Understanding the structure of neural connectomes with generative kernel embeddings

Recent efforts achived to completely map the connectivity of invertebrate brains, termed connectomes, at the scale of single neurons and synapses. Such connectome are high-dimensional, so identifying functional roles of neural circuit structures in a whole connectome by exhaustive screening is in general unfeasible. In order to characterize their peculiar structural features and function, it is crucial to extract relevant low-dimensional representations of these. Network embedding provides a powerful framework for connectome analysis by account for neurons' spatial embedding and tendency to cluster. Addressing these needs while including the connectomes traits (heterogeneous, asymmetrical, and weighted nature) remains a significant challenge

We propose an embedding approach tailored to model the structural constraints of neural connectomes. We provides a tractable likelihood for the synaptic connections between each neurons pairs based on a dual-space latent embedding with learnable distance kernels. It makes possible to: 1) extract a low-dimensional latent space; 2) account for synaptic weights and asymmetries by using two spaces for afferent and efferent connections; 3) learn the most appropriate kernel and embedding dimension; 4) identify specific latent space features to uncover hidden structural features; 5) generate artificial connectomes with realistic structural features. We validate our model on synthetic data and on the full adult connectome of the Drophila megalonaster