Inferring community architectures of multisensory pathways in Drosophila via unsupervised network embedding
Understanding the complex architecture and functions of neural circuits is central to unraveling the mechanisms of multisensory integration. In this study, we analyzed the structural properties of the Drosophila adult brain to infer community structures within multisensory pathways. We adopt a network embedding method developed by ourselves, the Bidirectional Heterogeneous Graph Neural Network with Random Teleport (BHGNN-RT), designed to generate vector representations of neurons in a directed, heterogeneous brain connectome. This approach takes advantage of both structural connectivity and network heterogeneity features, enabling effective clustering of neurons and revealing hierarchical community architectures in olfactory and broader multisensory systems. We applied BHGNN-RT to the fly brain connectome to examine connectivity-based community organization in major neuronal classes along multisensory pathways, revealing distinct neural groups with unique connectivity patterns in the antennal lobe, lateral horn, mushroom body, and other brain regions. Further analysis showed how different neural groups contribute to the integration of sensory information in olfactory and multisensory systems. We also investigated the bilateral symmetry of the olfactory pathway, shedding light on how sensory signals are processed with ipsilateral and contralateral connections to ensure robust perception. Our findings demonstrate the utility of graph representation learning in analyzing the structural connectivity of complex neural systems. The insights gained from BHGNN-RT provide a deeper understanding of the community architecture in the Drosophila brain and contribute to a broader comprehension of the mechanisms underlying multisensory integration.