flymsg

Simulate and view in 3D the first complete central nervous system of a male fruit fly: MaleCNS v1.0 (Janelia FlyEM + Google, CC-BY 4.0), with 165,122 traced neurons, 25.6 M connections and 124 M synapses across the brain, optic lobes and ventral nerve cord.

flymsg turns the wiring diagram into a spiking network, checks the simulation against circuits whose function is known from experiments, and streams the real neuron shapes into a browser to replay the activity.

The same pages are published at https://gianlucamazza.github.io/flymsg/.

Gallery

Paused frames from the page flymsg viz writes, 1400×900, adaptive quality off, with the geometry loaded. Shot with scripts/shoot.mjs on desktop OpenGL (GL=gl). ?chrome=0 does not build the control panel and hides the performance readout; the colour legend stays. A replay colours neurons by group unless a frame says otherwise, keeps silent neurons desaturated, and draws the neuropil shells around the somata in docs/live/cns/. An anatomy view draws the shells and does not replay spikes. Rebuild the scenes with flymsg viz --sim LC4_R, --sim pIP10 and --types pIP10 dPR1 dMS9. The page takes ?t=, ?paused, ?color=, ?chrome=, ?neuropils=, ?somata= and ?select=<bodyId>; clicking a neuron lists its partners. Rendering is described in viz.

Each scene also runs live in the browser on the published site: the page streams the neuron surfaces from the public MaleCNS volumes, and the scenes are the committed exports in docs/live/ (rebuild them with scripts/build_live_scenes.py). Around every scene the site draws all 140,024 somata of the volume as faint points coloured by region, so the whole CNS is visible. A desktop GPU helps.

Looming, LC4_R, 120 ms, stimulus on. The pulse lasts 300 ms. The stimulus is the 55 LC4_R neurons. The response fills the brain and the ventral nerve cord. Open live.

Song pathway, pIP10, 150 ms, stimulus on. Both pIP10 neurons are the cyan stimulus. Recruited neurons are magenta. Open live.

Same replay, coloured by the sex annotation. "Potentially …" calls join their class. Magenta is male-specific, and both pIP10 neurons are in it. Yellow is dimorphic, and both dMS9 neurons are in it. Grey has no dimorphism label; both dPR1 neurons are there, and they carry fru+dsx on the other field. Enrichment for the whole cases is in dimorphism. Open live.

Anatomy, no replay. One colour per type for the six neurons, drawn inside all 114 neuropil shells. pIP10 is male-specific, dMS9 is sexually dimorphic, and dPR1 is fru+dsx with no dimorphism label. Open live.

Intervention, LC4_R with both DNp01 silenced. The failed silence check, against the intact replay: TTMn_R still fires. Open the comparison.

Residual partner, DNp02 or DNp06 silenced as well. With DNp01 already silenced, TTMn_R goes from 7 spikes to 0 (DNp02, DNp06). The pre-registered scan named both (v2) and a replication on fresh seeds named them again (v3).

An edge through the eye. An ON edge moves front to back across the right eye under the locked T4/T5 v4 protocol at its strongest lamina gain (configuration 11). Photoreceptors, L1–L3 and the graded medulla glow by their potential, violet where they hyperpolarize, so the lamina's inversion of contrast is visible; spiking neurons of the T4/T5 pathway glow by spikes. A replay of one configuration, not the verdict, which is the locked run's (eye). Open live.

Left eye and both eyes. The same looming drive from the left (LC4_L) or from both sides. MaleCNS is one animal, not a symmetric template: TTMn_L fires about half as often as TTMn_R. Right against left, both eyes.

What it does

• Explore the wiring: annotations and partners of any cell type (info); the strongest chain between two cell types, plus alternative routes (path).
• Simulate: a whole-CNS leaky integrate-and-fire model after Shiu et al. (Nature 2024) (sim). On the same connectome the engine reproduces their published brian2 model within noise.
• Validate: a battery of known circuits with controls, dose-response and latency-order checks (validate). Model choices go through pre-registered rules (calibrate, select-model).
• Check the wiring: one shuffle of who connects to whom, each neuron keeping its degree and its outgoing weights (wiring-null); four pre-registered null families with many seeded draws (nulls). Every run that writes outputs also writes a manifest (commit, dataset hashes, config, seeds, packages, platform).
• Ask questions about sex differences: how much simulated responses run through dimorphic neurons, and what silencing them does (dimorphism); the same model on the female FlyWire brain (--dataset fafb).
• Test causal claims: silencing scans and follow-ups locked before they run, each with a matched null and a stated failure rule (silence-scan, partner-followup), and one measured plasticity rule (mb-plasticity).
• See it: a three.js page streams source-resolution neuron surfaces, skeletons and neuropils from the public data, and replays the simulated activity (viz). The frames are in the gallery.

Results so far

The index lives on the roadmap. Each row is a verdict already recorded on the linked page. Negative results stay in the findings log.

Quick start

Requires uv (Python 3.13) and, for the browser tests, Node ≥ 22 (CI: Node 24).

git clone https://github.com/gianlucamazza/flymsg && cd flymsg
uv sync
uv run flymsg fetch     # ~1.1 GB of public Feather tables into data/raw/, no account needed
uv run flymsg build     # compact to data/*.parquet (~10 s)

uv run flymsg path LC4 TTMn --alt 3
uv run flymsg sim LC4_R --duration 600 --stim-ms 300 --seeds 3 --superclass descending_neuron vnc_motor
uv run flymsg viz --sim LC4_R && python -m http.server -d runs/viz 8000   # open http://localhost:8000

Neurons are addressed by cell type (LC4), instance (LC4_R) or bodyId (10001). The data directory defaults to ./data (--data DIR or FLYMSG_DATA to change it).

