Reconstructing Cortical Networks: Case of Directed Graphs with High Level of Reciprocity

The problem of prediction of yet uncharted connections in the large scale network of the cerebral cortex is addressed. Our approach was determined by the fact that the cortical network is highly reciprocal although directed, i.e. the input and output connection patterns of vertices are slightly different. In order to solve the problem of predicting missing connections in the cerebral cortex, we propose a probabilistic method, where vertices are grouped into two clusters based on their outgoing and incoming edges, and the probability of a connection is determined by the cluster affiliations of the vertices involved. Our approach allows accounting for differences in the incoming and outgoing connections, and is free from assumptions about graph properties. The method is general and applicable to any network for which the connectional structure is mapped to a sufficient extent. Our method allows the reconstruction of the original visual cortical network with high accuracy, which was confirmed after comparisons with previous results. For the first time, the effect of extension of the visual cortex was also examined on graph reconstruction after complementing it with the subnetwork of the sensorimotor cortex. This additional connectional information further improved the graph reconstruction. One of our major findings is that knowledge of definitely nonexistent connections may significantly improve the quality of predictions regarding previously uncharted edges as well as the understanding of the large scale cortical organization.