Space-Independent Modular Structure of Brain Functional Networks

The brain’s intrinsic organization into functional networks has been assessed using imaging techniques, such as functional magnetic resonance imaging (fMRI). In several recent studies, the dynamic functional connectivity (dFC) of these networks was analyzed using a graph theory approach to extract features that characterize their topology over time. The question arises whether the features captured can be explained by the spatial constraints determined by the brain’s underlying structure[1], or if functional coactivation is to some extent responsible for the patterns found. We address this by investigating modular configuration in resting-state fMRI (rs-fMRI) data through community detection using a spatially informed null model.
A previously described rs-fMRI dataset acquired from 9 healthy subjects at 7T was used[2]. We started by identifying the graph’s nodes through parcellation of the data into 68 brain regions using the Desikan atlas. The edges were then obtained for each time point (TR=1s) and subject, by computing dFC between pairs of brain regions using phase coherence[3], and then by thresholding, while keeping the giant component structure and the main topological characteristics of the network. To analyze the community structure, we applied the Louvain algorithm to the time points for which the topology deviated significantly from the random case, using a rewiring null model. We obtained 6-7 communities with an average modularity Qoriginal = 0.693, which was significantly greater than Qrewiring = 0.297 for this null model, indicating there is a community structure for these networks. To investigate whether this modular configuration can be explained solely by the network’s spatial constraints, we then used a degree-constrained spatial null model[4], for which the modularity was slightly lower than for the networks being analyzed (Qspatial = 0.625). However, since the difference between these two was still statistically significant for most time points, we applied a modified Louvain algorithm[4] that regresses out spatial restrictions. This resulted in an average of 6 communities with modularity Qmodified = 0.071, for these selected time points. In conclusion, although the topology and community structure of the rs-fMRI dFC networks is mainly explained by the spatial embedding, some significant and non-negligible degree of functional specialization of the community structure can still be detected, which is space-independent.
References
[1] J. A. Roberts et al., The contribution of geometry to the human connectome, NeuroImage, 124 (2016) 379-393.
[2] J. Wirsich et al., EEG and fMRI connectomes are reliably related: a simultaneous EEG-fMRI study from 1.5T to 7T, bioRxiv, 10.1101/2020.06.16.154625
[3] J. Cabral et al., Cognitive performance in healthy older adults relates to spontaneous switching between states of functional connectivity during rest, Sci Rep 7, 5135 (2017)
[4] R. Cazabet et al., Enhancing Space-Aware Community Detection Using Degree Constrained Spatial Null Model, CompleNet 2017, Dubrovnik, Croatia, pp.26118 – 55