Identifying significant gene‐environment interactions using a combination of screening testing and hierarchical false discovery rate control

H. Robert Frost, Li Shen, Andrew J. Saykin, Scott M. Williams, Jason H. Moore, the Alzheimer's Disease Neuroimaging InitiativeGenetic 2016 · DOI

TL;DR

This paper introduces an enhanced method, SPUR, for identifying gene-environment (GxE) interactions in large-scale genetic studies. By combining score-statistic prescreening with hierarchical false discovery rate control, SPUR significantly improves statistical power and the ability to detect biologically relevant interactions, even with limited sample sizes.

What Problem Does This Paper Address?

Detecting gene-environment (GxE) interactions is crucial for understanding complex diseases, but it's a significant challenge in genome-wide association studies (GWAS). The sheer number of genetic markers (millions) necessitates stringent multiple hypothesis correction, which drastically reduces statistical power. Existing methods, while improving on single-step approaches, often struggle with computational feasibility for truly genome-wide data and lack formal methods for assessing the significance of individual interactions. This paper aims to overcome these limitations by developing a more powerful, scalable, and statistically rigorous GxE interaction detection framework.

Methods

The core of this work is an extension of a previously developed method called SPUR (Sequential Penalized and Unpenalized Regression). The key enhancements are:

  • Score Statistic Prescreening: To handle the massive scale of GWAS data, a computationally efficient prescreening step is introduced. This step uses a score statistic to quickly filter out a large proportion of genetic markers (SNPs) that are unlikely to be involved in GxE interactions, reducing the number of markers that need to be analyzed in subsequent, more computationally intensive steps. This is crucial for making the method applicable to millions of SNPs.
  • Hierarchical False Discovery Rate (FDR) Control: To formally assess the significance of individual GxE interactions identified in the testing stage, the authors incorporate a hierarchical FDR approach. This method allows for the simultaneous evaluation of multiple filter statistics (e.g., marginal association and gene-environment correlation) while maintaining appropriate control over the false discovery rate, providing a more robust way to identify significant interactions.
  • Generalized Exposure Handling: The method is extended to handle both binary and continuous environmental exposures, increasing its applicability to a wider range of studies.

The overall workflow involves:

  1. Prescreening: Using score statistics to select a subset of candidate SNPs.
  2. Filtering: Applying penalized regression (elastic net) using either marginal association or gene-environment correlation filters on the candidate SNPs.
  3. Testing: Fitting an unpenalized logistic regression model with the filtered SNPs to test for GxE interactions.
  4. Significance Assessment: Using an omnibus test for the overall GxE effect and hierarchical FDR for individual interaction tests.

Datasets

The paper evaluates its method using:

  • Simulated Data:
    • Simple Simulation Design: Used to characterize type I error control and power, simulating 2,000 SNPs and 1,000 subjects. This design was adapted from previous work (Dai et al., 2012).
  • Disease-Based Simulation Design: Based on genetic architectures of breast cancer, type 2 diabetes, and rheumatoid arthritis, adapted from Aschard et al. (2012a). This aims to provide a more realistic evaluation of performance.
  • Real Data:
    • Alzheimer's Disease Neuroimaging Initiative (ADNI) Study Data: This dataset was used to identify biologically plausible SNP-education interactions related to Alzheimer's disease status. The analysis focused on approximately 400,000 SNPs and 572 subjects (412 cases, 160 controls).

Key Insights and Ideas

  • The Power of Prescreening: The score statistic prescreening step is critical for enabling the analysis of truly genome-wide data. Without it, the computational burden of penalized regression on millions of SNPs would be prohibitive.
  • Hierarchical FDR is Key for Multiple Filters: The ability to formally assess individual interactions using hierarchical FDR is a significant improvement. It allows researchers to confidently identify specific GxE pairs without sacrificing statistical rigor, especially when using multiple filtering strategies.
  • Filter Choice Matters, But SPUR Mitigates Risk: The simulations clearly show that the choice of filter (marginal association vs. gene-environment correlation) can dramatically impact power. However, SPUR's ability to evaluate multiple filters in parallel with hierarchical FDR offers a robust solution, minimizing the penalty for not picking the "optimal" filter beforehand.
  • Biologically Plausible Findings in ADNI: The application to ADNI data successfully identified several education-SNP interactions that have known or plausible links to Alzheimer's disease, demonstrating the method's real-world utility.

Key Results

  • Type I Error Control: The extended SPUR method demonstrated acceptable type I error control across all simulation designs, with only a slight inflation in some cases, which was deemed acceptable given the substantial power gains.
  • Statistical Power:
    • In simple simulations, SPUR significantly outperformed standard screening-testing methods, particularly when using the appropriate filter. The global SPUR test (evaluating all interactions simultaneously) showed the highest power.
  • In disease-based simulations, power was generally lower but SPUR still showed superior performance compared to one-step and standard screening-testing methods.
  • The hierarchical FDR variant of SPUR provided impressive relative power, only slightly lower than SPUR using the optimal filter.
  • ADNI Analysis: SPUR identified several education-SNP interactions that were biologically plausible and relevant to Alzheimer's disease, including associations with genes like TOMM40, TXNL1, GRIA1, SULF1, SPATA13, and MIR633. Notably, six of the top ten genes associated with significant interactions had a known or plausible link to AD.

Limitations and Future Work

  • Computational Feasibility: While the prescreening step significantly improves scalability, fitting penalized regression models for tens of thousands of markers can still be computationally intensive.
  • Filter Selection: Although hierarchical FDR helps mitigate the risk of choosing the wrong filter, understanding the optimal conditions for each filter remains an area of interest.
  • Future Directions: The authors plan to:
  • Apply SPUR to a wider range of GWAS datasets.
  • Explore methodological enhancements, such as supporting higher-order interactions and different marker codings.
  • Investigate the concurrent evaluation of different interaction scales.

Summary updated

Abstract

Although gene-environment (G× E) interactions play an important role in many biological systems, detecting these interactions within genome-wide data can be challenging due to the loss in statistical power incurred by multiple hypothesis correction. To address the challenge of poor power and the limitations of existing multistage methods, we recently developed a screening-testing approach for G× E interaction detection that combines elastic net penalized regression with joint estimation to support a single omnibus test for the presence of G× E interactions. In our original work on this technique, however, we did not assess type I error control or power and evaluated the method using just a single, small bladder cancer data set. In this paper, we extend the original method in two important directions and provide a more rigorous performance evaluation. First, we introduce a hierarchical false discovery rate approach to formally assess the significance of individual G× E interactions. Second, to support the analysis of truly genome-wide data sets, we incorporate a score statistic-based prescreening step to reduce the number of single nucleotide polymorphisms prior to fitting the first stage penalized regression model. To assess the statistical properties of our method, we compare the type I error rate and statistical power of our approach with competing techniques using both simple simulation designs as well as designs based on real disease architectures. Finally, we demonstrate the ability of our approach to identify biologically plausible SNP-education interactions relative to Alzheimer's disease status using genome-wide association study data from the Alzheimer's Disease Neuroimaging Initiative (ADNI).
BibTeX
@inproceedings{frost2016identifying,
  author = {H. Robert Frost and Li Shen and Andrew J. Saykin and Scott M. Williams and Jason H. Moore and the Alzheimer's Disease Neuroimaging Initiative},
  title = {{Identifying significant gene‐environment interactions using a combination of screening testing and hierarchical false discovery rate control}},
  booktitle = {Genetic},
  year = {2016},
  doi = {10.1002/gepi.21997},
}
Figure