Coming up short: Generative network models fail to accurately capture long-range connectivity

Generative network models (GNMs) have been proposed to identify the mechanisms/constraints that shape the organization of the connectome. These models parameterize the formation of interregional connections using a trade-off between connection cost and topological complexity or biophysical similarity. Despite their simplicity, GNMs can generate synthetic networks that capture many topological properties of empirical brain networks. However, current models often fail to capture the topography (i.e., spatial embedding) of many such properties, such as the anatomical location of network hubs. In this study, we investigate a diverse array of GNM formulations and find that none can accurately capture empirical patterns of long-range connectivity. We demonstrate that the spatial embedding of longer-range connections is critical in defining hub locations and that it is precisely these connections that are poorly captured by extant models. We further show how standard measures used for model optimization and evaluation mask these and other differences between synthetic and empirical brain networks, highlighting the need for care when interpreting GNMs and metrics. Overall, our findings demonstrate common failure modes of GNMs, identify why these models do not fully capture brain network organization, and suggest ways the field can move forward to address these challenges.