Source-linked AI summary
Navigation of brain networks
Caio Seguin, Martijn P. van den Heuvel, Andrew Zalesky
TL;DR
The paper asks whether navigation offers a parsimonious model of neural communication without requiring global knowledge of network topology. It measures navigation efficiency across mammalian connectomes and finds near-optimal, resource-efficient routing that relates to functional connectivity, while noting unresolved biological plausibility and measurement limits.
Problem
Neural communication in large-scale brain networks lacks a well-established routing model that avoids biologically unrealistic global knowledge of network topology.
Method
The study defines navigation as locally choosing the directly connected node closest to the target and evaluates its efficiency across binary, weighted, and distance-based mammalian connectomes.
Results
Navigation was near-optimally efficient across mammalian connectomes, remained efficient across three network representations, and explained variation in functional connectivity.
Takeaways & Limitations
Brain wiring and spatial embedding are conducive to effective decentralized communication, supporting navigation as a viable neural communication strategy.
Takeaways & Limitations
Navigation assumes nodes know relative spatial positions of neighboring nodes and targets, whose biological plausibility remains to be established.
Abstract
from arXiv · showhide
Understanding the mechanisms of neural communication in large-scale brain networks remains a major goal in neuroscience. We investigated whether navigation is a parsimonious routing model for connectomics. Navigating a network involves progressing to the next node that is closest in distance to a desired destination. We developed a measure to quantify navigation efficiency and found that connectomes in a range of mammalian species (human, mouse and macaque) can be successfully navigated with near-optimal efficiency (>80% of optimal efficiency for typical connection densities). Rewiring network topology or repositioning network nodes resulted in 45%-60% reductions in navigation performance. Specifically, we found that brain networks cannot be progressively rewired (randomized or clusterized) to result in topologies with significantly improved navigation performance. Navigation was also found to: i) promote a resource-efficient distribution of the information traffic load, potentially relieving communication bottlenecks; and, ii) explain significant variation in functional connectivity. Unlike prevalently studied communication strategies in connectomics, navigation does not mandate biologically unrealistic assumptions about global knowledge of network topology. We conclude that the wiring and spatial embedding of brain networks is conducive to effective decentralized communication. Graph-theoretic studies of the connectome should consider measures of network efficiency and centrality that are consistent with decentralized models of neural communication.
Introduction
Neural communication in large-scale brain networks remains poorly understood, motivating navigation as a decentralized routing model for connectomics. The study develops a navigation-efficiency measure and evaluates it across mammalian connectomes, rewired topologies, spatial embeddings, network traffic, and functional connectivity.
- Understanding how neural information is routed through complex white matter networks remains an open systems-neuroscience challenge.
- Shortest-path routing requires network-wide knowledge of topology, an assumption challenged by the decentralized organization of biological systems.
- Navigation routes information by repeatedly selecting the directly connected node closest in distance to the target, but it can fail or produce inefficient paths.
- The study develops a navigation-efficiency measure and applies it to publicly available macaque, mouse, and human connectomes.
- Rewiring connections or repositioning nodes drastically reduces navigation efficiency, while evolving networks toward randomized or clustered topologies yields minimal improvement.
- Compared with shortest-path routing, navigation distributes network resources more uniformly and produces stronger correlations with functional connectivity.
Results
Navigation successfully routes information through mammalian connectomes with near-shortest-path efficiency, while topology and spatial embedding jointly support this performance. Navigation also distributes traffic more broadly and its geometric path lengths explain variation in functional connectivity.
- Navigation reached 89%, 94%, and 96% of human node pairs at 10%, 15%, and 20% connection density, respectively.
- 83% Ewei_R efficiency at 15% connection density showed that navigation approached weighted shortest-path performance in the human connectome.
- 61% median Ewei_R and 79% median Edis_R efficiency on 5-hop paths showed that performance extended beyond directly connected or short-hop pairs.
- Navigation efficiency improved by less than 5% after slight randomization, while clusterization progressively reduced performance, indicating near-optimal topology.
- Navigation centrality was more uniformly distributed than betweenness centrality, with the 50% least connected nodes mediating 26% of navigation paths.
- Navigation path lengths correlated significantly with functional connectivity and predicted it more strongly than shortest-path lengths or direct Euclidean distance in the right hemisphere.
Discussion
The study concludes that connectome wiring and spatial embedding support efficient decentralized navigation, while navigation distributes network resources more uniformly and relates to functional connectivity. Its efficiency remains near-optimal across communication regimes and close to peak across tested topological rewiring, but relies on local spatial information.
- Discussion: Navigation uses network resources more uniformly than shortest paths, potentially reducing concentration of communication traffic.Navigation centrality was more uniformly distributed across nodes than weighted betweenness centrality.
- Discussion: Navigation paths were efficient simultaneously in binary, weighted, and distance-based connectomes, indicating robustness across communication regimes.The three regimes emphasize different attributes of network communication.
