Abstract The relationship of Alzheimer's disease (AD) to the brain has been widely studied. Early diagnosis of AD is challenging for multiple reasons, including the presence of variability in the clinical and pathological features within affected individuals. In recent work we have shown a large improvement in sensitivity and stability of brain- based biomarkers of various disorders using approaches that leverage the joint information across multiple mo- dalities. In this work we focus on three promising directions including first the joint fusion of multiple functional networks with brain structure by using a novel approach called parallel multilink joint independent component analysis. Secondly, we will develop an approach to leverage a generative framework for capturing dimensional changes in AD using an approach based on variational autoencoders, high order clustering, and meta-mapping between latent space variables and colors. Finally, we will extend this work to move towards a low dimensional prediction of AD risk by combining multiple large clinical datasets with large population datasets such as the UKbiobank data and others. The proposed aims have the potentially to significantly increase the sensitivity of imaging-based methods for predicting AD in unaffected individuals.