From Phenotype To Genotype: A Structured Sparse Learning Framework For Imaging Genetics Studies

Abstract

Sparsity is one of the intrinsic properties of real-world data, thus sparse representation based learning models have been widely used to simplify data modeling and discover predictive patterns. By enforcing properly designed structured sparsity, one can unify specific data structures with the learning model. We proposed several novel structured sparsity learning models for multi-modal data fusion, heterogeneous tasks integration, and group structured feature selection. We applied our new structured sparse learning methods to the emerging imaging genetics studies by integrating phenotypes and genotypes to discover new biomarkers which are able to characterize neurodegenerative process in the progression of Alzheimer's disease and other brain disorders. Different to traditional association studies, our new structured sparse learning models can elegantly take advantage of the useful information contained in biomarkers, cognitive measures, and disease status, where, crucially, the interrelated structures within and between both genetic/imaging data and clinical outcomes are gracefully exploited by our newly designed convex sparse regularization models.We empirically evaluate our new methods on the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort to identify Alzheimer's disease (AD) risky biomarkers, where we have achieved not only clearly improved prediction performance for cognitive measurements and diagnosis status, but also a compact set of highly suggestive biomarkers relevant to AD.