Computational Tools for SNP Interactions - How Good Are They?

It is no trivial task to sift through huge amounts of SNP data to detect interactions between SNPs that can be relevant to identify propensity for a certain disease or a phenotype trait of interest, especially because many times it also involves the influence of environmental aspects. In a previous work, we analyzed the impact of different epistatic models on the accuracy of exhaustive computational methods. Those methods have good accuracy, but they are by nature, highly computationally demanding, hence not well suited for large population size or large number of SNPs, as found in genome-wide studies. In this paper, we report the results of a comparative study of methods for detecting epistatic interactions, based on recent trends, namely greedy and Bayesian computational approaches. Our experiments reveal that all methods have better performance in scenarios with higher values for heritability and minor allele frequency (MAF). In general, in terms of accuracy, BOOST outperformed the other methods studied. Even presenting an statistically significantly better performance, BOOST could not reach 40% accuracy when there were 50 or more SNPs, for cases with heritability 0.01 and MAF 0.2, even with a large number of individuals.