Anatomical and Functional Gradients Shape Dynamic Functional Connectivity in the Human Brain

Xiaolu Kong, Ru Kong, Csaba Orbán, Peng Wang, Shaoshi Zhang, Kevin Anderson, Avram J. Holmes, John D. Murray, Gustavo Deco, Martijn P. van den Heuvel, B.t. Thomas YeobioRxiv 2021 · DOI

TL;DR

This paper introduces a novel computational model of the human brain that accurately captures dynamic functional connectivity (FC) by incorporating anatomical and functional gradients. This model reveals that the brain exhibits multi-stable states, with sensory-motor regions playing a key role in switching between these states, and suggests a link between these dynamics and the brain's underlying gene expression patterns.

What Problem Does This Paper Address?

Understanding how the complex patterns of spontaneous brain activity arise from the underlying neural circuitry is a fundamental challenge in neuroscience. While previous computational models have successfully replicated static functional connectivity (FC) – the average synchrony between brain regions over minutes – they have largely failed to capture the dynamic, moment-to-moment fluctuations in this connectivity, known as time-varying FC dynamics. This limitation hinders our ability to understand how the brain transitions between different functional states and how these dynamics relate to brain organization and function. This paper addresses this gap by developing a more sophisticated model that accounts for regional variations in local circuit properties, informed by anatomical and functional brain gradients.

Methods

The core of this work is the development of a Parametric Mean Field Model (pMFM). This model builds upon existing dynamic mean field models (MFMs) by incorporating spatially heterogeneous local circuit parameters. Instead of assuming uniform properties across the cortex, the pMFM allows parameters like recurrent connection strength, external input, and noise amplitude to vary across brain regions. Crucially, these variations are parameterized as linear combinations of two key neuroanatomical gradients: the T1w/T2w myelin map (an anatomical measure) and the principal resting-state FC gradient (a functional measure).

The model's parameters were optimized using a Covariance Matrix Adaptation Evolution Strategy (CMA-ES) algorithm. This optimization aimed to minimize the discrepancy between simulated and empirical static FC (measured by Pearson's correlation) and time-varying FC dynamics (measured by the Kolmogorov-Smirnov distance). The model was trained and validated on data from the Human Connectome Project (HCP).

To investigate the drivers of dynamic FC, the authors employed causal perturbation analysis. They manipulated specific regions within the simulated model to observe the impact on the system's state transitions.

Datasets

The study utilized data from the Human Connectome Project (HCP) S1200 release. This dataset includes resting-state fMRI (rs-fMRI), diffusion MRI, and T1w/T2w structural MRI data from 1052 participants. The data was randomly divided into training (N=351), validation (N=350), and test (N=351) sets.

Additionally, the study leveraged the Allen Human Brain Atlas (AHBA) dataset, which provides gene expression data from postmortem human brain samples. This dataset was used to investigate the relationship between functional connectivity dynamics and molecular markers.

Key Insights and Ideas

  • Anatomical and Functional Gradients are Crucial: The study demonstrates that simply having spatially heterogeneous parameters isn't enough; these heterogeneities must be informed by anatomical (myelin) and functional (FC gradient) information to accurately model brain dynamics.
  • Brain Exhibits Multi-Stability: Both empirical and simulated data show evidence of distinct, recurring brain states, characterized by different patterns of functional connectivity. This suggests the brain isn't just a static network but dynamically switches between configurations.
  • Sensory-Motor Regions Drive State Transitions: Causal manipulation experiments revealed that perturbing sensory-motor regions can reliably "kick" the system from an incoherent state to a more coherent one, highlighting their critical role in driving these dynamic shifts.
  • Gene Expression Links to Dynamics: The spatial distribution of sensory-motor drivers (regions most involved in FC dynamics) aligns with specific gene expression patterns, particularly those related to parvalbumin and somatostatin interneurons. This suggests a potential link between the brain's molecular makeup and its dynamic functional behavior.

Key Results

  • The pMFM, incorporating both anatomical and functional gradients, generated significantly more realistic static FC and FCD compared to previous models.
  • Empirical and simulated FC dynamics showed remarkably similar sharp transitions, indicating the presence of multiple attractors (stable states).
  • Time-varying regional fMRI amplitude was strongly correlated with multi-stability in FC dynamics, with higher amplitude during coherent states and lower amplitude during incoherent states.
  • Causal perturbation of sensory-motor regions led to a significant transition towards a more coherent FC state (p = 6e-14).
  • The spatial distribution of sensory-motor drivers showed strong correlations with PVALB-SST gene expression (r=0.72 empirical, r=0.65 simulated) and the first principal component of gene expression.

Limitations and Future Work

  • Potential Degeneracy in Parameters: The study notes that recurrent connection strength might exhibit potential degeneracies, meaning different parameter combinations could lead to similar model behavior. This could explain some contradictions in existing literature.
  • Specificity of Gene Expression: While PVALB-SST showed a strong correlation, the authors acknowledge that this association might be embedded within a broader pattern of gene expression variation. Further research is needed to fully disentangle the specific contributions of different gene expression patterns.
  • Future Work: The authors suggest that future work could explore the role of these dynamics in cognitive function and neurological disorders, given the link between gene expression and brain states. The development of more sophisticated models that can capture these complex interactions is also a promising avenue.

Summary updated

Abstract

Abstract Large-scale biophysical circuit models can provide mechanistic insights into the fundamental micro-scale and macro-scale properties of brain organization that shape complex patterns of spontaneous brain activity. By allowing local synaptic properties to vary across brain regions, recent large-scale circuit models have demonstrated better fit to empirical observations, such as inter-regional synchrony averaged over several minutes, i.e. static functional connectivity (FC). However, most previous models do not capture how inter-regional synchrony patterns vary over timescales of seconds, i.e., time-varying FC dynamics. Here we developed a spatially-heterogeneous large-scale dynamical circuit model that allowed for variation in local circuit properties across the human cortex. We showed that parameterizing local circuit properties with both anatomical and functional gradients was necessary for generating realistic static and dynamical properties of resting-state fMRI activity. Furthermore, empirical and simulated FC dynamics demonstrated remarkably similar sharp transitions in FC patterns, suggesting the existence of multiple attractors. We found that time-varying regional fMRI amplitude tracked multi-stability in FC dynamics. Causal manipulation of the large-scale circuit model suggested that sensory-motor regions were a driver of FC dynamics. Finally, the spatial distribution of sensory-motor drivers matched the principal gradient of gene expression that encompassed certain interneuron classes, suggesting that heterogeneity in excitation-inhibition balance might shape multi-stability in FC dynamics.
BibTeX
@misc{kong2021anatomical,
  author = {Xiaolu Kong and Ru Kong and Csaba Orbán and Peng Wang and Shaoshi Zhang and Kevin Anderson and Avram J. Holmes and John D. Murray and Gustavo Deco and Martijn P. van den Heuvel and B.T. Thomas Yeo},
  title = {{Anatomical and Functional Gradients Shape Dynamic Functional Connectivity in the Human Brain}},
  year = {2021},
  doi = {10.1101/2021.03.15.435361},
  howpublished = {bioRxiv},
}
Figure