Real MaleCNS wiring. Toy dynamics. Buttons are current injectors.
FLYBOARD is an arcade window into the full 166,700-neuron male fruit-fly central nervous system — the MaleCNS v1.0 connectome released by Janelia FlyEM. Six arcade buttons inject simulated current directly into the actual resolved cell populations of the real connectome, whose 25.58M directed synapses all run live on your GPU-free CPU. The glow on the brain is not an animation: it is the voltage/spike field of a leaky-integrator simulation crossing that real wiring.
| neurons | 166,700 whole-CNS cells, typed and 3D-positioned |
| synapses | 25,582,938 directed synapse counts (input-normalized per postsynaptic cell) |
| data | MaleCNS v1.0 (Dorkenwald et al., Cell 2026) — Janelia FlyEM + collaborators |
| runtime | ~4x real-time on the full net · ~113x in --lite |
| stack | Python 3.12+ · pygame-ce · numpy · scipy · numba · MIT |
- Screenshots
- What the buttons do
- Physics
- Quick start
- Options
- Controls
- Requirements
- Performance
- Repository layout
- Inspiration & provenance
- License
 |
 |
| Idle — untouched MaleCNS wiring sits dark, near-critical. | FOOD — taste cells (SNta*) are injected; the wave is real voltage propagation. |
 |
 |
| MATE — the courtship circuit (pIP1 et al.) lights up. | ORGASM — "FAKE REWARD BURST", bounded to a 1%-of-net dopamine subset. |
 |
 |
| DEATH — the whole brain freezes. Not biological death. | RESET — clears death, back to near-critical idle. |
MaleCNS measured wiring. Toy dynamics. Buttons are current injectors. Not feelings. Not a human orgasm. Not biological death.
The footer of every frame says exactly that. The connectome is real; the dynamics are a toy — a leaky integrator tuned to sit near-critical (dark at rest, localized measured waves on demand) because the same real matrix is globally bistable at canonical settings (see Physics). There is no camera module anywhere in the codebase.
Buttons inject a constant I_button into the resolved MaleCNS cell IDs for a
HOLD window (default 250 ms). The glowing cloud is the voltage field.
| Button | Key | Drives these real MaleCNS cells | Effect |
|---|---|---|---|
| FOOD | 1 |
taste cells, SNta* family prime (2799 cells, 57 types) |
sweet/gustatory burst |
| PAIN | 2 |
nociceptor class, includes ppk receptor cells (257 cells) | aversive localized wave |
| LICK | 3 |
proboscis extension motor neurons (68 cells) | reflexive feeding motor burst |
| MATE | 4 |
courtship circuitry incl. pIP1 (3165 cells) | courtship ensemble lights |
| ORGASM | 5 |
the real dopaminergic population (4443 cells) | FAKE REWARD BURST — capped, see below |
| DEATH | d |
— | freezes every voltage to 0; the whole brain goes dark |
| RESET | r |
— | clears death, restores near-critical idle |
Cell counts come from data/out/hits.json, written after every cache load, with
requested_types / found_count / missing_types / proxy_used per button. If a
button resolves to zero exact cells (e.g. on the LITE subsample) it still fires as an
orange PROXY, driving a superclass subset instead.
MaleCNS idle sits SIM_BG below threshold, and a large drive can ignite the entire
CNS. The reward burst is therefore bounded by construction: ORGASM drives at most
cap × n reward cells (cap = 0.01, i.e. ~1,667 of the DA population), enforced in
Sim.inject(..., bounded_extra=True) and reported live as reward_driven_fraction.
Light-up score = (spikes during hold + 300 ms) − (rolling 1 s baseline), counting
only cells that beat their baseline by ≥ 1 spike. Rendered per type and per
superclass, so a burst's identity is legible as real fly types (LC10a, ORN_DA1,
TmY21, …) rather than "something lit up".
Per 20 ms tick:
v <- exp(-dt/0.1) · v + SIM_W_GAIN · W·spikes + SIM_BG + noise + retina + I_button
spike if v >= 1: v = 0
W is the untouched MaleCNS weight table (synapse counts, column-normalized per
postsynaptic cell). Shipped constants: SIM_W_GAIN = 1.0, SIM_BG = 0.110,
noise 0.06.
Why not the "obvious" stronger drive? At the canonical 1.5 / 0.180, the real
25.58M-edge matrix is globally bistable: rest sits at v_ss = 0.18/(1 − e^(−0.2)) ≈ 0.993,
just 0.007 V under threshold — any single spike (even a noise blip) snowballs into a
permanent whole-CNS strobe of ~166k cells/tick. The shipped constants run the same
untouched wiring in the near-critical regime: dark idle, button bursts are measured,
localized waves, and the leaderboard means something.
