Interweaving structure and function in human brain networks

The human brain is a complex organ. Recent progress in neuroimaging has enabled the mapping of large-scale brain regions and their connections. In parallel, advances in the study of complex networks and systems have impacted diverse research areas. The field of Network Neuroscience leverages these developments to model the brain as a network where nodes and edges are estimated from neuroimaging data. Structural networks summarize the anatomical links between delimited brain regions, and functional networks represent the statistical links between the physiological activity of the same regions. Structural and functional connectivity form distinct and complementary components of the complex system that is the brain. In this thesis, we first use machine learning to illustrate the complementarity of structural and functional connectivity in the detection of schizophrenia in brain networks. Secondly, we propose that observed functional connectivity can be explained by a controllable linear dynamics spreading on the structural network. We derive a model in which different sets of brain regions controlling the dynamics produce coactivation patterns corresponding to different functional states. This model provides a principled way to identify potential control centers in the brain. The identification of control regions is then applied in two case studies where we show that they significantly differ i) between schizophrenic patients and healthy controls and ii) during meditation with respect to resting-state. Next, we focus on the individual-level variability of functional connectivity. We build on previous work that applied data-driven methods to extract the individual fingerprint of functional networks. We show that applying degree-normalization to functional networks modulates the influence of strongly connected nodes and systematically improves several fingerprinting metrics. Eventually, we explore an information-theoretic measure called transfer entropy in order to detect directed functional interactions between brain regions. We show on a bivariate synthetic example that transfer entropy is able to capture time-varying and delayed interactions in non-stationary signals. In empirical data, we illustrate the challenge of interpreting the results provided by the estimation of transfer entropy.