Whole-brain modeling of dynamic causal circuits in human cognition using amortized variational inference

Human neuroscience faces a significant challenge in understanding the dynamic mechanisms behind cognition. To address this, researchers have developed a new computational framework called Multivariate Dynamical Systems Identification with Amortized Variational Inference (MDSI-AVI). This innovative approach aims to capture the complex, asymmetric, and context-dependent whole-brain directed interactions, while also accounting for the variability in hemodynamic responses seen in functional magnetic resonance imaging (fMRI) data.

MDSI-AVI employs a unique combination of simulation-based inference through both forward and reverse variational inference. This enables the framework to overcome the limitations of traditional variational methods when dealing with high-dimensional data. By averaging out the uncertainty in hemodynamic response parameters through forward simulation, MDSI-AVI produces well-calibrated posterior estimates of directed connectivity, even when working with networks containing hundreds of nodes.

These estimates scale efficiently and are able to maintain their accuracy despite the complexity of the data.

The framework was applied to data from the Human Connectome Project, which included 728 participants. The results revealed new insights into working memory mechanisms, specifically highlighting the dorsal anterior insula as a critical hub that influences activity throughout the entire brain. The study also demonstrated task-dependent modulation of causal influences, where the salience network actively drives the activity of the frontoparietal network.

This, in turn, differentially influences the default mode and sensorimotor networks, depending on the level of working memory load.

The connections between these different brain networks were found to be task-dependent and could accurately distinguish between various task conditions. Furthermore, these whole-brain causal interactions were shown to predict working memory performance with a high degree of accuracy. The framework's ability to consistently produce reproducible results across different whole-brain parcellations establishes MDSI-AVI as a robust and valuable tool for advancing our understanding of circuit dynamics in cognition and related neurological disorders.