3. Whole-brain modelling of low-dimensional manifold modes reveals organizing principle of brain dynamics

COEX 

In recent years, considerable efforts have been invested to reduce the complexity of brain activity recordings through the application of nonlinear dimensionality reduction algorithms. Besides their practical use for data analysis, low dimensional manifold representations emerge as a fundamental principle underlying information processing in neural circuits. This work addresses two interconnected questions concerning the role of manifolds in large-scale brain dynamics. First, we ask whether the neural manifolds exist independently of the dimensionality of the brain activity recordings; second, we investigate whether the interactions between manifold coordinates can be modelled to provide meaningful insights into the dynamics of whole-brain activity during rest and cognition. To tackle these questions, we used deep autoencoders to extract manifolds of reduced dimensionality underlying functional magnetic resonance imaging (fMRI) recordings at different brain parcellations. Next, we constructed computational models of non-linear oscillators coupled according to the effective functional connectivity of the obtained manifolds, taking into account the level of non-equilibrium dynamics quantified by the non-reversibility of the signals. We found that the underlying dimensionality of whole-brain fMRI recordings has an optimal value around 10 regardless of the parcellation. Furthermore, modelling the dynamics of the coupled manifold coordinates allowed us to more faithfully reproduce whole-brain activity during cognition compared to models in the original state space with a higher number of dimensions. We then leveraged the generative capabilities of the model to perform in silico perturbations in the reduced network space. These findings support the key role of manifolds as an organizing principle of brain function at the whole-brain scale, also suggesting that future modelling efforts should strive to address the dynamics of low dimensionality representations.