling Framework

The authors propose a multi-step statistical approach to handle multiple correlated phenotypes:

  1. Joint Modeling & Transformation: Multi-response data is transformed into a uni-response format, allowing the model to explicitly accommodate the correlation structure between different phenotypes within the same subject.
  2. Group Lasso Penalty: The core of the method is the application of a Group Lasso penalty. Here, a "group" is defined as the set of regression coefficients for a single SNP across all measured outcomes. This encourages "group-level sparsity," meaning the model either selects a SNP as relevant to the entire suite of outcomes or excludes it entirely.
  3. Computational Efficiency: To manage the scale of GWAS data, the authors developed a Group Cyclical Coordinate Descent (GCD) algorithm. This iterative optimization process is made computationally feasible through block-wise orthogonalization of transformed covariates.

Simulation Study

To validate the model, the authors generated synthetic datasets with 500 subjects and up to 10,000 SNPs, simulating various levels of correlation between response variables.

Illustration of the structured coefficient vector used in simulation scenarios to represent the genetic effects of SNPs across multiple outcomes.

The simulations demonstrated that the joint modeling approach consistently achieved higher True Positive rates and lower Sum of Squared Errors (SSE) than individual Lasso analyses. The advantage was most pronounced when the phenotypes exhibited moderate to strong correlation, confirming that the model effectively uses phenotypic correlation as a source of information.

Real-World Application: Heterogeneous Stock Mice

The method was applied to a real-world dataset involving 2,202 mice and 13,459 SNP markers. The analysis focused on two highly correlated immunological traits: the CD4/CD8 ratio and the CD4:CD3 index.

Manhattan plot showing the absolute values of estimated coefficients (|β|) across chromosomes for the CD4:CD3 index, highlighting the SNPs identified by the joint modeling approach.

The joint method identified 45 SNPs, with 38 reaching statistical significance. In terms of predictive power, the joint model outperformed individual Lasso analysis by approximately 28%, reducing the Prediction Mean Squared Error (MSE) from 2.33 to 1.66. This suggests that joint modeling captures a more accurate representation of the underlying genetic signals.

Discussion and Limitations

The proposed method offers a robust framework for pleiotropic analysis in GWAS. However, it relies on the assumption that the same set of SNPs affects all outcomes. While simulations showed resilience in "partially matched" scenarios, the method's performance is optimized for truly shared genetic bases. Additionally, the approach may struggle with very rare variants (low Minor Allele Frequency) due to covariate sparsity. Future research will focus on scaling the algorithm to handle dozens of phenotypes and exploring alternative group-type penalties like Group SCAD.

Summary updated

Abstract

Genome-wide association studies have been extensively conducted, searching for markers for biologically meaningful outcomes and phenotypes. Penalization methods have been adopted in the analysis of the joint effects of a large number of SNPs (single nucleotide polymorphisms) and marker identification. This study is partly motivated by the analysis of heterogeneous stock mice dataset, in which multiple correlated phenotypes and a large number of SNPs are available. Existing penalization methods designed to analyze a single response variable cannot accommodate the correlation among multiple response variables. With multiple response variables sharing the same set of markers, joint modeling is first employed to accommodate the correlation. The group Lasso approach is adopted to select markers associated with all the outcome variables. An efficient computational algorithm is developed. Simulation study and analysis of the heterogeneous stock mice dataset show that the proposed method can outperform existing penalization methods.
BibTeX
@inproceedings{liu2012analysis,
  author = {Jin Liu and Jian Huang and Shuangge Ma},
  title = {{Analysis of Genome-Wide Association Studies with Multiple Outcomes Using Penalization}},
  booktitle = {PLoS},
  year = {2012},
  doi = {10.1371/journal.pone.0051198},
}
Figure