Comparative analysis of machine learning models for shortlisting SNPs to facilitate detection of marginal epistasis in GWAS

Epistasis, an essential genetic element causing phenotypic diversity, is frequently characterized as the interaction between two or more genes. Previous models could identify marginal epistatic interactions by mapping variants that have nonzero marginal epistatic effects. However, these models fail short of identifying individual interaction partners. To reduce the computational burden of the existing epistasis detection algorithms without compromising the detection of exact epistatic partners, strengths of various machine learning algorithms were exploited as a filtering strategy. Seven machine learning strategies were compared for shortlisting marginally associated SNPs that includes AdaBoost, artificial neural network, 3 random forest, stepwise regression, ridge regression, lasso and elastic net. Datasets were simulated for different combinations of heritability and minor allele frequencies, and performances of different algorithms were evaluated using power and precision measures. We found that ridge regression model outperformed the other models in shortlisting marginal epistasis-related SNPs. Thus, it is expected that epistasis detection tools will benefit by adding a filtering stage using ridge regression for efficient detection of marginal epistasis in large genomic datasets.