Improved Multi-task SCCA for Brain Imaging Genetics via Joint Consideration of the Diagnosis, Parameter Decomposition and Network Constraints

Brain imaging genetics develops rapidly, aiming to identify bi-multivariate associations between genetic loci and neuroimaging quantitative traits (QTs). The multi-task Sparse Canonical Correlation Analysis (MTSCCA) is a popular and effective technique in this area since it obtains superior results than those single-task based SCCA methods. Unfortunately, the most existing MTSCCA methods are either unsupervised or incapable of identifying the shared and specific patterns of multimodal neuroimaging QTs simultaneously. In this paper, we propose a novel diagnosis guided MTSCCA to identify the association between genetic and imaging phenotypic markers. Our method has three merits. First, it follows the same modeling paradigm of previous MTSCCA. This enables it to incorporate multimodal imaging QTs jointly, thereby facilitating a more comprehensive identification of genetic factors. Second, our method utilizes the parameter decomposition which could identify both modality-shared and -specific imaging QTs, and further uncovers their genetic mechanisms. Third, we also employed a new network constraint which could find out potentially meaningful brain imaging networks. Compared with conventional SCCA methods including both single-task and multi-task ones, the proposed method has improved or comparable correlation coefficients, and obtains a clean imaging pattern of good meaning. In addition, these results on the Alzheimer’s disease neuroimaging initiative (ADNI) cohort show that our method selects meaningful biomarkers, indicating that it could offer a significant addition to brain imaging genetic studies.