Performance of epistasis detection methods in semi-simulated GWAS

Clement Chatelain, Clément Chatelain 0002, Guillermo Durán, Vincent Thuillier, Franck AugéBMC Bioinformatics 2018 · DOI

TL;DR

This paper tackles the challenge of finding gene-gene interactions (epistasis) in large genetic datasets (GWAS), which is crucial for understanding complex diseases but computationally demanding and prone to false positives. The authors developed a realistic simulation pipeline and rigorously compared five leading epistasis detection methods, finding that DSS offers a good balance of accuracy and speed, especially for interactions between genes with low linkage disequilibrium.

What Problem Does This Paper Address?

Genome-Wide Association Studies (GWAS) have been instrumental in identifying genetic variants associated with diseases. However, these studies often explain only a fraction of the "missing heritability," leading researchers to believe that interactions between genes (epistasis) play a significant role. Detecting epistasis is computationally intensive due to the vast number of gene pairs to test, and it's challenging to control for false positives, especially when genes are physically close on chromosomes (high linkage disequilibrium - LD). While many methods exist, there's a lack of independent, comprehensive comparisons using realistic simulated data that accounts for complex LD patterns. This paper aims to fill that gap by evaluating several popular epistasis detection methods under realistic conditions.

Methods

The core of this work is a novel three-step simulation pipeline designed to generate semi-simulated GWAS data that closely mimics real-world scenarios.

  1. Realistic Genotype Simulation: They used European samples from the 1000 Genomes Project as a template to simulate genotypes for 100,000 individuals, preserving the complex Linkage Disequilibrium (LD) structure found in human populations.
  2. Disease Model Simulation: They simulated disease phenotypes based on predefined genetic models, including additive effects and various types of epistasis (gene-gene interactions). This allowed them to control the ground truth of which interactions are truly associated with the disease.
  3. Case-Control Cohort Generation: They sampled cases and controls from the simulated population to create realistic GWAS cohorts.

Using this pipeline, they evaluated five popular bivariate epistasis detection methods: fastepi, GBOOST, SHEsisEpi, DSS, and IndOR. Their performance was assessed based on false positive rate control, statistical power, and computational speed (leveraging GPU acceleration).

Datasets

The primary dataset used for simulation was the European (EUR) samples from the 1000 Genomes Project phase 3 to establish realistic LD patterns. For validation, they used a real GWAS dataset for Type 2 Diabetes (T2D) from the Welcome Trust Case Control Consortium (WTCCC).

Key Insights and Ideas

  • LD is a major challenge: The presence of LD between causal SNPs significantly impacts the false positive rates of some methods, particularly fastepi and IndOR.
  • Simulation is crucial: Evaluating epistasis methods requires realistic simulations that capture complex LD structures, as simplified simulations can lead to misleading conclusions.
  • No single "best" method: Different methods excel under different conditions. DSS is generally strong, especially when LD is low, while fastepi shows some power in high LD scenarios but struggles with FPR control.
  • Computational feasibility: With GPU acceleration, exhaustive genome-wide epistasis analysis is now computationally feasible within hours for large datasets.
  • Complementary approaches: Combining methods like DSS and GBOOST might offer a more robust approach to detecting different types of epistasis.

Key Results

  • False Positive Rate (FPR): DSS, GBOOST, and SHEsisEpi demonstrated good control of the false positive rate. fastepi and IndOR showed an increased FPR when causal SNPs were in LD.
  • Statistical Power: DSS was the most powerful method in most scenarios with no or weak LD between causal SNPs. fastepi was the only method with measurable power in scenarios with strong LD (r² = 0.5), though its FPR control was poor.
  • AUC: DSS generally achieved the highest Area Under the ROC Curve (AUC) across most scenarios, indicating superior overall performance in distinguishing true positives from false positives.
  • Computational Performance: fastepi and SHEsisEpi were the fastest, followed by DSS, with IndOR and GBOOST being the slowest. However, all methods were capable of analyzing large GWAS datasets within a reasonable timeframe using GPUs.
  • WTCCC T2D Analysis: GBOOST and DSS detected distinct sets of interacting SNP pairs and genes, suggesting they capture different biological signals and could be used in combination.

Limitations and Future Work

  • IndOR's FPR and Numerical Issues: The paper notes that IndOR's FPR can be inflated, partly due to numerical issues arising from inverting covariance matrices, especially in smaller cohorts. They also suggest that IndOR's different definition of epistasis might contribute to its performance.
  • fastepi's FPR Control: While fastepi showed some power in high LD, its poor FPR control makes it less reliable in such scenarios.
  • Future Directions: The authors suggest expanding the evaluation to include gene-based epistasis methods and exploring how to translate statistical interactions into biological mechanisms. They also highlight the need for more robust methods for interpreting interactions within genes or genomic regions, and the potential for future large GWAS to detect weaker epistasis effects.

Summary updated

BibTeX
@article{chatelain2018performance,
  author = {Clement Chatelain and Clément Chatelain and Guillermo Durán and Vincent Thuillier and Franck Augé},
  title = {{Performance of epistasis detection methods in semi-simulated GWAS}},
  journal = {BMC Bioinformatics},
  year = {2018},
  doi = {10.1186/s12859-018-2229-8},
}
Figure