Brain Geometry and Connectome Data
Relevant source files
The following files were used as context for generating this wiki page:
Brain Geometry and Connectome Data comprises the structural and topological information derived from the MaleCNS v1.0 dataset. This data provides the anatomical basis for the neural models used to operate the simulated fly. The project utilizes actual neuron skeletons, neuropil meshes, and measured synaptic connectivity to represent the central nervous system (CNS).
Sources: README.md, apps/website/public/media/attribution.md
Data Acquisition and Provenance
The project acquires data from the MaleCNS collaboration, which includes HHMI Janelia, University of Cambridge, MRC Laboratory of Molecular Biology, and Google Research. The data consists of measured connectivity, body annotations, soma locations, neuron skeletons, and neuropil surfaces. Scripts automate the download of these assets from Google Cloud Storage while verifying their integrity using SHA-256 hashes.
Sources: scripts/acquire_connectome.py:10-21, apps/website/public/media/attribution.md
Source Files and Verification
| Role | Filename | Description |
|---|---|---|
| Annotations | body-annotations-male-cns-v1.0-minconf-0.5.feather | Metadata for neurons including status and superclass. |
| Transmitters | body-neurotransmitters-male-cns-v1.0.feather | Predicted neurotransmitters for each neuron. |
| Weights | connectome-weights-male-cns-v1.0-minconf-0.5.feather | Synaptic weight counts between neuron pairs. |
Sources: scripts/acquire_connectome.py:12-16, reports/2026-09-09/graph-manifest.json
Connectome Graph Preparation
The system processes the raw MaleCNS tables to create a directed graph representing the connectome. It filters the dataset to retain only neurons with the status Traced. This produced a subset of 165,122 neurons and 25,563,197 edges in the pilot run. The graph is stored as a Compressed Sparse Row (CSR) matrix where rows represent postsynaptic bodies and columns represent presynaptic bodies.
Sources: scripts/prepare_graph.py:53-62, reports/2026-09-09/graph-manifest.json
Graph Processing Flow
The following diagram illustrates the transformation from raw data tables to the final sparse graph artifact.
This flow ensures that only high-confidence, traced neural paths are used for model training. Reachability analysis calculates the steps required to transition from sensory neurons (e.g., ol_sensory, vnc_sensory) to motor neurons (vnc_motor).
Sources: scripts/prepare_graph.py:100-142, reports/2026-09-09/graph-manifest.json
CNS Geometry
CNS Geometry represents the physical structure of the fly's brain and ventral nerve cord (VNC). The acquire_cns_geometry.py script fetches neuron skeletons and neuropil region meshes. It downsamples skeletons using a stride of 4 while preserving branch endpoints to maintain structural fidelity at a lower computational cost.
Sources: scripts/acquire_cns_geometry.py:72-83, reports/2026-09-09-clean-videos/geometry-manifest.json
Geometry Data Components
The geometry is exported in a compressed .npz format containing the following arrays:
- neurite_vertices: 3D coordinates of neuron skeleton points.
- neurite_edges: Connectivity indices for skeleton segments.
- roi_vertices / roi_faces: Mesh data for neuropil compartments (e.g., fullbrain, VNC).
- selected_body_ids: A class-stratified subset of neurons (e.g., 512 neurons) used for visualization.
Sources: scripts/acquire_cns_geometry.py:107-113, reports/2026-09-09-clean-videos/geometry-manifest.json
Neuron Selection Logic
The selection of neurons for rendering and geometry acquisition follows a stratified strategy to ensure representative coverage of the CNS.
This logic prioritizes neurons that exhibit significant activation states in previous simulation runs while maintaining diversity through random seeding.
Sources: scripts/acquire_cns_geometry.py:46-63
Technical Summary
The Brain Geometry and Connectome Data module provides the high-fidelity anatomical framework necessary for neuro-mechanical simulation. By combining verified synaptic connectivity with downsampled physical geometry, the system enables the mapping of model activation states onto biological structures.
Sources: apps/website/public/media/attribution.md, reports/2026-09-09/graph-manifest.json