- Discussion: 97%–100% EbinR navigation efficiency was observed in dense human and non-human connectomes at 50%–60% connection density.Navigability increased with connection density, while density-matched null networks performed significantly worse.
- Discussion: A progressive regularity–randomness rewiring spectrum improved human navigation performance by only ∼1% for N = 360, placing the brain near peak performance.The analysis did not exhaust other possible connectome attribute dimensions, such as modular–centralized organization.
- Discussion: Navigation path lengths significantly predicted functional connectivity even after controlling for Euclidean distance.The association therefore cannot be entirely attributed to navigation paths approximating physical distance.
- Discussion: Navigation approximates shortest-path efficiency without requiring global knowledge of network topology, supporting decentralized communication models.Alternative diffusion, communicability, and spreading models may assume broadcast signaling across the network, whereas navigation routes information toward a destination.
- Discussion: Navigation assumes that nodes know the relative spatial positions of neighboring nodes and the target, although this is more realistic than global topological knowledge.Further work is needed to establish whether the spatial-information assumption is biologically plausible.
- Discussion: Non-invasive neuroimaging generally cannot directly observe information transfer between distant neural elements, motivating future stimulation or virtual-lesion tests.These approaches could compare navigation and shortest-path communication strategies through local electrophysiological or simulated functional changes.
Materials and Methods
The methods represent connectomes with weighted adjacency and length matrices, transform connection weights into lengths, and evaluate navigation by repeatedly selecting the spatially closest neighbor. Navigation efficiency and path comparisons are computed across node pairs and network representations.
- Materials and Methods: Weighted connectomes are represented by W, where Wij denotes the connection weight between nodes i and j.Weights quantify relationship strength, such as streamline counts or labelled-neuron fractions.
- Materials and Methods: Connection lengths convert weights into travel costs, enabling computation of short communication paths.Strong connections are mapped to short lengths and weak connections to long lengths.
- Materials and Methods: Lwei = −log10(W/max(W)) transforms weighted connections monotonically while attenuating extreme weights.For human streamline-count connectomes, an added +1 prevents the largest weight from mapping to zero length.
- Materials and Methods: Navigation repeatedly advances from the current node to the neighbor with the shortest Euclidean distance to the target.The process stops when the target is reached or when a node is revisited, which marks failure.
- Materials and Methods: Navigation path lengths depend on the connection-length definition, although the navigated paths depend on topology and node positions.The asymmetric navigation path-length matrix was computed by navigating every ordered node pair.
- Materials and Methods: Navigation-to-shortest-path efficiency ratios compare navigation path lengths with shortest paths across binarized and weighted connectomes.The analyses also include distance-based comparisons.
- Materials and Methods: Human, macaque, and mouse datasets were publicly available, and the navigation implementation was released online.The cited implementation is hosted at the authors’ GitHub repository.
Supporting Information: Navigation of brain networks
The supporting document is authored by Caio Seguin, Martijn P. van den Heuvel, and Andrew Zalesky.
- Supporting Information: Navigation of brain networks: Caio Seguin, Martijn P. van den Heuvel, and Andrew Zalesky are listed as the authors.
Overview
The supporting information describes data acquisition, preprocessing, network measures, methods, and supplementary analyses for assessing result robustness and universality.
- Overview: The supporting document details connectivity-data acquisition and preprocessing, network measures, applied methods, and supplementary replication analyses.These analyses are described as assessing the robustness and universality of the main results.
- Overview: The document is identified as arXiv:1801.07938v1, dated 24 January 2018.
Materials and Methods
The study assembled mammalian and human connectomes from structural and functional imaging or tracer data, then evaluated navigation across resolutions, density thresholds, and null network models. Human analyses also tested alternative cortical parcellations and preprocessing choices.
- Macaque and mouse connectomes represented interregional anatomical projections derived from tracer-based datasets.
- Human cortical parcellations varied in resolution and included custom subdivisions of a 68-region atlas plus the Glasser Atlas.
- Subcortical regions were excluded because their voxel-based representations require parcellation methods different from those applied to cortical surface meshes.
- Functional connectivity was computed from resting-state fMRI time series and averaged across the same 75 participants.
- Analyses thresholded connection weights across 5%–60% densities and compared empirical networks with rewired, spatially randomized, progressively randomized, and progressively clusterized null models.
Supplementary analyses
Supplementary analyses examined how navigation performance varies with connection-length definitions, parcellation resolution, shortest-path hop count, node centrality, and functional connectivity. They also assessed path-length transformations and resource distribution.
- Dividing streamline counts by Euclidean distance improved navigation performance, reaching SR = 96% in the tested human structural networks.
- Shortest-path-hop analyses compare navigation efficiency ratios with matching-hop shortest paths while distinguishing successful and failed navigation paths.
- Navigation centrality and binary betweenness were positively correlated (r = 0.54), but navigation centrality was more uniformly distributed across nodes.
- Log-transforming navigation path lengths better differentiated short paths and produced stronger correlations with functional connectivity.
- Supplementary path-structure analyses stratified non-rich-club routes by source–target distance and navigation-path hop count.