Seven experiments · one connectome · 2026

The fly's wiring diagram, put to work.

Flybrain is a set of experiments built on MaleCNS v1.0, the synapse-level map of a male fruit fly's central nervous system. Some turn the wiring into things you can play: an arena where small flies run on real sub-circuits, a browser studio, a garden seen through a fly's eyes. Others ask hard questions of it and report the answers as measured, including the null results.

Live FlyWorld sub-circuit10,203 neurons · 401,680 edges
loading 2.5 MB of synapses…
sensory cellsinterneuronsdescending → motor
decoded bodystill
DNp01 0.0 Hz
Leaky integrate-and-fire, 1 ms steps, event-driven. The same arrays the Python game steps on a GPU, run here on one browser thread (…). Layout is synaptic depth from the sensory cells, not anatomy. Full panel →
165,122traced neurons in MaleCNS v1.0
25,563,197directed neuron-to-neuron connections
124,025,046synapses between them
CC BY 4.0FlyEM · HHMI Janelia · Cambridge · MRC LMB · Google

Projects

01 — 10 · play, learn, inspect, read
  1. 01FlyWorldPython · panda3d · torchA garden arena where six small flies are driven by a 10,203-neuron escape-and-feeding circuit extracted from the connectome, and you drive a bigger fly to herd them onto fruit. Loom at them and the giant fiber fires; they jump. Stimulate DNp09 and they walk.Play
    brain in your browser
    desktop game
  2. 02FlyWorld StudioReact · three.js · TypeScriptA browser exploration studio: fly a procedural avatar through a small world, open its teaching brain, then switch to a real-data workspace with bounded MaleCNS circuit packages, provenance inspection and native-coordinate anatomy.Play
    runs in the browser
  3. 03FlyVisionPython · pygame · numpyA first-person garden sampled through 1,709 published eye directions from a microCT reference fly, plus a structural atlas that recovers 96.1% of withheld visual column coordinates from wiring alone.Play
    desktop · previews here
  4. 04Connectome language modelPyTorch · preregistered protocolDoes fly wiring help a tiny recurrent network predict text? A 512-neuron subgraph against degree-preserving rewired controls, five seeds, 35 runs, frozen before the held-out test. Answer: no advantage established.Read
    figures & results
  5. 05FlycodesPython · reservoir readout · bounded judgeCan the complete 165,122-neuron graph, used as a fixed reservoir with a trained readout, write code? Trained on 246 curated MBPP examples, scored on LeetCode-style tasks. It scored 0/24, same as its no-recurrence control.Explore
    results platform
  6. 06MaleCNS atlasPython · SQLite · static reportA read-only inspection of the exact 512-neuron, 3,061-edge graph used in the language experiment: verified identity against the published bulk product, contact matrices, partner lookups, and eight native-space skeletons.Explore
    static report
  7. 07Sugar circuit auditPython · neuPrint · Tastekin et al. 2026A reconstruction and audit of the gustatory sensorimotor pathway from a 2026 Cell paper against MaleCNS v1.0: 95 target bodies resolved, 746,910 raw edge weights cross-checked, every discrepancy logged, and an interactive circuit explorer.Explore
    circuit explorer
  8. 08Steering circuitPyTorch · rules fixed in advanceTrained until it stops improving, does a 443-neuron fly steering circuit keep a robot on target better than a direct connection? 260 runs. Not clearly: its loops act as a short memory that cuts obstacle hits when the target is hidden and adds wall hits, and rewired copies do just as well.Read
    figures & results
  9. 09Forward-only trainingPython · node perturbation · pilot gateCan the same circuit learn without backpropagation steadily enough to compare learning rules? A 10-run pilot found results swinging between random starts about four and a half times more than planned for, past the limit set in advance, so the main study stopped itself.Read
    figures & results
  10. 10Fly-inspired searchPython · numpy · DockerFlyHash and the fly Bloom filter against the methods real systems use, the fly's own odour wiring and the standard search benchmark. The filter wins only at memory budgets where nothing works; FlyHash answers about one question for every 200 the best index does.Read
    figures & results

Results

08 — 10 · measured, mostly negative
Slope chart. Post collisions: 13.2% for the circuit and 13.3% for the direct connection with the target visible; 18.2% against 21.4% with it hidden. Wall collisions: 7.4% against 4.5% visible, 25.5% against 22.8% hidden.Slope chart. Post collisions: 13.2% for the circuit and 13.3% for the direct connection with the target visible; 18.2% against 21.4% with it hidden. Wall collisions: 7.4% against 4.5% visible, 25.5% against 22.8% hidden.
The clearest finding of projects 08 to 10: the fly circuit's loops give a simulated robot a short memory, which helps it miss posts while its target is hidden and makes it hit walls more often, by about the same 2.8 points whether the target is hidden or not. With the target hidden the two effects roughly cancel out, and the circuit is not clearly better than a direct connection with no loops.

Data: stage1_headroom.md · stage1b_occlusion.md · all results and the four verdicts →

How the work was done

protocols · ledgers · gates

Each project started with a written protocol and ended with a status ledger. Learning experiments were preregistered in the local sense: an analysis plan hashed before training, development selection on validation data only, a frozen configuration, and a single explicit command that opened the test split. Where a comparison came out flat, the page says so.

The three latest experiments (08 to 10) went further: each was preregistered with its analysis code frozen and fingerprinted, and with gates decided in advance, the results that would count as a win, as ruling an effect out, or as a reason to stop. Two of them ended at such a gate. That discipline, more than any single number, is what this work establishes.

Every number on these pages is copied from a saved artifact in the repository: a run manifest, an audit JSON, a figure's sidecar. Nothing is re-derived for the website. The boundaries are stated too: synapse counts are structural weights, not physiological ones; the arena flies are artificial models constrained by real wiring, not emulated animals; the first-person garden is sampled through a female reference eye that has not been registered to the male connectome.

One dataset, one specimen

All seven projects read the same three files from the MaleCNS v1.0 release: body annotations, body neurotransmitter predictions and the traced-to-traced connection table at confidence ≥ 0.5. The release is pinned by file hash in every project's manifest. The browser demos on this site carry bounded, pre-extracted subsets; the raw tables (about 1.1 GB) stay on the machine that built them.

male-cns:v1.0 · UUID 4b2087c0fbe046bfaf0d60bc970e3e5d · connectome-weights-male-cns-v1.0-minconf-0.5.feather