Biologically-Aware Algorithms for Connectomics

Abstract

With billions of neurons and trillions of synapses in humans, the inner workings of the brain remain one of the great open mysteries of the universe. A complete understanding of the brains of even tiny organisms such as Caenorhabditis elegans remains elusive. The field of connectomics seeks to unravel the mysteries of brains by analyzing their circuitry at a synaptic level. In pursuit of this goal, neuroscientists extract and image brain tissue from various species, trace the neurons through the images, identify all synaptic connections, and construct a wiring diagram. A team of researchers spent a dozen years manually extracting the first nearly complete connectome from an image stack. This early success contained 302 neurons and 5,000 synapses. In the succeeding decades, automated algorithms have complemented and, at times, replaced manual human reconstruction efforts. More recently produced wiring diagrams, or connectomes, contain thousands of neurons with millions of synaptic connections.

Recent advancements in image acquisition using multi-beam electron microscopes enable neuroscientists to capture one terabyte of raw image data every hour. Although these engineering achievements open the door for imaging larger brain volumes from more evolved species, human labor, particularly that which requires expert knowledge, has become the bottleneck. Therefore, researchers have increasingly relied on automated solutions to reconstruct neurons from the raw image data and extract the wiring diagrams. The previous decade has seen a deluge of new algorithms that confront the challenges of converting the raw image data into connectomes for further analysis. Despite the incredible success of these algorithms, there is still room for improvement in accuracy and efficiency.

By guiding algorithm design with existing biological knowledge, we can improve accuracy, reduce complexity, and generate results more faithful to the neuronal data. We demonstrate these principles with four biologically-aware algorithms that confront various problems across the connectomics pipeline. We explicitly tailor our solutions for connectomic data and outperform existing methods that are generally agnostic to the underlying biology. This dissertation considers just a tiny fraction of the numerous challenges when working with connectomics data. However, we believe that biologically-aware algorithms like those presented here can make inroads into tackling a wide range of problems in connectomics.