Unsupervised network embedding for Drosophila adult brain connectome
Drosophila fly has served as a model system for structural and functional studies in animals for several decades. However, its complex brain anatomy has challenged investigations regarding the structural organization and functional roles in nervous systems. Recently, graph theoretical analyses gradually become standard tools in modeling anatomical connectivity in the brain, while the Drosophila adult brain itself is a heterogeneous graph in nature. In this work, we proposed a network embedding strategy for this heterogeneous brain connectome. Inspired by its hierarchical layout and neuronal weight distribution, the network embedding was produced based on the Relational GCN framework, and the updating for embeddings was conducted with increment and decrement for information flow. The embedding results were optimized under Deep Graph Infomax (DGI), which aims to learn a node encoder that maximizes the mutual information between local patches and the global representation of the graph. Along with the final embedding results, these increment and decrement components can also be used to represent neural properties in the downstream clustering task. We performed KNN methods to cluster the neural embedding, and the clustering results were evaluated with Silhouette Score. Our embedding strategy would provide a new framework for brain network analysis. Meanwhile, clustering analysis can shed light on new neural classes and unreported topological features from the connectivity perspective.