The most appropriate group to review this material would be a Senior Neuroinformatics and Connectomics Research Consortium. This group would consist of principal investigators in computational neuroscience, high-throughput electron microscopy (EM) specialists, and neuroanatomists focused on human cortical architecture.
This research represents a milestone in human connectomics: the nanoscale reconstruction of a 1 $mm^3$ fragment of human temporal cortex (dataset H01). Utilizing high-throughput serial section electron microscopy (EM), the authors generated 1.4 petabytes of data, encompassing approximately 57,000 cells and 150 million synapses. The study details the computational pipeline—including multiresolution flood-filling networks for segmentation and U-Net classifiers for synapse prediction—required to manage data at this scale. Key findings include a glia-to-neuron ratio of 2:1, the discovery of a bimodal directional orientation in Layer 6 "triangular" neurons, and the identification of rare but potent multisynaptic connections where single axons establish up to 50 synapses with a single target. The H01 dataset and associated analysis tools (Neuroglancer, CREST, VAST) are provided as an open-access resource for the neuroscientific community.
Technical Summary: Human Cerebral Cortex Reconstruction (H01 Dataset)
[0:00] Data Scale and Volume: The study reconstructed 1 $mm^3$ of human temporal cortex, producing a 1.4 petabyte dataset. The volume contains ~57,000 cells, 230 mm of vasculature, and ~150 million synapses.
[1:05] Methodology – Acquisition and Alignment: Tissue was obtained via neurosurgical resection (epilepsy access), rapidly fixed, and sectioned at 33.9 nm. Imaging was performed via multibeam scanning EM at 4x4 nm resolution. Fine-scale alignment utilized optical flow fields to correct for drift and jitter across 5,019 sections.
[2:00] Segmentation and Error Correction: 3D reconstruction employed multiresolution flood-filling networks (FFN). To mitigate merge errors (e.g., axon-dendrite crossovers), the team utilized automated subcompartment classification (axon vs. dendrite) to apply targeted "cuts" in the agglomeration graph.
[2:30] Synaptic Prediction: Automated classifiers identified ~150 million synapses. Post-correction estimates suggest a distribution of 67.1% excitatory and 32.9% inhibitory synapses. Machine learning (ResNet-50) was utilized to distinguish synapse types based on EM ultrastructure.
[3:45] Analytical Tooling: The project released several specialized tools:
CREST: Program for exploring synaptic pathways and connectivity chains.
VAST: Manual voxel painting and skeletonization tool.
[4:20] Cellular Composition and Layering: Neuropil volume breakdown: Unmyelinated axons (40.2%), dendrites (25.8%), and glia (15.5%). Glia outnumber neurons 2:1. Neuronal density is ~16,000/$mm^3$, significantly lower than mouse association cortex.
[5:15] Synaptic Architecture: Excitatory synapse density peaks in Layers 1 and 3; inhibitory density peaks in Layer 1. Pyramidal neurons exhibit compartmentalization (inhibitory inputs on the soma/AIS, excitatory on distal spines), a pattern not observed in interneurons.
[6:00] Layer 6 Triangular Neurons: Analysis of "compass" cells in Layer 6 revealed a bimodal distribution of basal dendrites. These dendrites orient in mirror-symmetrical anterior-posterior directions, suggesting a previously unknown structural organization in deep cortical layers.
[7:00] Rare Multisynaptic Connections: While 96.49% of axonal inputs consist of a single synapse, the study identified rare "strong" connections. Some axons provide >50 synapses to a single partner. These are not incidental (as per Peters' Rule) but represent purposeful, high-weight physiological inputs.
[8:30] Discussion and Future Implications: The study proves the viability of rapid immersion fixation for human connectomics. It acknowledges the caveat of using epileptic tissue but provides a baseline for "engramics"—the study of the physical instantiation of memory and experience in human neural circuits.