Constructing an integrated picture of human brain function requires understanding how the effects of molecular and genetic factors propagate upwards, through many intervening layers of structure and interaction, to influence behavioral, psychiatric and cognitive traits. Projects such as the BRAIN Initiative (BI) recognize that building such a picture requires the convergent efforts of experts across genetics, genomics, neuroscience, and clinical studies, and have created resources to aid the integration of data from these disciplines. However, the challenge of combining experimental methods and theoretical models spanning vast length/time scales remains significant. One of the more promising avenues of addressing this challenge is the use of interpretable deep-learning approaches to learn high-dimensional structure inherent in data. By embedding constraints from known biological structure, investigators can relate the models’ internal representations to identifiable factors from neuroscience. This proposal will draw on the extensive resources in BI archives, along with other public resources, to integrate data from genetics, functional genomics, and neuroimaging. Through secondary analysis on this data we will build deep, multilevel polygenic models of high-level traits, such as cognitive, affective and psychiatric traits. We will trace the mechanisms underlying such traits to specific regions, cell types, functional connectivity patterns and structural imaging features. Additionally, by embedding biological structure at intermediate levels (tissue and cell-type gene regulatory networks; structural/functional constraints from MRI data), we will build models that improve on additive heritability measures of polygenic risk. In the process, we will harmonize BI data with other publicly available brain omics and imaging datasets. We will deposit all resources and models into relevant BI archives. The proposal is framed as follows. First, we will combine genetics with genomics-based networks from multiple brain regions and cell types, and develop predictive models of region- and cell-type-specific omics variation. These will be included in an interpretable deep model of cognitive and psychiatric traits (Aim 1). Second, we will learn predictive models of structural and functional imaging features from genetic predictors, which will likewise be embedded in interpretable deep models of high-level traits (Aim 2). Third, an integrated, polygenic model will be built by combining both functional-genomics- and neuroimaging-based features, allowing the impact of both subcomponents to be assessed. Furthermore, we will extend our previous work to develop compression-based interpretability methods, which allow a network to be coarse-grained and interpreted at varying levels of resolution. Such interpretation will include the exploration of subphenotypic structure in psychiatric disorders and interactions between traits (Aim 3). We expect the proposed approach to have wide-ranging implications, including insights into mechanistic underpinnings of brain function, new frameworks for integrative multilevel analysis, and the development of methods and resources for future research.

Although cognitive, behavioral and psychiatric traits all exhibit substantial amounts of heritability, knowledge of the mechanisms by which genetic variation is linked to such traits remains limited, partly because of the difficulty in building models reflecting the many levels of structure in the human brain across which genetic variation is filtered. We propose to develop an interpretable deep-learning framework to allow data across genetics, functional genomics, and neuroimaging domains to be combined in learning deep, multilevel polygenic models of high-level traits, while harmonizing data across BRAIN Initiative and other large-scale resources. Our predictive modeling framework will embed known and learned biological structural constraints, and permit investigators to relate the models’ internal representations to desired levels of analysis, by developing network-compression based coarse-graining and interpretability methods.