A principled approach to community detection in interareal cortical networks

Abstract

Interareal structural connectivity explored by tract-tracing reveals super-dense, weighted, directed, and spatially embedded complex networks. These properties make the extraction and alignment of the community structure of these networks with brain function challenging. Here, we overcome these difficulties using a principled, information theoretic approach that distinguishes connectivity profiles using the Hellinger distance. Applying it to retrograde tract-tracing data in the macaque, we show that the cortex at the interareal level is organized into a nested hierarchy of link-communities alongside a node-community hierarchy. We find that the [1/2]-Renyi divergence of connection profiles, a non-linear transform of the Hellinger metric, follows a Weibull-like distribution and scales linearly with interareal distances, establishing a quantitative expression between functional organization and cortical geometry. Finally, we show that the loop entropy along the hierarchy is maximized at a community description level that is neither excessively detailed, nor non-specific, defining a "Goldilocks" level that we hypothesize, is optimal for system-wide efficient information processing. Author SummaryUnderstanding how functional organization emerges from connectivity alone is a central challenge in neuroscience. The cortex forms a dense network, making it difficult to identify meaningful groups of areas using standard methods. Here, we introduce a principled information-theoretical approach that compares the connectivity patterns of brain areas to uncover their community structure and organization. By first detecting the link communities and then deriving the grouping of areas from it, our method reveals a hierarchical structure based on patterns of reciprocal communication. Applied to the retrograde tract-tracing database of the macaque cortical network, this approach identifies communities that share similar interaction patterns, offering new insights into large-scale brain organization. More broadly, it provides a general framework for analyzing complex networks with dense, directed, and weighted connections.