Heritability and Cognitive Relevance of Structural Brain Controllability

Abstract

Cognition and behavior are thought to emerge from the connections and interactions among brain regions. The precise nature of these relationships remains elusive. Here we use tools provided by network control theory to determine how the structural connectivity profile of brain regions may shape individual variation in cognition. In a cohort of healthy young adults (n = 1066), we computed two fundamental brain regional control patterns, average and modal controllability, which index the degree of influence of a region over others. We first established that regional brain controllability measures were both reproducible and heritable. Regions with controllability profiles theoretically conducive to facilitating multiple cognitive operations were over-represented in higher-order resting-state networks. Finally, variation in regional controllability accounted for about 50% of interindividual variability in multiple cognitive domains. We conclude that controllability is a biologically plausible property of the structural connectome and provides a mechanistic explanation for how brain structural architecture may influence cognitive functions.

Keywords: cognition; controllability; diffusion magnetic resonance imaging; structural connectivity.

Figure 1

Overview of the workflow. (a) We analyzed diffusion magnetic resonance imaging data obtained from 1066 healthy participants of the HCP. (b) Fiber tractography was performed in each participant. The structural connectome was characterized using a parcellation of 512 brain regions. (c) A weighted structural connectivity matrix of size 512 × 512 was constructed representing the streamline counts between pairs of regions. (d) Average (AC) and modal controllability (MC) at each brain region were computed for each participant. (e) Test–retest reliability of regional AC and MC in a dataset of 44 HCP participants that had repeat scans was assessed using the ICC. (f) The heritability of regional AC and MC was computed. (g) Brain regions were clustered based on AC and MC measures. (h) Multivariate PLSs was used to quantify the association between measures of regional controllability and multidomain cognitive measures.

Figure 2

Regional average (AC) and modal controllability (MC). The ranked mean value of the AC of each of the 512 brain regions in the (a) entire sample (n = 1066) and in the (b) baseline and (c) repeat scans of a reproducibility subsample (n = 44). The ranked mean value of the MC of each of the 512 brain regions in the (d) entire sample (n = 1066) and in the (e) baseline and (f) repeat scans of a reproducibility subsample (n = 44). The ranked controllability values are projected onto the cortical surface for ease of visualization. L: Left. Additional information is provided in Supplementary Figure S2 and Tables S4 and S5.

Figure 3

Heritability of structural brain controllability. Heritability of (a) Average Controllability and (b) Modal Controllability for each of the 512 brain regions. L: Left. Additional details are provided in Supplementary Figure S5 and Tables S8 and S9.

Figure 4

Regional controllability profiles and cognitive systems. (a) Dendrogram resulting from hierarchical clustering of 512 brain regions based on their AC and MC. (b) A two-cluster solution was supported by the DB criterion. (c) Matrix of the controllability measures for the two-cluster solution (red and green). Each row is a brain region and each column is a measure. (d) Violin plots of nodal AC and MC of the two different clusters. (e) Radar plots showing the percentage of brain regions in each cluster localized in the cognitive systems. AUD, auditory network; CEN, central executive network; DMN, default mode network; SAL, salience network; SMN, somatosensory network; VIS, visual network. L: Left.

Figure 5

Regional controllability and cognitive function. AC (panels a–d): The PLSs model identified two significant LVs LV1 and LV2 which, respectively, accounts for 50% and 19% of the covariance between AC and cognition. For each LV, (panels a and c) we show the correlations between AC scores and cognitive variables and (panels b and d) we map the bootstrap ratios of the weights of AC of the corresponding regions. MC (panels e–h): The PLS model for MC identified two significant LVs LV1 and LV2 which, respectively, account for 48% and 15% of covariance between MC and cognition. For each LV, (panels e and g) we show the correlations between MC scores and cognitive variables and (panels f and e) we map the bootstrap ratios of the weights of MC of the corresponding regions. In all models, a region with positive bootstrap ratio contributes positively to the controllability-cognition covariation, while a region with a negative bootstrap ratio contributes negatively to the controllability-cognition covariation. Bootstrap ratios for all LVs were thresholded at bootstrap ratios| > 3, P < 0.01. The names of the cognitive variables as coded in the HCP database are provided in Supplementary Table S3. L: left.