Bootstrap Inference with Neural-Network Modeling for Gene-Disease Association Testing

Abstract

Estimates derived from neural network modeling are used to test the significance of single nucleotide polymorphisms (SNPs) in the categorization of case control status in genetic association studies. Our artificial neural network (ANN) model of gene-disease correlation is represented by a fully connected 3-layered feedforward neural network with input nodes, corresponding to the number of studied SNPs, a hidden layer and a single output unit for the disease status. We used an evaluation procedure that measures the predictive significance of a single SNP, based on the change in the error function when the input is removed from the network. Two ANNs, one with all inputs and the other with a tested input removed are run in parallel and the change in error is calculated as a function of the relative out-of-sample performance of these two networks. With the help of a bootstrap technique (resampling with replacement) the valid inference of the tested input variables is derived. We report on the performance of the procedure as evaluated on simulated SNP datasets of varying complexity, on a real study dataset and in comparison to an SVM implementation of the procedure