Dr. Suprateek Kundu (MD Anderson Cancer Center) will present at the October SI Virtual Working Group Meetings (2023-2024) on Tuesday, October 24 at 1:30pm EST (12:30pm CST/11:30am MST/10:30am PST).
Bayesian Longitudinal Tensor Models for High-dimensional Imaging Genetics Analysis
Imaging data often serve as effective endotypes for discovering genetic signatures driving mental disorders. The overwhelming majority of literature has focused on penalized approaches for analyzing image-gene associations, although some limited Bayesian approaches have also been proposed. However, existing methods routinely ignore modeling the spatial information in the brain image when analyzing image-gene associations. Our goal is to propose a novel spatially-aware Bayesian tensor-based analysis for inferring longitudinal voxel-wise image-gene associations that is able to map longitudinal trajectories of neurodegeneration. Some innovative features of the proposed method include
- incorporating the spatial configuration of voxels;
- gene network information to model the regression coefficients;
- the ability to pool information across visits to produce more reliable estimates.
This is made possible via structured tensor-based representation for the coefficient matrices that are modeled under suitable prior distributions incorporating gene network information. We develop an efficient Markov chain Monte Carlo (MCMC) to implement the proposed approach, and propose a joint credible regions approach for inferring significant features. Extensive simulation studies illustrate the clear advantages compared to routinely used cross-sectional approaches, spatially-naive methods, as well as methods that do not incorporate gene network knowledge. An application of the proposed approach to ADNI-1 data set reveals distinct spatial patterns of image-gene association between two groups separated by low and high rates of cognitive decline, with such genetic signatures potentially indicative of cognitive decline in Alzheimer’s Disease.