Scientific outputs use execution-bound directories and automatic resume; the CLI prints the final path. A clean committed worktree is required. See run integrity for output paths, checkpoints and historical locks.

Commands

The 3D page takes URL options such as ?t=<ms>&paused, ?color=dimorphism, ?detail=, ?adaptive=0, ?neuropils=0|1, ?somata=0 (hide the whole-CNS soma layer), ?budget=<M triangles>, ?bench=<s> and ?xr=1 (experimental WebXR observer) (viz).

How it works

1. Data. The MaleCNS tables list every traced neuron with its annotations (cell type, predicted transmitter, sex-related labels) and every connection with its synapse count. flymsg keeps the 165,122 traced neurons and the connections among them (data).
2. Model. Each neuron is a leaky integrate-and-fire unit. A connection of n synapses adds ±n · w_syn to its target after 1.8 ms. The sign follows the presynaptic transmitter: acetylcholine excites; GABA, glutamate and histamine inhibit; modulators have no fast effect. Stimulated neurons receive Poisson input spikes. All parameters are Shiu et al.'s, except w_syn, which is divided by 1.43, the lower quartile of the MaleCNS/FAFB synapse-density ratio (median 1.81; model).
3. Validation. Each case stimulates a population and checks the expected target: - it fires reliably; - it fires far more than when the same number of random, similar neurons is stimulated; - it responds more to stronger input; - successive stages fire in order; - activity dies out after the stimulus.
4. Rules before results. Criteria, calibrations, model selections and every confirmatory experiment are written down in validation and a versioned, hashed lock in validation/*.yaml before they run; the runner refuses a lock edited without a new version, and a failure is recorded rather than tuned away. Every new lock names the records in validation/evidence.yaml that hold the source, units and uncertainty of its parameters.
5. 3D view. The export holds only metadata and activity. The page reads the public multi-resolution meshes and skeletons directly from Google Cloud Storage, refining detail near the camera within a triangle budget (viz).

Limits

• One animal per dataset. There is no measure of individual variability, and sex differences cannot be separated from reconstruction differences. flymsg audit flags incompletely reconstructed bilateral neurons (3.2 % of types, including MN9_R).
• Wiring only, by default. The default model has no gap junctions, no graded release and no neuromodulation. sim.Gap and sim.Graded exist and stay off unless a lock names them. A spike crosses a junction only when that lock sets spike_mv and spike_ms (model). One optional plasticity rule (KC → MBON11) changes nothing unless called. Transmitters are machine predictions, and 2.6% of synapses carry no sign. What of the brain is in the model is counted in roadmap. The Giant Fiber behaves like a shak-B² mutant lacking its electrical synapse.
• Qualitative rates. Use the model to rank which circuits a stimulus recruits and in what order, not to predict exact firing.
• Few, short cases. The validation cases are mostly 1–2-hop feed-forward chains.

Documentation

• Evidence-gated delivery milestones: S0–S4 scientific gates, H1/H2 independent hardware tracks, issue ownership, acceptance and explicit no-go rules.

A fact has one home. The roadmap is the index of verdicts. A protocol stays in its own page and in validation/*.yaml. A negative result is one line in the findings log. A public source enters provenance when its commit is on main. The changelog is the release note. A present-tense sentence describes the default model today. A dated sentence stays under its date.

Development

uv run pytest            # unit tests on synthetic networks + browser-module tests (node --test)
uv run pytest -m slow    # plus the full validation battery (needs the data) and a live GCS read
uv run ruff check && uv run ruff format --check src tests scripts

• tests/reference_sim.py keeps the previous NumPy simulator as an oracle: the compiled kernel must reproduce its spikes exactly.
• scripts/shiu_reference.py runs the published brian2 model for the engine check. It needs brian2 in its own environment.
• scripts/shoot.mjs screenshots the 3D view headless (GL=gl for the real GPU).
• scripts/vendor-js.sh refreshes the vendored three.js and lil-gui.
• scripts/bench_viewer.py measures the 3D view on the gallery scenes and the heaviest comparisons (any live scene by name) and refuses a loaded machine; scripts/brain_coverage.py prints what of MaleCNS the model uses; scripts/archive_results.py copies cited run outputs into results/; scripts/build_figures.py redraws docs/img/*.svg from them.
• The code carries no lint or formatter suppressions (noqa, fmt: skip, …); a test fails if one appears. A run a doc cites starts from a commit on main (architecture).

Citation

If you use flymsg, cite the data and the model it builds on (see also CITATION.cff):

• MaleCNS: Berg, S. et al. Sexual dimorphism in the complete Drosophila male central nervous system connectome. Cell 189, 5504–5526 (2026). doi:10.1016/j.cell.2026.08.015. CC-BY 4.0.
• Model: Shiu, P. K. et al. A Drosophila computational brain model reveals sensorimotor processing. Nature 634, 210–219 (2024). doi:10.1038/s41586-024-07763-9
• For --dataset fafb: FlyWire (Dorkenwald et al. 2024; Schlegel et al. 2024; Matsliah et al. 2024; Berg et al., the MaleCNS paper above), as listed in data.

License

Code: MIT. The connectome data are not part of this repository; they are downloaded from their public sources under their own licences. Vendored three.js and lil-gui are MIT (src/flymsg/viz_static/vendor/). See NOTICE for what the licence does not cover.

Exact Quest and desktop performance work follows the exact observer decision. Approximate diagnostics do not qualify original geometry or physical XR.