Connectome structure constrains neural activity across scales

Abstract

Connectomes can be represented at many resolutions, but coarse-graining risks discarding structure relevant to neural dynamics. Here we ask at what resolution a connectome preserves what it predicts about neural activity. In C. elegans, direct synaptic connectivity poorly predicts whether stimulating one neuron evokes a response in another. A parameter-free predictor based on paths through the connectome performs substantially better, with most predictive information carried by two-step paths and by the arrangement rather than the number of connections. Coarse-graining neurons into biological cell classes preserves these predictions despite discarding path structure, whereas random partitions degrade them. Connectivity alone identifies alternative coarse-grainings that preserve prediction down to about nine groups corresponding to nervous-system subsystems. Partitions inferred from newly hatched animals preserve adult predictions, and lineage-based partitions in Drosophila preserve path structure better than random coarse-graining. Functional prediction can therefore survive at resolutions far coarser than those required to preserve structure.