Identifying gene-gene interactions using penalized tensor regression.

Mengyun Wu,Jian Huang,Shuangge Ma
DOI: https://doi.org/10.1002/sim.7523
2018-01-01
Statistics in Medicine
Abstract:Gene-gene (GxG) interactions have been shown to be critical for the fundamental mechanisms and development of complex diseases beyond main genetic effects. The commonly adopted marginal analysis is limited by considering only a small number of G factors at a time. With the main effects, interactions hierarchical constraint, many of the existing joint analysis methods suffer from prohibitively high computational cost. In this study, we propose a new method for identifying important GxG interactions under joint modeling. The proposed method adopts tensor regression to accommodate high data dimensionality and the penalization technique for selection. It naturally accommodates the strong hierarchical structure without imposing additional constraints, making optimization much simpler and faster than in the existing studies. It outperforms multiple alternatives in simulation. The analysis of The Cancer Genome Atlas (TCGA) data on lung cancer and melanoma demonstrates that it can identify markers with important implications and better prediction performance.
What problem does this paper attempt to address?