./run-flyboard bootstraps a Python venv (.venv/) and installs requirements on
first use, so a fresh machine needs only Python 3.12+ and ~2.5 GB of disk.
git clone git@github.com:NullLabTests/flybrain.git
cd flybrain
./run-flyboard fetch # download ~1.3 GB of Janelia MaleCNS v1.0 tables + soma shards
./run-flyboard build # build data/out/flyboard_graph.npz (166,700 cells) + manifest
./run-flyboard # full 166k net, arcade UIAlready have the cache? fetch and build skip work that is done and report it.
No MaleCNS cache at all? ./run-flyboard --lite falls back to a synthetic fixture so
the UI always opens.
./run-flyboard [--lite] [--source SEL] [--step] [--noise F] [--hold-ms MS]
[--drive F] [--seed N] [--cite] [--build-only] [--headless]
| Option | Meaning |
|---|---|
--lite |
4096-cell LITE BRAIN subsample spanning all superclasses (~113x real-time) |
--source SEL |
eye-input pane: synth:idle (default) · synth:pattern · image:PATH · video:PATH |
--step |
force STEP mode — burst on button, then frozen lit cloud |
--noise F |
per-tick noise sigma (default 0.06; raise for a twitchier net) |
--hold-ms MS |
button injection hold window in ms (default 250) |
--drive F |
DRIVE fraction 0..1 (default 0.5); also the +/− keys and panel slider |
--seed N |
RNG seed (defaults to a fixed seed per process) |
--cite |
print paper + data citations and exit |
--build-only |
build the graph cache and exit (no UI) |
--headless |
run the UI on a dummy video driver (used by tests / screenshot tooling) |
| Input | Action |
|---|---|
| 1–5 / d / r | FOOD / PAIN / LICK / MATE / ORGASM / DEATH / RESET |
| mouse click on panel button | same as the key |
+ / − or the DRIVE slider |
injection strength ±10% |
| p / STEP | STEP mode toggle |
| s / SLICE | hide the optic lobes |
| drag | rotate the 3D soma cloud |
| mouse wheel | zoom |
| click a glowing cell | pin it: type · superclass · side (last few shown) |
| Esc / q | quit |
Python 3.12+. Core: numpy, scipy, pygame-ce, numba, PyYAML, pyarrow,
pandas. Data tooling: neuroglancer + tensorstore (soma reading), imageio
(optional, only for --source video:). Everything installs automatically into
.venv/ on first run-flyboard invocation; a system Python with scientific wheels
(apt/system-site-packages) keeps the first build fast.
- Full net (166,700 cells): ~4x real-time on a mid-range laptop; the whole UI (render, meters, leaderboard) draws from pure numpy + numba — no GPU needed.
- LITE (4,096 cells): ~113x real-time — bursts resolve in well under a second.
- If real-time ever drops below 0.25x, the UI auto-engages STEP mode (burst on button, then frozen lit cloud) so the arcade never lags out.
flyboard/ python package
__init__.py shipped constants, footer disclaimer, ORGASM cap
build.py Janelia feather tables -> CSR graph (real, safe numbers)
soma.py neuroglancer AnnotationReader -> measured soma positions
data.py fetch / status of the remote Janelia cache
graph.py graph loading, LITE subsample, region codes, colors
resolve.py button -> MaleCNS type -> cell IDs (+ hits.json, proxies)
sim.py leaky-integrator physics, hold + measure windows
source.py eye-input pane (synth / image / video) [no camera]
render.py software 3D point-cloud renderer (pure numpy)
stats.py "who lit up" leaderboard
ui.py 3-pane pygame UI
cite.py paper + data citations
tests/ headless spec tests, including scripted mouse clicks
scripts/ headless screenshot gallery generator
presets.yaml button -> type/selector definitions
docs/screens/ README screenshots (generated, not hand-drawn)
FLYBOARD is a small love letter to the 2026 MaleCNS v1.0 release:
Dorkenwald, S., et al. (2026). Sexual dimorphism in the complete Drosophila male central nervous system connectome. Cell, 30(11), 3090–3090. Data: Janelia Research Campus / Google Research FlyEM, Cambridge, and collaborators.
- The anatomy is real. Every resolved button population is a typed MaleCNS cell
group: taste
SNta*cells, ppk nociceptors, proboscis motor neurons, pIP1 courtship neurons, the dopamine projection set. - Soma positions are measured. The MaleCNS soma-points shards are read through the neuroglancer annotation reader — no hand-rolled binary parsing.
- The arcade idea is borrowed with love. Reward buttons that light a brain nod to the classic "pleasure button" thought experiment (Olds & Milner, 1954) — and to every demo where the glow is honestly computed rather than faked.
The fly deserves better than a strobe. FLYBOARD keeps the real wiring at its own critical edge so a press shows you which corner of the MaleCNS actually reacts.
MIT. The FLYBOARD code is MIT-licensed; the MaleCNS v1.0 data remains copyright its respective owners (Janelia FlyEM et al.), whose terms apply to the data itself.