Gene-Gene Interaction Tests Using SVM and Neural Network Modeling

Abstract

Artificial neural networks (ANN) and support vector machine (SVM) modeling offer promise in the analysis of genotype-phenotype correlation in genetic association studies. In particular, we are interested in studying single nucleotide polymorphisms (SNPs) as genetic markers as predictors of a dichotomous disease outcome. The problem we are investigating is that of gene-gene and gene-environment interactions as determinants of the expression of complex diseases. This study builds on our previous work for a single gene testing procedure developed and presented earlier (Matchenko-Shimko and Dube, 2006). As for single SNPs pre-selection (Matchenko-Shimko and Dube, 2006), we rely on ANN sensitivity analysis algorithms to detect potential pairs of interacting SNPs associated with the disease outcome. The statistical test for SNP interaction is computed using a bootstrap technique and is based on the measure of the predictive significance of two SNPs from the change in the ANN error function (SVM regression error) when these two SNPs are removed from the ANN or SVM genotype-phenotype models. To investigate the power to detect and test gene-gene interactions we simulated genotypes including two interacting loci with low marginal effects, incomplete penetrance and phenocopies according to three different models of interaction