Leveraging the genetic correlation between traits improves the detection of epistasis in genome-wide association studies

Julian Stamp, Alan DenAdel, Daniel Weinreich, Lorin Crawford

TL;DR

This paper introduces mvMAPIT, a novel statistical framework that leverages correlations between multiple traits to significantly improve the detection of genetic interactions (epistasis) in genome-wide association studies (GWAS). By jointly analyzing traits, mvMAPIT offers increased power to identify variants involved in complex genetic architectures, outperforming traditional single-trait approaches.

What Problem Does This Paper Address?

Genome-wide association studies (GWAS) have been instrumental in identifying genetic variants associated with complex traits and diseases. However, a significant portion of the heritability for these traits remains unexplained, often referred to as "missing heritability." Epistasis, the phenomenon where the effect of one gene depends on the presence of one or more other genes, is a major contributor to this missing heritability. Traditional methods for detecting epistasis often involve computationally intensive, exhaustive searches for pairwise interactions, which can be underpowered, especially for traits with low heritability or when dealing with a vast number of genetic variants. Furthermore, most existing methods analyze traits one at a time, neglecting the potential power gains from jointly modeling correlated phenotypes. This paper addresses the challenge of efficiently and powerfully detecting epistatic variants by developing a multivariate approach that capitalizes on the shared genetic architecture across multiple traits.

Methods

The core innovation is the multivariate MArginal ePIstasis Test (mvMAPIT). This framework extends the univariate "MArginal ePIstasis Test" (MAPIT) by formulating it as a multivariate linear mixed model (mvLMM). Instead of testing each variant for interaction effects with all other variants on a single trait, mvMAPIT considers the interaction effects across multiple traits simultaneously.

Key methodological aspects include:

  • Marginal Epistasis: mvMAPIT focuses on detecting "marginal epistasis," meaning it identifies variants that have a significant interaction effect with any other variant, without needing to pinpoint the exact interacting partner. This reduces the combinatorial explosion of pairwise tests.
  • Multivariate Variance Component Estimation: The authors developed a novel algorithm for estimating variance components in the multitrait setting, enabling efficient parameter inference and P-value computation. This is crucial for handling the increased complexity of multiple phenotypes.
  • Leveraging Trait Correlations: The power of mvMAPIT stems from its ability to exploit the correlation structure between the epistatic effects across different traits. When an interaction has a similar effect (or correlated effects) on multiple traits, mvMAPIT can detect it more effectively.
  • Combinatorial Hypothesis Testing: mvMAPIT employs a three-step hypothesis testing procedure:
  • Univariate MAPIT tests on each trait individually.
  • A "covariance test" to assess shared epistatic effects between pairs of traits.
  • A combination of P-values from these steps using Fisher's method, harmonic mean, or Cauchy combination tests to derive an overall marginal epistatic P-value. This approach aims to aggregate signals and improve power.
  • Scalability: The authors claim their approach is scalable to moderately sized GWAS, with computational costs increasing with the number of traits but remaining manageable.

Datasets

The authors utilized two primary datasets for evaluation:

  1. Simulated Data: A simulation scheme was adapted from previous studies, using real genotypes from chromosome 22 of the Wellcome Trust Case Control Consortium (WTCCC) 1 study. This allowed for controlled generation of synthetic quantitative traits under various genetic architectures (varying heritability, additive vs. epistatic effects, and correlation structures) to assess the performance of mvMAPIT.
  2. Protein Sequence Data: Data from Phillips et al. (2021) on binding affinity of two anti-influenza antibodies (CR6261 and CR9114) to different influenza strains. This dataset represents a biological system where mutations can lead to epistatic effects on binding affinity.
  3. Mouse Heterogeneous Stock Data: Genotype and phenotype data from approximately 2,000 mice from the Wellcome Trust Centre for Human Genetics. This dataset includes 15 hematological phenotypes, representing a real-world scenario with correlated traits and varying genetic architectures.

Key Insights and Ideas

  • Joint modeling is powerful: The fundamental insight is that analyzing multiple correlated traits together can dramatically increase the statistical power to detect genetic associations, including epistasis, compared to analyzing each trait in isolation.
  • Correlated epistatic effects are key: mvMAPIT's advantage is most pronounced when epistatic interactions have correlated effects across traits. This suggests that shared biological pathways or mechanisms might be influencing multiple phenotypes simultaneously.
  • Marginal epistasis simplifies the search: By focusing on marginal epistasis, the method avoids the computational burden of testing every possible pairwise interaction, making it more feasible for large-scale studies.
  • A unified framework for complex traits: The mvMAPIT framework provides a flexible way to investigate the genetic underpinnings of traits that are not isolated but rather part of a correlated biological system.

