Connectivity inference for petascale neural circuit reconstruction

Abstract

The reconstruction of neural circuits from electron microscopy (EM) has great promise to improve our understanding of biological and artificial intelligence. Automated recon­struction systems designed to study EM image volumes have been used to process larger datasets in recent years. Most of this work has focused on the reconstruction of neuron morphology, while recent efforts have also produced systems that infer the synaptic con­nectivity between reconstructed cells. Here, we demonstrate developments of this approach through a series of analyses and technological improvements. First, we perform an analysis that is based solely on neuronal morphology in the mammalian retina. We develop statisti­cal tests to validate potential cell types, helping shape our understanding of the information the retina sends to the brain. Next, we develop a machine learning approach for connectiv­ity inference in large EM image volumes, and we apply this approach to help produce the largest connectivity map to date. We then analyze similar semi­automated reconstructions of mouse visual cortex to explore their value with current technology. The first finds that synapse density, mitochondrial density, and dendritic diameter correlate with one another within layer 2/3 cells of the mouse primary visual cortex. A separate analysis finds some evidence that synapse size follows a bimodal distribution, with potential implications for memory stability and neural network organization. Lastly, we build a basic prototype of a sparse inference engine to improve current morphology reconstructions by detecting errors in large EM volumes.