A General Framework of Brain Region Detection And Genetic Variants Selection in Imaging Genetics

Abstract: Imaging genetics is a growing field that employs structural or functional neuroimaging techniques to study individuals with genetic risk variants potentially linked to specific illnesses. This area presents considerable challenges to statisticians due to the heterogeneous information and different data forms it involves. In addition, both imaging and genetic data are typically high-dimensional, creating a "big data squared" problem. Moreover, brain imaging data contains extensive spatial information. Simply vectorizing tensor images and treating voxels as independent features can lead to computational issues and disregard spatial structure. This paper presents a novel statistical method for imaging genetics modeling while addressing all these challenges. We explore a Canonical Correlation Analysis based linear model for the joint modeling of brain imaging, genetic information, and clinical phenotype, enabling the simultaneous detection of significant brain regions and selection of important genetic variants associated with the phenotype outcome. Scalable algorithms are developed to tackle the "big data squared" issue. We apply the proposed method to explore the reaction speed, an indicator of cognitive functions, and its associations with brain MRI and genetic factors using the UK Biobank database. Our study reveals a notable connection between the caudate nucleus region of brain and specific significant SNPs, along with their respective regulated genes, and the reaction speed.