Global features of functional brain networks change with contextual disorder

Abstract

It is known that features of stimuli in the environment affect the strength of functional connectivity in the human brain. However, investigations to date have not converged in determining whether these also impact functional networks' global features, such as modularity strength, number of modules, partition structure, or degree distributions. We hypothesized that one environmental attribute that may strongly impact global features is the temporal regularity of the environment, as prior work indicates that differences in regularity impact regions involved in sensory, attentional and memory processes. We examined this with an fMRI study, in which participants passively listened to tonal series that had identical physical features and differed only in their regularity, as defined by the strength of transition structure between tones. We found that series-regularity induced systematic changes to global features of functional networks, including modularity strength, number of modules, partition structure, and degree distributions. In tandem, we used a novel node-level analysis to determine the extent to which brain regions maintained their within-module connectivity across experimental conditions. This analysis showed that primary sensory regions and those associated with default-mode processes are most likely to maintain their within-module connectivity across conditions, whereas prefrontal regions are least likely to do so. Our work documents a significant capacity for global-level brain network reorganization as a function of context. These findings suggest that modularity and other core, global features, while likely constrained by white-matter structural brain connections, are not completely determined by them.

Fig. 1

Markov processes used to generate the four auditory series. Each node in the transition matrix corresponded to a pure tone. Line weights represent transition probability. The entropy of the Markov processes ranged from random (Markov entropy = 2.0) to highly ordered (Markov entropy = 0.81).

Fig. 2

Computing single node set consistency (SNSC) between the Highly ordered and Random conditions. Per participant in each of these two conditions (1a and 1b), we generated 100 modularity partition solutions (2a and 2b), in which every voxel was in a module. For each condition, the partition with the maximal Q was identified. For these two partitions, we calculated the proportion of within-module voxels that maintained both in Random and Highly ordered, as specified in Eq. (2) (see text) (3). This proportion defined the voxel's SNSC value. The group-level data were the median value for each voxel, across participants.

Fig. 3

Modularity in the four experimental conditions. Independent of graph density, modularity values (Q) patterned similarly across conditions (A) as did mean module number (B). Within each density level, modularity (Q) differed significantly across the 4 conditions. The mean number of modules differed for the 5%, 12%, and 15% edge density thresholds. Error bars here and in subsequent figures capture within-participant error (Loftus and Masson, 1994).

Fig. 4

Partition similarity of functional networks in the Random and Highly ordered conditions. Due to the stochastic nature of the modularity optimization algorithm, between-condition partition similarity is presented via a distribution (gray) and compared to a null distribution constructed from single matrix (black). A: Adjusted Rand Index values. B: Normalized Mutual Information values. For both measures, the null distribution differed significantly from that found for the comparison of the two conditions (both Kolmogorov–Smirnov tests, D = 1, p < .00001).

Fig. 5

Parameter estimates and aggregate plots for degree distributions by condition. A. Log–log plot of the cumulative degree distribution by condition for a randomly chosen participant. Colored lines in each condition depict fit of an exponentially truncated power law to the data. B. Best-fit parameter estimates for an exponentially truncated power law distribution were estimated per participant per condition and analyzed at the group level. These parameters differed significantly across conditions, and the Random condition had the lowest cutoff point, representing a degree distribution that held fewer voxels with high numbers of connections.

Fig. 6

Single node set consistency (SNSC) across the Random and Highly ordered conditions. SNSC is a node-level measure of the proportion of within-module connections that maintain across two partitions. Due to the stochastic nature of the modularity optimization algorithm, between-condition SNSC is presented via a distribution (gray) and compared to a null distribution constructed from single matrix (black). The gray distribution shows SNSC values when evaluating partitions constructed from the Random and Highly ordered conditions of the participant group. The black distribution is the null distribution and is derived from pairs of partitions generated from the same binarized matrix. The difference between distributions was significant (Kolmogorov–Smirnov test, D = .93, p < .00001), indicating considerably less stability in node-level connectivity across conditions compared to the within-condition null distribution.

Fig. 7

Group-median single-node set consistency (SNSC) values at the single-node level. Warmer colors (higher SNSC) indicate voxels that maintain stronger module membership across the Random and Highly ordered conditions. Note that all values were much lower than the upper limit on SNSC values, which approached unity, as seen in Fig. 6. LH: left hemisphere. RH: right hemisphere.

Fig. 8

Areas where average global connectivity decreased as condition disorder increased. A. Significant clusters where average connectivity strength (average functional connectivity of each voxel to every other voxel; Cole et al., 2010) decreased with input disorder (individual voxel p < .01, FWE p < .05). B. Corresponding graphs for clusters in A. Ho: Highly ordered, So: Some order, Ar: Almost random, R: Random. SF/PM: superior frontal/premotor, IT/Crb: inferior temporal/cerebellum, BG–Th: basal ganglia–thalamus, S/MFGrh: superior/middle frontal gyrus right hemisphere, BG/Ins: basal ganglia/insula, PMrh: premotor right hemisphere, Crb: cerebellum (not shown in Panel A), MFGlh: middle frontal gyrus left hemisphere.