Abstract

This toolkit provides programmatic access to the FlyWire connectome with fafbseg-py, interoperable objects for navis, and a PyTorch module for morphology/connectivity embeddings. The goal is reproducible, scalable analysis with clear routes to deployment (Docker + CI).

Methods

Data access & transforms

• fafbseg-py to fetch meshes/skeletons/annotations.
• Coordinate transforms handled via library APIs to ensure stable cross-version queries.
• Returned objects are navis-compatible for downstream morphology and graph operations.

Embeddings (PyTorch)

• Construct node/edge features from morphology or connectivity tables.
• Train a shallow model to learn embeddings for similarity/retrieval tasks.
• Provide unit tests and deterministic seeds for reproducibility.

Engineering

• Dockerfile, requirements.txt / environment.yml.
• GitHub Actions for lint/tests.
• Tutorial notebooks convertible to CI-smoke tests.

Results

• End-to-end pull of N neurons + M neuropils with stable transforms.
• Learned embeddings separate neuron families in 2D projections.
• Latency reduced by X% after caching key queries.

Discussion

A unified, reproducible access layer + embeddings unlock faster exploratory analyses and ML baselines. The navis bridge lets researchers pivot between analysis and visualization seamlessly; the PyTorch block sets the stage for contrastive or supervised tasks at scale.

Data & Reproducibility

• Env: conda env create -f environment.yml (or pip install -r requirements.txt)
• Docker: docker build -t flywire-toolkit . → docker run -p 8888:8888 flywire-toolkit
• CI: See badge and workflow in repo
• Notebooks: notebooks/01_fetch.ipynb, 02_embeddings.ipynb

References

• FlyWire + fafbseg‑py docs; navis docs; PyTorch docs.