Blending generative models with deep learning for multidimensional phenotypic prediction from brain connectivity data

Network science as a discipline has provided us with foundational machinery to study complex relational entities such as social networks, genomics, econometrics etc. The human brain is a complex network that has recently garnered immense interest within the data science community. Connectomics or the study of the underlying connectivity patterns in the brain has become an important field of study for the characterization of various neurological disorders such as Autism, Schizophrenia etc. Such connectomic studies have provided several fundamental insights into its intrinsic organisation and implications on our behavior and health. This thesis proposes a collection of mathematical models that are capable of fusing information from functional and structural connectivity with phenotypic information. Here, functional connectivity is measured by resting state functional MRI (rs-fMRI), while anatomical connectivity is captured using Diffusion Tensor Imaging (DTI). The phenotypic information of interest could refer to continuous measures of behavior or cognition, or may capture levels of impairment in the case of neuropsychiatric disorders. We first develop a joint network optimization framework to predict clinical severity from rs-fMRI connectivity matrices. This model couples two key terms into a unified optimization framework: a generative matrix factorization and a discriminative linear regression model. We demonstrate that the proposed joint inference strategy is successful in generalizing to prediction of impairments in Autism Spectrum Disorder (ASD) when compared with several machine learning, graph theoretic and statistical baselines. At the same time, the model is capable of extracting functional brain biomarkers that are informative of individual measures of clinical severity. We then present two modeling extensions to non-parametric and neural network regression models that are coupled with the same generative framework. Building on these general principles, we extend our framework to incorporate multimodal information from Diffusion Tensor Imaging (DTI) and dynamic functional connectivity. At a high level, our generative matrix factorization now estimates a time-varying functional decomposition. At the same time, it is guided by anatomical connectivity priors in a graph-based regularization setup. This connectivity model is coupled with a deep network that predicts multidimensional clinical characterizations and models the temporal dynamics of the functional scan. This framework allows us to simultaneously explain multiple impairments, isolate stable multi-modal connectivity signatures, and study the evolution of various brain states at rest. Lastly, we shift our focus to end-to-end geometric frameworks. These are designed to characterize the complementarity between functional and structural connectivity data spaces, while using clinical information as a secondary guide. As an alternative to the previous generative framework for functional connectivity, our representation learning scheme of choice is a matrix autoencoder that is crafted to reflect the underlying data geometry. This is coupled with a manifold alignment model that maps from function to structure and a deep network that maps to phenotypic information. We demonstrate that the model reliably recovers structural connectivity patterns across individuals, while robustly extracting predictive yet interpretable brain biomarkers. Finally, we also present a preliminary analytical and experimental exposition on the theoretical aspects of the matrix autoencoder representation