TL;DR
This paper introduces LT-MAPIT, a novel statistical method for identifying genetic interactions (epistasis) in case-control studies. It extends existing "marginal epistasis" approaches to handle the complexities of binary disease outcomes and case-control study designs, offering improved power and accuracy compared to previous methods.
What Problem Does This Paper Address?
Genome-wide association studies (GWAS) aim to find genetic variants linked to diseases. While most focus on individual variants (additive effects), genetic interactions (epistasis) are known to play a crucial role in disease development. Traditional methods for detecting epistasis involve exhaustively searching all possible pairs or higher-order combinations of genetic variants. This approach is computationally intensive and suffers from a massive multiple testing burden, leading to low statistical power, especially in case-control studies. A more recent approach, "marginal epistasis," identifies variants involved in any interaction without pinpointing specific partners, reducing the search space. However, existing marginal epistasis methods were designed for quantitative traits and lacked power when applied to binary case-control data. This paper addresses this gap by developing a method that properly accounts for the binary nature of case-control data and the specific challenges of case-control study designs.
Methods
The core innovation is the Liability Threshold Marginal Epistasis Test (LT-MAPIT). This method extends the concept of marginal epistasis to case-control studies by employing a liability threshold mixed model. This model treats the observed binary disease status (case/control) as a thresholded version of an underlying continuous "liability" trait. This allows the application of statistical methods developed for quantitative traits to case-control data while properly accounting for ascertainment bias (how cases and controls are selected).
Key methodological aspects include:
- Liability Threshold Model: This is the foundation for handling binary outcomes and case-control ascertainment.
- Marginal Epistasis: Instead of testing all pairwise interactions, LT-MAPIT tests each variant for its marginal epistatic effect – its combined interaction effect with all other variants. This significantly reduces the number of tests.
- Variance Component Estimation: The model estimates variance components related to additive effects, marginal epistatic effects, and residual errors.
- Approximation Methods: To handle the computational complexity of the likelihood function, the paper uses approximations, including a first-order Taylor expansion and Markov Chain Monte Carlo (MCMC) for estimating latent liabilities.
- MQS Approach: A Method of Moments approach (MQS) is used for parameter estimation and hypothesis testing, which is computationally efficient.
- Genomic Control: A recalibration approach based on genomic control is applied to ensure well-calibrated p-values.
Datasets
The study uses two main sources of data:
- Simulated Data: Synthetic genotypes and phenotypes were generated to rigorously evaluate the performance (Type I error control and power) of LT-MAPIT and compare it against other methods. These simulations allowed for controlled manipulation of parameters like disease prevalence and the proportion of variance explained by interactions.
- WTCCC (Wellcome Trust Case Control Consortium) Data: This publicly available dataset comprises genetic data from approximately 14,000 cases across seven common diseases (bipolar disorder, coronary artery disease, Crohn's disease, hypertension, rheumatoid arthritis, type 1 diabetes, and type 2 diabetes) and shared controls. This real-world data was used to apply LT-MAPIT and identify potential marginal and pairwise epistatic associations.
Key Insights and Ideas
- Bridging the Gap: The most significant insight is the successful adaptation of marginal epistasis methods to the challenging domain of case-control studies, which was previously a major limitation.
- Power for Case-Control: Naively applying existing marginal epistasis methods to case-control data leads to a significant loss of power. LT-MAPIT's liability threshold approach rectifies this.
- Computational Efficiency: By focusing on marginal effects, LT-MAPIT avoids the combinatorial explosion of testing all pairwise interactions, making it more computationally feasible for genome-wide scans.
- Prioritization Tool: LT-MAPIT can serve as an effective first step in a two-step approach, identifying candidate variants for further, more intensive pairwise interaction testing.
- MHC Region Significance: The application to WTCCC data highlights the continued importance of the Major Histocompatibility Complex (MHC) region in various autoimmune and inflammatory diseases, as identified by LT-MAPIT.
Key Results
- Type I Error Control: Simulations demonstrated that LT-MAPIT effectively controls the Type I error rate, producing well-calibrated p-values across different disease prevalences.
- Increased Power: LT-MAPIT showed significantly higher power to detect marginal epistatic effects in case-control studies compared to existing marginal epistasis methods and conventional exhaustive search approaches.
- Robustness: The method demonstrated robustness to variations in disease prevalence and the proportion of variance explained by additive versus epistatic effects.
- Real Data Application: LT-MAPIT successfully identified potential marginal epistatic associations in seven complex diseases from the WTCCC dataset, with findings near known disease-associated genes (e.g., CD200, PSAT1, IFNB1) and within the MHC region.
Limitations and Future Work
- Direct Pairwise Identification: Like other marginal epistasis methods, LT-MAPIT identifies variants involved in epistasis but does not directly pinpoint the specific interacting partners. The paper proposes a two-step approach (LT-MAPIT followed by focused pairwise testing) as a practical solution.
- Computational Cost: While more efficient than exhaustive searches, LT-MAPIT can still be computationally demanding, requiring fitting a separate model for each SNP. The authors note that scaling to even larger datasets might require new algorithms.
- Higher-Order Interactions: The current implementation focuses on pairwise interactions. The authors suggest that LT-MAPIT can be extended to detect higher-order interactions by introducing additional random effects terms.
- Prioritization Strategy: The two-step prioritization procedure is described as "ad hoc" and could potentially miss important epistatic associations. Future work could explore more integrated statistical approaches to jointly detect marginal and pairwise interactions.
- Alternative Models: The paper briefly discusses logistic mixed models as an alternative for case-control studies but notes their limitations in handling ascertainment and multiple variance components compared to the liability threshold model. Exploring extensions of LT-MAPIT using logistic mixed models is suggested as a future research avenue.
Summary updated