Neural Cartography: Methods for Mapping the Structure of Neural Networks

Abstract

Recent advancements in volumetric electron microscopy (vEM) allow neuroscientists to map the nervous system at unprecedented scale and resolution. This offers an opportunity to unravel structural organization, cellular morphology, and connectivity, providing new insights into circuit function. This dissertation presents new tools for scalable analytical solutions and circuit analyses to extract testable biological insights from vEM wiring diagrams. In the first section, I utilize cell-body morphology and connectivity to train a hierarchical model for cell-type prediction in a cubic millimeter vEM dataset of mouse visual cortex. This method bypasses the need for complete cell reconstructions, is adaptable to multiple cell-typing schemes, and is computationally inexpensive. Further, this method produced predictions for nearly 100,000 cells with 91% accuracy compared to expert-labeled ground truth. These predictions, publicly available with a new feature set, can be used for unsupervised search of rare cell-types and demonstrated the surprising sufficiency of the somatic region for cell identification.In the second section, I reconstructed the wiring diagram of tactile sensory neurons in the fly ventral nerve cord to elucidate how spatial information is organized within somatosensory circuits, specifically regarding spatially targeted leg grooming. Using genetic labeling and vEM, I defined the foreleg somatotopic map. Downstream connectivity revealed 60 interneurons receiving substantial synaptic input exclusively from tactile neurons. These interneurons exhibit unique axonal projections, diverse dendritic morphologies, and distinct postsynaptic partners. Optogenetic experiments and kinematic analyses demonstrated that activating distinct interneurons initiates spatially guided grooming strategies consistent with our structurally derived receptive field predictions. From these results I propose a four-layer circuit where interneurons form distinct functional modules, each sampling a portion of the leg to elicit spatially targeted grooming. The third section extends this spatial mapping analysis to chemosensory information from the fly leg, using a similar approach to examine how a distinct circuit structure might also maintain spatial information for a different sensory system within the same body segment. Overall, this dissertation examines neural network maps broadly, quantifying structural variability in cell bodies and connectivity across large populations. It then zooms in on specific sensory circuits, exploring how their structural organization informs our understanding of neural computations.