Resolving structural variability in network models and the brain

TL;DR

This study compares empirical brain network data with various synthetic models to identify mechanisms underlying the brain's complex connectivity, revealing that models constrained by anatomical regions best replicate real network properties.

Contribution

The paper introduces a comprehensive comparison of 13 synthetic network models with human brain connectivity data, highlighting the importance of anatomical constraints in modeling brain networks.

Findings

Models constrained by anatomical regions resemble real brain networks.

Single-property models often fail to replicate multiple network features.

Multiple mechanisms likely contribute to brain network development.

Abstract

Large-scale white matter pathways crisscrossing the cortex create a complex pattern of connectivity that underlies human cognitive function. Generative mechanisms for this architecture have been difficult to identify in part because little is known about mechanistic drivers of structured networks. Here we contrast network properties derived from diffusion spectrum imaging data of the human brain with 13 synthetic network models chosen to probe the roles of physical network embedding and temporal network growth. We characterize both the empirical and synthetic networks using familiar diagnostics presented in statistical form, as scatter plots and distributions, to reveal the full range of variability of each measure across scales in the network. We focus on the degree distribution, degree assortativity, hierarchy, topological Rentian scaling, and topological fractal scaling---in addition to several summary statistics, including the mean clustering coefficient, shortest path length, and network diameter. The models are investigated in a progressive, branching sequence, aimed at capturing different elements thought to be important in the brain, and range from simple random and regular networks, to models that incorporate specific growth rules and constraints. We find that synthetic models that constrain the network nodes to be embedded in anatomical brain regions tend to produce distributions that are similar to those extracted from the brain. We also find that network models hardcoded to display one network property do not in general also display a second, suggesting that multiple neurobiological mechanisms might be at play in the development of human brain network architecture. Together, the network models that we develop and employ provide a potentially useful starting point for the statistical inference of brain network structure from neuroimaging data.