Refined parcellation of the nervous system by algorithmic detection of hidden features within communities

The nervous system can be represented as a multiscale network comprised by single cells or ensembles that are linked by physical or functional connections. Groups of morphologically and physiologically diverse neurons are wired as connectivity patterns with a certain degree of universality across species and individual variability. Thereby, community detection approaches are often used to characterize how neural units cluster into such densely interconnected groups. However, the communities may possess deeper structural features that remain undetected by current algorithms. We present a scheme for refined parcellation of neuronal networks, by identifying local integrator units (LU) that are contained in network communities. An LU is defined as a connected subnetwork in which all neuronal connections are constrained within this unit, and can be formed for instance by a set of interneurons. Our method uses the Louvain algorithm to detect communities and participation coefficients to discriminate local neurons from global hubs. The sensitivity of the algorithm for discovering LUs with respect to the choice of community detection algorithm and network parameters was tested by simulations of different synthetic networks. The appropriateness of the algorithm for real-world scenarios was demonstrated on weighted and binary Caenorhabditis elegans connectomes. The detected LUs are distinctly localized within the worm body and clearly define functional groups. This approach provides a robust, observer-independent parcellation strategy that is useful for functional structure confirmation and potentially contributes to the current efforts in quantitative whole-brain architectonics of different species as well as the analysis of functional connectivity networks.