Evaluation of epistasis detection methods for quantitative phenotypes

Abstract Background Epistasis, or genetic interaction, has been increasingly recognized for its ubiquity and for its role in susceptibility to common human diseases, such as Alzheimer’s. A wide variety of epistasis detection tools are currently available with several studies comparing the performance of methods suitable for case-control data. However, there is limited understanding of how well these tools perform with quantitative phenotypes. Methods We identified six epistasis detection methods suitable for quantitative phenotype data: EpiSNP, Matrix Epistasis, MIDESP, PLINK Epistasis, QMDR, and REMMA. To evaluate these tools, we generated simulated datasets using EpiGEN. The datasets modeled various pairwise interactions between disease-associated SNPs, including dominant, multiplicative, recessive, and XOR interactions. Additionally, we assessed the BOOST and MDR algorithms on discretized (case-control) version of the datasets. These tools were then tested on the Adolescent Brain Cognitive Development (ABCD) dataset for the externalizing behavior phenotype. Results Each tool exhibited strong performance for certain interaction types, but weaker performance for others. MDR achieved the highest overall detection rate of 60%, while EpiSNP had the lowest overall detection rate of 7%. MDR and MIDESP performed best at detecting multiplicative interactions with detection rates of 54% and 41% respectively. Both MDR and MIDESP were also effective at detecting XOR interactions with detection rates of 84% and 50% respectively. PLINK Epistasis, Matrix Epistasis, and REMMA excelled at detecting dominant interactions, all achieving a 100% detection rate. On the other hand, EpiSNP was particularly effective at detecting recessive interactions with a detection rate of 66%. When analyzing the ABCD dataset, Plink Epistasis and Plink BOOST identified SNPs within the DRD2 and DRD4 genes, which have been previously linked to externalizing behavior. Conclusion Since no single method consistently outperforms others across all types of epistasis, and given that the specific types of epistasis present in a dataset are often unknown, it may be more effective to use multiple epistasis detection algorithms in combination to obtain comprehensive results.