Dynamic reconfiguration of frontal brain networks during executive cognition in humans

Abstract

The brain is an inherently dynamic system, and executive cognition requires dynamically reconfiguring, highly evolving networks of brain regions that interact in complex and transient communication patterns. However, a precise characterization of these reconfiguration processes during cognitive function in humans remains elusive. Here, we use a series of techniques developed in the field of "dynamic network neuroscience" to investigate the dynamics of functional brain networks in 344 healthy subjects during a working-memory challenge (the "n-back" task). In contrast to a control condition, in which dynamic changes in cortical networks were spread evenly across systems, the effortful working-memory condition was characterized by a reconfiguration of frontoparietal and frontotemporal networks. This reconfiguration, which characterizes "network flexibility," employs transient and heterogeneous connectivity between frontal systems, which we refer to as "integration." Frontal integration predicted neuropsychological measures requiring working memory and executive cognition, suggesting that dynamic network reconfiguration between frontal systems supports those functions. Our results characterize dynamic reconfiguration of large-scale distributed neural circuits during executive cognition in humans and have implications for understanding impaired cognitive function in disorders affecting connectivity, such as schizophrenia or dementia.

Keywords: dynamic network; flexibility; frontal cortex; graph theory; working memory.

Fig. 1.

Network reconfiguration during executive function. (A) We use a numerical n-back task consisting of 0-back and 2-back conditions. (B) We define 270 cortical and subcortical regions of interest (36), and (C) extract the mean time course from each region. (D) A sliding window comprising 15 volumes with no gap was applied to regional mean time courses, and for each window we estimated the functional connectivity between pairs of regions using coherence. This procedure resulted in a sequence of 114 time-ordered adjacency matrices. (E) Using a dynamic community detection algorithm (part 1 in panel), we identified network modules in each time window and tracked their evolution over time. (F) By estimating the probability that a brain region changes its allegiance to modules between any two consecutive time windows (part 2 in panel), we observed that whole-brain flexibility oscillated between unitask (2-back or 0-back only) and dual-task (2-back and 0-back in same time window) conditions.

Fig. 2.

Evolving network organization. (A) To determine whether cognitive systems are transiently or consistently recruited during task execution, we construct a modular allegiance matrix T by computing the contingency matrix N for each window: the element N_ij is equal to 1 if nodes i and j are in the same module and is equal to zero otherwise. We sum all contingency matrices for each condition to obtain the modular allegiance matrix T, whose elements T_ij indicate the fraction of time windows in which nodes i and j have been assigned to the same module. We then apply a community detection algorithm to T to obtain a “consensus partition” (25), which represents the common modular structure across all time windows. (B) The modular allegiance matrices for the 0-back condition (Left) and 2-back condition (Right). The letter beneath the block diagonal elements indicates the network module identified in the consensus partition: F, frontal; FP, frontal-parietal; FT, frontotemporal; H, hippocampal; O, occipital; P, parietal; PF, right prefrontal; S, subcortical; SM, somatomotor. (C) A mapping of the frontal modules obtained in B to their brain coordinates for the 0-back (Left) and 2-back condition (Right); labels are as in C. (D) Frontal systems show high integration during 2-back, whereas during 0-back occipital and parietal systems show an equally strong interaction. Error bars indicated SEs of mean over partitions belonging to either to 0-back or 2-back conditions.

Fig. 3.

Role of frontal systems in executive functioning. (A) The flexibility in the frontal cortex during 2-back is positively correlated with 2-back task accuracy, measured in percentage of right answers. (B) Integration, as the contribution to flexibility that is due to between-module reconfiguration, shows an even stronger association with task accuracy. (C and D) Furthermore, frontal integration is related to other cognitive measures as shown by correlation with the performance in the digit span test (backward, measured in number of correctly remembered items) and the performance in the trail-making test B (TMT-B, measured in seconds to task completion).

Fig. S1.

Distribution of flexibility across cognitive systems. (A) The flexibility during 0-back is different across network modules, with modules incorporating frontal nodes showing a higher flexibility than modules lacking frontal nodes. (B) The distribution of flexibility is even more pronounced during the 2-back condition, again with frontal modules showing higher flexibility than nonfrontal modules. Module labels are as follows: F, frontal; FP, frontal-parietal; FT, frontotemporal; H, hippocampal; O, occipital; P, parietal; PF, right prefrontal; S, subcortical; SM, somatomotor.

Fig. S2.

Distribution of integration across cognitive systems. Integration of different cognitive systems during 0-back (Left) and 2-back (Right). Black columns indicate original data; gray columns indicate integration from 1,000 permutations of consensus matrix labels. Error bars indicate SD over 1,000 permutations. Module labels are as follows: F, frontal; FP, frontal-parietal; FT, frontotemporal; H, hippocampal; O, occipital; P, parietal; PF, right prefrontal; S, subcortical; SM, somatomotor.

Fig. S3.

Community structure during 0-back and 2-back. To optimally visualize the differences in community structure between 0-back and 2-back conditions on a node-to-node basis, we have adapted the original node order presented in an earlier figure to the node order published by Power et al. (36).

Fig. S4.

Classical graph metrics. The mean weighted clustering coefficient (red) and pathlength (blue) for functional networks extracted from overlapping time windows. Values are averaged over subjects. Note that each functional networks was normalized by dividing each edge weight by the mean edge weight in the network; this procedure ensured that observed changes in the statistics were not driven by changes in mean edge weight, but by specific changes in the network topology. We observe an oscillatory pattern in these statistics that is consistent with that observed for network flexibility in the main manuscript.

Fig. S5.

Details of n-back paradigm. Stimuli were presented in blocks of either 0-back (Left) or 2-back (Right) conditions, without any additional control condition. In the 0-back condition, subjects were required to press the button on the response box corresponding to the number currently displayed on the presentation screen (red numbers next to screen images on the Left). In the 2-back condition, subjects were required to press the button on the response box corresponding to the number presented two stimuli before the number currently displayed on the presentation screen (red numbers next to screen images on the Right). Each condition block was repeated four times in an interleaved manner as shown on the Bottom.

Fig. S6.

Schematic of circuit dynamics during task execution. (A) At rest, brain functional networks are characterized by network modules, or putative cognitive systems, which contain sets of brain regions that are coherently active with one another (36, 37, 62). (B and C) A cognitive task paradigm may elicit several cognitive processes, during which modules may interact transiently to share information with one another or enable coherent information processing (2, 40). (D) The network dynamics of the task paradigm can be summarized by the flexibility of individual brain regions, which have partnered across modules, and the integration between modules that have been transiently coupled (3, 24).