Predicting Neural Activity from Connectome Embedding Spaces

Describing the relationship between activity and connectivity in neural circuits is fundamental to understanding brain function. The connectome of cortical networks is high-dimensional, whereas neuronal activity is often low-dimensional relative to the number of neurons. Consequently, only a small fraction of the information contained in the connectome is relevant to activity. Can this relevant information be identified? In this study, we propose that the activity of each excitatory neuron in the sensory cortex can be approximated as a point in an embedding space defined by its connectome. Using the MICrONS dataset, which provides millimeter-scale, nanometer-resolution data on proofread and coregistered neurons, we perform a correlation analysis and identify a statistically significant alignment between morphological and functional similarities. Topological analysis reveals that the structured representation spaces of both neuronal activity and connectome share a low-dimensional, hyperbolic geometry with exponential scalability. Based on these findings, we develop a simple linear model to reconstruct in vivo neuronal activity by embedding anatomical affinities using multidimensional scaling (MDS). This approach achieves an explanation ratio of 68% compared to directly using activity similarity, outperforming the 56% obtained using the full connectome. Our results high-light a clear and robust structure-function coupling: although geometry-compatible dimensionality reduction methods discard much of the connectome’s detailed information, they prove more effective in predicting neuronal activity, suggesting that synaptic connections may encode an abstract low-dimensional organization.