An interactive explorer of the Drosophila visual–motor connectome.
▲ Real EM-traced neuron morphology from the MaleCNS connectome. Click to open the live site, or watch the full-quality recording.
How can visual information travel through the fruit-fly nervous system towards neurons involved in movement control? This project traces one structural pathway — from the R1–R6 photoreceptors to DNg13, a descending neuron of the locomotor control population — through the open MaleCNS connectome, and presents it as interactive 3D neuron reconstructions, a weighted connectivity graph and evidence-linked information about every neuron.
Everything shown is a structural account of wiring: which neurons connect to which, and how strongly. It does not simulate neural firing and does not predict behaviour.
| View | What it does |
|---|---|
| 3D connectome | Real EM-traced neuron skeletons, coloured by pathway stage, inside a soma point-cloud of the whole CNS. Click a neuron to inspect it. |
| Journey | A step-by-step animated walk from retina → lamina → medulla → lobula → central brain → descending neuron, with the 3D view following along. |
| Circuit | A simplified weighted graph of the pathway; arrow thickness is the fraction of the target's input supplied by that source. |
| Routes | Ranked alternative routes, path lengths, bottleneck neurons, and a node-removal experiment showing how structural routes change. |
| Explore | Pick any of 30 senses (filterable by modality) and any of 480 movement neurons and get the real route, its strength and how it ranks — draw it in the 3D view, play it neuron by neuron, and see which senses share a gateway into that neuron. |
| Context | Whether the route is stronger than chance: rank against all descending neurons, against all sensory modalities, against weight-shuffled graphs, plus a threshold-robustness sweep. |
| Sex | The whole analysis re-run on the female FlyWire connectome, with the female brain viewable in 3D: which connections have a female counterpart, how their strengths compare, which routes are conserved, and the null test replicated in the female brain. |
| Neuron detail | Cell type, class, connection strengths, predicted neurotransmitter, sex status, and an evidence label on every claim, with links to neuPrint. |
Existing connectome explorers will return a path between almost any two neurons, because a dense recurrent network links nearly everything within a few hops. So "a path exists" is weak evidence — and this project measures how weak.
The R1–R6 → DNg13 route is real, reproducible and stable across thresholds. Whether it is special depends on what you compare it against, and the comparisons do not agree:
- DNg13 ranks 312 of 480 descending neurons for connection strength from the photoreceptors — below median.
- R1–R6 ranks 191 of 333 sensory types into DNg13 — whose strongest structural sensory inputs are mechanosensory, gustatory and olfactory, not visual.
- Against randomised wiring (500 runs each), 76% of globally shuffled graphs beat the route — but only 6.8% when the shuffle preserves each neuron's own input profile and randomises only which partner supplies which share.
That disagreement is the substantive result. The two null models preserve different properties of the network and therefore answer different questions. Under the structure-preserving null, 6.8% of randomised networks performed at least as well as the observed route, making it suggestive but short of conventional significance. Under the global weight shuffle, 76.2% did. Reporting both shows how strongly the conclusion depends on the assumptions built into the null model.
The same ranking independently surfaces DNp01 (the Giant Fiber) and the other known visually-driven descending neurons at the top — about 140× stronger than DNg13 — which is a useful check that the method works.
The pathway also is not structurally identical across the two datasets, and this is computed ratherthan read off annotations. Re-running the identical analysis on the female FlyWire connectome: 25 of 31 connections have a female counterpart, and all 6 that don't involve only LoVP92 and VES200m — exactly the two male-specific neurons. The strongest male route (via the male-specific LoVP92) does not exist in the female brain, which reaches DNg13 another way. The null result replicates too — DNg13 ranks 264/443 female descending neurons, below median as in the male.
Full write-up: docs/methods-and-findings.md.
The deployed site is at https://sight-to-action.netlify.app. To run it locally:
npm install --prefix web
npm run dev --prefix webThen open http://localhost:5173. The app reads pre-computed JSON files, so it needs no backend.
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txtDownload the connectome tables (~560 MB, public, no API token):
BASE=https://storage.googleapis.com/flyem-male-cns/v1.0/connectome-data/flat-connectome
curl -o data/raw/body-annotations-male-cns-v1.0-minconf-0.5.feather $BASE/body-annotations-male-cns-v1.0-minconf-0.5.feather
curl -o data/raw/body-neurotransmitters-male-cns-v1.0.feather $BASE/body-neurotransmitters-male-cns-v1.0.feather
curl -o data/raw/connectome-weights-male-cns-v1.0-minconf-0.5-significant-only.feather $BASE/connectome-weights-male-cns-v1.0-minconf-0.5-significant-only.featherOptionally add the female connectome for the sex comparison (~880 MB):
mkdir -p data/raw/flywire
curl -L -o data/raw/flywire/proofread_connections_783.feather \
"https://zenodo.org/records/10676866/files/proofread_connections_783.feather?download=1"
curl -L -o data/raw/flywire/flywire_annotations.tsv \
"https://raw.githubusercontent.com/flyconnectome/flywire_annotations/main/supplemental_files/Supplemental_file1_neuron_annotations.tsv"The build skips the female comparison if these are absent.
