Structural covariance networks are coupled to expression of genes enriched in supragranular layers of the human cortex

Abstract

Complex network topology is characteristic of many biological systems, including anatomical and functional brain networks (connectomes). Here, we first constructed a structural covariance network from MRI measures of cortical thickness on 296 healthy volunteers, aged 14-24 years. Next, we designed a new algorithm for matching sample locations from the Allen Brain Atlas to the nodes of the SCN. Subsequently we used this to define, transcriptomic brain networks by estimating gene co-expression between pairs of cortical regions. Finally, we explored the hypothesis that transcriptional networks and structural MRI connectomes are coupled. A transcriptional brain network (TBN) and a structural covariance network (SCN) were correlated across connection weights and showed qualitatively similar complex topological properties: assortativity, small-worldness, modularity, and a rich-club. In both networks, the weight of an edge was inversely related to the anatomical (Euclidean) distance between regions. There were differences between networks in degree and distance distributions: the transcriptional network had a less fat-tailed degree distribution and a less positively skewed distance distribution than the SCN. However, cortical areas connected to each other within modules of the SCN had significantly higher levels of whole genome co-expression than expected by chance. Nodes connected in the SCN had especially high levels of expression and co-expression of a human supragranular enriched (HSE) gene set that has been specifically located to supragranular layers of human cerebral cortex and is known to be important for large-scale, long-distance cortico-cortical connectivity. This coupling of brain transcriptome and connectome topologies was largely but not entirely accounted for by the common constraint of physical distance on both networks.

Keywords: Allen Human Brain Atlas; Cortical thickness; Gene expression; Structural brain network; Transcriptomic brain network.

Fig. 1

A schematic overview of data analysis. A: Cortical thickness was estimated from R1-weighted MRI scans on 296 healthy young participants. The left hemisphere (LH) was sub-divided into regional nodes by a fine-grained parcellation of 152 nodes based on the Desikan-Killiany atlas. B: Reflection of AIBS gene expression profiles sampled in the right hemisphere (RH, yellow dots) onto the LH (red dots) to increase sample density. C: Regional gene expression profiles from the AIBS Human Brain Atlas were mapped to the same cortical parcellation template to estimate regional expression profiles and a co-expression matrix. The gene expression matrix was thresholded to construct a binary graph which had a modular community structure. D: Modules of the structural covariance network, and cytoarchitectonic classes defined by von Economo, were used to test the hypothesis that whole genome co-expression and HSE gene co-expression were greater between nodes in the same topological module or cytoarchitectonic class.

Fig. 2

Global topology of gene co-expression (left) and structural covariance networks (right). A: Gene co-expression matrix and structural covariance matrix identically ordered in alignment with the modular community structure of the transcriptional network. B: Global topological metrics estimated in TBN, SCN and comparable random networks: C_p = clustering coefficient; L_p = path length; E_loc = local efficiency; E_glob = global efficiency; σ = small-world; a = assortativity; Q = modularity. Error bars represent standard deviations. C: Top 10% of nodes with highest degree (hub nodes) D: Rich club coefficient curves for TBN, SCN and random networks.

Fig. 3

Nodal topology and connection distance of structural covariance network (SCN) and transcriptional brain network (TBN). A. (left) Degree distribution of both networks (solid lines) and a comparable random network (dashed line); (right) connection distance distribution of both networks (solid lines) and a comparable random network (dashed line). B. (left) Effect of inter-regional distance on structural covariance; (middle) effect of inter-regional distance on gene co-expression; and (right) gene co-expression versus structural covariance. C. (left) Degree versus connection distance in the SCN; (middle) degree versus connection distance in the TBN; and (right) nodal degree in TBN versus nodal degree in SCN.

Fig. 4

Community structure of the structural covariance network and transcriptomic brain network. A: Modular decomposition of the whole genome transcriptional brain network (TBN; 7 modules; left); modular decomposition of the structural covariance network (SCN; 9 modules; middle); modular decomposition of the transcriptional brain network restricted to HSE genes (8 modules, right); and two alluvial diagrams showing how regions were aligned to the same or different modules in the SCN and the two transcriptional networks;. B: (left) Distribution of whole genome co-expression between regions assigned to the same random modules compared to the whole genome co-expression between regions assigned to the same modules of the empirical SCN; (right) co-expression of the whole genome and the HSE gene set within each module of the SCN, within each cytoarchitectonic class of the von Economo (VE) atlas, and within comparable random modules. Error bars represent standard deviations.

Fig. 5

Structural covariance and HSE gene co-expression. A: HSE gene expression (top) and SCN nodal degree (bottom). B: HSE gene expression in different cytoarchitectonic classes defined by the von Economo atlas. C: Nodal degree in structural covariance network versus HSE gene expression. D: Mean nodal degree in different cytoarchitectonic classes defined by the von Economo atlas. E: Co-expression (whole genome, blue; or HSE genes only, red) versus structural covariance. Dashed vertical line indicates the threshold value of structural covariance used to define a binary graph of edges and non-edges. F: Correlations between structural covariance and (i) whole genome co-expression (scalar value represented as a dashed blue vertical line), (ii) HSE gene co-expression (scalar value represented as a dashed red vertical line), (iii) co-expression of each of the 5917 gene ontologies defined in the MSigDB collection (represented in grey as a probability density). Each value of the distribution represents the proportion of MSigDB gene sets showing the corresponding correlation with SCN. G: Co-expression (whole genome, blue; HSE genes only, red; and each of the 5917 gene ontologies, grey distribution) for SCN edges. H: Co-expression (whole genome, blue; HSE genes only, red and each of the 5917 gene ontologies, grey distribution) for SCN non edges (unconnected regions).