The economic impact of substance and schizophrenia amounts to hundreds of billions of dollars, not to mention the countless lives impacted both directly and indirectly. Schizophrenia (SZ) and substance use disorders (SUD) are both extremely complex, both with substantial genetic and environmental components and with some shared aspects. In addition, schizophrenia patients tend to have increased levels of substance use which further complicates our understanding of the diagnosis. Most studies of SZ and SUD which incorporate imag ing and genetic data still ignore most of the information provided by the data by only analyzing a small number of genetic factors or brain regions. To characterize the available information, we are in need of approaches that can deal with high-dimensional data exhibiting interactions at multiple levels, while providing interpretable solu- tions. In the previous funding period we developed methods for pairwise coupling of high-dimensional genetic and imaging data which provided a powerful way to analyze the full information in joint data sets. However this is just the tip of the iceberg because in order to understand the complex interchange of biological pathways, brain networks, and behavior we need approaches that can handle more than two types of data. In this pro- posal we will focus on three key areas. First-building on our previous successes-we will develop new meth- ods that can robustly capture complex relationships between multiple types of data (e.g. genetic codes -> methylation -> brain function -> behavior). Then, we will develop new approaches for the effective use of reliable prior information and provide a set of methods that optimally balance between prior information (model- based) and information readily available from the data (data-driven). And finally, we will combine the strengths of two domains of research, the tractability of data-driven decompositions such as independent component analysis (ICA) with the flexibility of multi-layered learning, to develop an approach we call deep independence networks. This will allow us to capture indirect, but important, relationships among modalities, while also taking advantage of the full data available. The methods we develop will provide a very desirable framework allowing investigators to infer relations in high dimensional data and will provide a much needed set of data analysis tools to the community. We will continue to focus on two important applications where integrating such data is especially important, schizophrenia and addiction, which also share some comorbidity. Focusing on two differ- ent disorders will help us to further generalize the algorithms developed and evaluate shared and distinct as- pects of these two disorders. By combining 1) the extensive data made available by our collaborators, 2) de- velopment of computational approaches for fusing high dimensional data, and 3) the conceptual models we have developed for schizophrenia and alcohol use disorder and results from ongoing studies, we are poised to fill an important gap in the field and produce new tools that have applicability to a wide variety of diseases.
Schizophrenia (SZ) and substance use disorders (SUD) impact countless lives but we have struggled with understanding the complex neuronal and genetic changes associated with these inter-related disorders. We have access to large polymorphism arrays, epigenetic arrays, functional and structural imaging, electroen- cephalography, and behavioral measures, but current methods for analyzing genetic, imaging, and behavioral data are still in their infancy and in many cases researchers rely on simple correlations among a small subset of variables. Motivated by several key questions in schizophrenia and substance use, we will develop a family of methods which will fill an important gap and enable investigators to ask questions like 'which overlapping biological pathways (derived from genetic and epigenetic information) are linked to networks of brain function and structure that jointly predict the degree of substance use?'. The approaches we develop will be of general use and made easily accessible to the research community. 37
Showing the most recent 10 out of 231 publications