Key Results

  • Type I Error Control: Simulations demonstrated that mvMAPIT effectively controls Type I error rates and produces well-calibrated P-values under the null hypothesis (i.e., when only additive effects are present).
  • Improved Power: mvMAPIT showed significantly improved power to detect epistatic variants compared to univariate MAPIT, especially when epistatic effects were correlated across traits (e.g., $v_a = 0.8$).
  • Real-world application success:
    • In the antibody binding affinity study, mvMAPIT successfully identified known epistatic mutations that affect binding to influenza strains, including those that univariate methods missed.
  • In the mouse hematology data, mvMAPIT identified significant epistatic SNPs for several trait pairs, often driven by the covariance component, and these findings were consistent with previous literature.
  • Scalability: The method is presented as scalable to moderately sized GWAS, with computational costs increasing with the number of traits.

Limitations and Future Work

  • Identifying specific interaction pairs: While mvMAPIT identifies SNPs involved in epistasis, it does not directly pinpoint the exact interacting partners. The authors suggest a two-step approach (mvMAPIT followed by explicit interaction testing) as a current workaround, but a unified approach is a future direction.
  • Computational cost for very large datasets: For biobank-scale studies with hundreds of thousands of individuals, the current implementation might still be computationally demanding. Future work could explore more efficient fitting of multivariate linear mixed models, potentially using randomized methods.
  • Gene-by-environment/sex interactions: The framework could be extended to incorporate gene-by-environment (GxE) or gene-by-sex (GxSex) interactions by modifying the covariance matrix.
  • Application to complex trait types: The authors envision applying mvMAPIT to more complex trait types beyond simple quantitative traits, such as longitudinal data, metabolic profiles, or morphological traits, which often have inherent dependencies between their components.
  • Caveat of same-locus additive effects: The authors acknowledge that inferred interactions can sometimes be confounded by additive effects of untyped variants in the same genomic region, a common challenge in GWAS.

Summary updated

Abstract

Epistasis, commonly defined as the interaction between genetic loci, is known to play an important role in the phenotypic variation of complex traits. As a result, many statistical methods have been developed to identify genetic variants that are involved in epistasis, and nearly all of these approaches carry out this task by focusing on analyzing one trait at a time. Previous studies have shown that jointly modeling multiple phenotypes can often dramatically increase statistical power for association mapping. In this study, we present the "multivariate MArginal ePIstasis Test" (mvMAPIT)-a multioutcome generalization of a recently proposed epistatic detection method which seeks to detect marginal epistasis or the combined pairwise interaction effects between a given variant and all other variants. By searching for marginal epistatic effects, one can identify genetic variants that are involved in epistasis without the need to identify the exact partners with which the variants interact-thus, potentially alleviating much of the statistical and computational burden associated with conventional explicit search-based methods. Our proposed mvMAPIT builds upon this strategy by taking advantage of correlation structure between traits to improve the identification of variants involved in epistasis. We formulate mvMAPIT as a multivariate linear mixed model and develop a multitrait variance component estimation algorithm for efficient parameter inference and P-value computation. Together with reasonable model approximations, our proposed approach is scalable to moderately sized genome-wide association studies. With simulations, we illustrate the benefits of mvMAPIT over univariate (or single-trait) epistatic mapping strategies. We also apply mvMAPIT framework to protein sequence data from two broadly neutralizing anti-influenza antibodies and approximately 2,000 heterogeneous stock of mice from the Wellcome Trust Centre for Human Genetics. The mvMAPIT R package can be downloaded at https://github.com/lcrawlab/mvMAPIT.
BibTeX
@inproceedings{stamp2023leveraging,
  author = {Julian Stamp and Alan DenAdel and Daniel Weinreich and Lorin Crawford},
  title = {{Leveraging the genetic correlation between traits improves the detection of epistasis in genome-wide association studies}},
  booktitle = {G3},
  year = {2023},
  doi = {10.1093/g3journal/jkad118},
}
Figure