In the field of Alzheimer?s and related disorder, there has been very little work focusing on imaging genomics biomarker approaches, despite considerable promise. In part this is due to the fact that most studies have fo- cused on candidate gene approaches or those that do not capitalize on capturing (and amplifying) small effects spread across many sites. Even for genome wide studies, the vast majority of imaging genomic studies still rely on massive univariate analyses. The use of multivariate approaches provides a powerful tool for analyzing the data in the context of genomic and connectomic networks (i.e. weighted combinations of voxels and genetic variables). It is clear that imaging and genomic data are high dimensional and include complex relationships that are poorly understood. Multivariate data fusion models that have been proposed to date typically suffer from two key limitations: 1) they require the data dimensionality to match (i.e. 4D fMRI data has to be reduced to 1D to match with the 1D genomic data, and 2) models typically assume linear relationships despite evidence of non- linearity in brain imaging and genomic data. New methods are needed that can handle data that has mixed temporal dimensionality, e.g., single nucleotide polymorphisms (SNPs) do not change over time, brain structure changes slowly over time, while fMRI changes rapidly over time. Secondly, methods that can handle complex relationships, such as groups of networks that are tightly coupled or nonlinear relationships in the data. To ad- dress these challenges, we introduce a new framework called flexible subspace analysis (FSA) that can auto- matically identify subspaces (groupings of unimodal or multimodal components) in joint multimodal data. Our approach leverages the interpretability of source separation approaches and can include additional flexibility by allowing for a combination of shallow and ?deep? subspaces, thus leveraging the power of deep learning. We will apply the developed models to a large longitudinal dataset of individuals at various stages of cognitive impair- ment and dementia. Using follow-up outcomes data we will evaluate the predictive accuracy of a joint analysis compared to a unimodal analysis, as well as its ability to characterize various clinical subtypes including those driven by vascular effects including subcortical ischemic vascular dementia versus those that are more neuro- degenerative. We will evaluate the single subject predictive power of these profiles in independent data to max- imize generalization. All methods and results will be shared with the community. The combination of advanced algorithmic approach plus the large N data promises to advance our understanding of Alzheimer?s and related disorders in addition to providing new tools that can be widely applied to other studies of complex disease. 3
It is clear that multimodal data fusion provides benefits over unimodal analysis, however existing approaches typically require the data to have matched dimensionality, leading to a loss of information. In addition, most models assume linear relationships, despite strong evidence of nonlinear relationships in the data. We propose to develop new flexible models to capture multi-scale brain imaging and genomics data which we will use to study a large data set of individuals with Alzheimer?s disease and Alzheimer?s disease related disorders. 2
