Identifying global features of the mammalian connectome using network science and machine learning

Abstract

The connectome is the full set of connections in the brain, and studying it can show how structure supports function. In practice, however, tract-tracing data are incomplete. Even for macaque and mouse, only part of the cortex has been injected, leaving many missing values. In this project, machine learning was used to fill in those gaps and build complete connectivity matrices. To check if the models learned meaningful wiring rules, a re-prediction step was added: known connections were masked and then predicted again, so the output could be compared to the real data. Network analysis was then applied, focusing on triad motifs, and the results were compared to degree-preserving and distance-based null models. For macaque, the re-predicted networks kept the same motif structure as the original data, while the mouse results were less consistent. When comparing across species, some motif patterns were shared while others diverged. These results suggest that the modelscan recover important structure, but the outcome depends on data quality. The stronger performance on macaque shows the promise of the approach, while the weaker performance on mouse highlights its current limitations.