Learning the wiring rules of a mammalian cortical column

Abstract

Characterization of neural circuits architecture typically relies on measurable neuronal features such as morphology, molecular identity, and spatial location. While generative models lever-aging these properties have proven accurate, they remain constrained by available measurements and our assumptions regarding the prospective features. Here, we present an alternative approach using representational learning and use it to model the circuitry of a column of the mouse primary visual cortex. Our framework learns jointly low-dimensional embeddings of neurons in an abstract feature space alongside wiring rules that predict synaptic connectivity. These embedding-based models accurately predict individual synapses, connectivity degrees, and network motif statistics -- outperforming standard generative models that depend on detailed cell-type classifications -- using only a handful of embedding dimensions and wiring rules. Crucially, the learned representations prove interpretable, recapitulating cortical depth, cell type, and dendritic morphology. The resulting wiring blueprint is both simple and biologically meaningful, suggesting that cortical connectivity follows surprisingly parsimonious logic. This framework offers a general and exportable tool for learning minimal generative models of connectomes.