Epistasis detection using a permutation-based Gradient Boosting Machine

Abstract

Detecting single nucleotide polymorphism (SNP) epistasis contributes to understand disease susceptibility and discover disease pathogenesis underlying complex disease. In this paper, we propose an approach called permutation-based Gradient Boosting Machine (pGBM) to detect pure epistasis by estimating the power of a GBM classifier which is influenced by permuting SNP pairs. pGBM is based on two permutation strategies and gradient boosting machine model. To extend pGBM to detect pure epistasis well on unbalanced dataset, average AUC difference value is chosen as the metric that quantifies the SNP interactions intensity. The experiment results demonstrate that our method has a high success rate with both balanced/unbalanced simulation and real dataset. In addition, pGBM shows great potential to detect pure SNP epistasis to uncover more complex disease pathogenesis.