Then run the notebook, which regenerates every file the app consumes (skeletons are fetched from the public bucket on first run and cached):
jupyter nbconvert --to notebook --execute --inplace notebooks/01_extract_pathway.ipynbThe app copies data/processed/*.json into place automatically whenever you
run npm run dev or npm run build, so there is no manual sync step.
The site is a static build with no backend. On Netlify the settings come from
netlify.toml — base directory web, build npm run build,
publish web/dist. The analysis outputs are committed under data/processed/
and copied into the bundle at build time, so a deploy needs no extra steps.
python -m pytest tests/ -m "not needs_data" # fast, no data download needed
python -m pytest tests/ # includes checks against the real connectomeThe suite pins down what the route ranking actually computes (relative weights sum to one per target, score is their product along a path, hop limits hold), and cross-checks the fast route finder used by the Explore tab against the exact k-shortest-paths implementation used everywhere else. That equivalence had previously only been verified by hand — and when it was made automatic it immediately caught a real bug, described in the commit history.
src/sight_to_action/
data.py loaders for the MaleCNS flat tables
analysis.py type-level graph, route ranking, bottlenecks, removal experiment
nulls.py null models: rank vs other DNs/senses, weight-shuffled graphs
explorer.py precomputes every sense -> every descending neuron for the app
female.py rebuilds the same graph on FlyWire and compares male vs female
pipeline.py assembles the pathway dataset (evidence + sex labels)
skeletons.py SWC download, parsing and decimation
build.py one reproducible build of every artefact
notebooks/ the analysis, executed with outputs
tests/ pytest suite; synthetic fixtures, no data download required
web/ React + Three.js (R3F) + Cytoscape.js application
docs/ methods and findings write-up
data/processed/ small derived JSON consumed by the app (committed)
data/raw/ downloaded connectome tables and skeletons (gitignored)
Body-to-body connectivity is collapsed to a type-level directed graph. Each
edge carries the total synapse count and a relative weight — the fraction
of the target type's input coming from that source — so that routes are not
dominated by simply large cell types. Routes are ranked by the product of
relative weights along the path (a shortest path under −log cost), searched
within a corridor of types that can lie on a ≤5-hop path. Bottlenecks are
types shared by many alternative routes; the removal experiment deletes a
type and re-runs the search.
Male: MaleCNS v1.0, CC-BY, produced by HHMI Janelia, the Cambridge Drosophila Connectomics Group, Google Research and collaborators — https://male-cns.janelia.org/download/.
Female: FlyWire/FAFB whole-brain connectome, snapshot 783, CC-BY — connectivity from Zenodo 10676866, annotations from flyconnectome/flywire_annotations (Schlegel et al., Nature 2024; Dorkenwald et al., Nature 2024).
- Validated R1–R6 → DNg13 pathway extracted from real connectivity
- Interactive 3D EM-traced neuron skeletons
- Animated step-by-step journey through the pathway
- Route ranking, path lengths, alternative routes, bottleneck analysis
- Structural node-removal experiment
- Male/female comparison computed on the female FlyWire connectome
- Statistical null models and threshold-robustness analysis
- Interactive explorer over all 30 senses x 480 descending neurons
- Any route drawn and played back in 3D at real soma positions
- Male and female brains both viewable in 3D, with routes drawn in either
- Four-level evidence labelling with links to source data
- Reproducible notebook, documented provenance, methods write-up
- Public deployment
- Short demonstration video
This is stated plainly because the commit history shows an AI co-author trailer, and I would rather be direct about this.
What I brought. The choice of project and question; the decision to trace a specific sensory-to-motor pathway rather than build another connectome browser; the scope, the endpoints (R1–R6 → DNg13) and the requirement that every claim be labelled by how well it is evidenced. I directed each round of work, judged the output, and rejected what was not good enough. The 3D view also went through several rewrites on my feedback, and I found the bugs that led to real fixes.
How AI assisted. Wrote most of the code, and proposed the analysis methods and then reviewed and iterated with me. It also carried out the data engineering: locating the public datasets, the download pipeline, the SWC decimation and the web application.
What I checked. That the endpoints and cell types exist in the data; that the neurotransmitter predictions match known biology (R1–R6 histaminergic, L1 glutamatergic) as a sanity check on the pipeline; that the reported findings follow from the outputs; and that the caveats are stated. The test suite exists precisely so these claims do not rest on trust — writing it immediately exposed a bug affecting 17% of routes in the Explore tab.
What I would not claim. The science here is a reproduction, not a discovery. The visual pathway it recovers is textbook, the sexual dimorphism in this dataset is published, and the tooling category is well served by neuPrint, Codex and Virtual Fly Brain. What I think is genuinely uncommon is the combination: a route, a statistical test of whether that route is remarkable, the same analysis repeated on a second connectome, and an explicit evidence label on every statement.
Code is MIT (see LICENSE). The connectome datasets are the work of others and remain under their own CC-BY terms, which require attribution to the original authors — see the licence file for details.