mondk/fly-brain — FlyWire Brain Flight Sim

A simulation of the sensorimotor flight-control circuitry of the fruit fly (Drosophila melanogaster) based on the real connectome (FlyWire), running LIF neuron models in accordance with the whole-brain study by Shiu et al., Nature 2024 (UC Berkeley team — "simulating the entire fly brain on a laptop"), placed in a 3D environment to observe the brain piloting the fly in a closed-loop manner.

Running

cd D:\opencode-file\fly-brain
python -m http.server 8000
# open browser: http://localhost:8000/viewer/

Requirements: modern browser, 1 GB free RAM. Simulates 30,212 neurons / 272,318 synapses in real-time (60 FPS on a standard laptop; event-driven engine, no memory allocation within the loop, ensuring no lag).

What is REAL (based on published data)

Item	Source
3,794,615 synapse edges (pre, post, neuropil region, synapse count)	FlyWire v630, via Murthy Lab mirror (connections.csv.gz, ~26 MB download)
Neurotransmitter labels per neuron (ACh/GABA/Glut/DA/SER/OCT)	Predictions by Eckstein et al., Cell 2024 (neurons.csv.gz)
Cell type / super-class / side for 127,979 neurons	Schlegel et al., Nature 2024 (classification.csv.gz)
Synapse distribution per neuron across 79 neuropil regions	FlyWire (neuropil_synapse_table.csv.gz)
Weight signs: GABA/Glut = inhibitory, ACh/DA/SER/OCT = excitatory	Convention from Shiu et al., Nature 2024
LIF parameters: V₀=−52 mV, Vth=−45 mV, τm=20 ms, τsyn=5 ms, w=0.275 mV/synapse, refractory period 2.2 ms	Shiu et al. + original code philshiu/Drosophila_brain_model (model.py)
Key neurons present in the data: DNp01 (giant fiber escape, 2 neurons), LC4/LPLC2/LC11/LC13 (looming detection), R1-6/R7-8 (visual), JO-B/JO-E (Johnston's organ/wind sensing), EPG (head direction/compass, 50), PFL3 (central complex steering, 24), Kenyon cells/MBON/DAN (learning–memory), ORN/ALPN (olfactory), DN descending (1303), MN motor	verified directly in classification.csv.gz

Flight circuit comprising 30,212 neurons selected based on anatomical criteria: visual (VIS 10,526) + olfactory (OLF 3,425) + mechanical/wind (MECH 2,779) + gustatory (TASTE 438) + central complex (CX 2,859) + mushroom body (MB 5,681) + lateral horn (737) + descending DN (1,303) + ascending feedback (2,364) + motor (100). Edges retained if ≥5 synapses (standard FlyWire noise filtering). ## What is Model-based / Artificial (read to avoid misunderstanding)

1. 3D neuron positions are estimates, NOT raw microscopy coordinates: x-axis = left/right side of the main neuropil (real data), y/z = PCA based on actual synapse distribution profiles + jitter. Raw EM coordinates are gigabytes in size, exceeding this session's 100 MB budget.
2. Sensory gain: Poisson drive to sensory inputs with 15 mV EPSP/hit (Shiu used a saturating model, f=250). This is an artificial parameter to compensate for the absence of the cell body/VNC and the fact that only a portion of the brain is simulated.
3. Arousal baseline (6–12 Hz) into DN/CX/MB (artificial spontaneous activity—real flies also exhibit spontaneous activity, but this specific figure was manually selected and tuned to achieve a baseline DN rate of ~2.5 Hz).
4. Steering: the optomotor/obstacle-avoidance reflex derived from left/right visual disparity is engineered (acting as a proxy for the HS/VS pathways not yet separated in the subcircuit); actual DN and PFL3 signals are superimposed (assuming ipsilateral alignment). The visual→CX→DN pathway in this simplified LIF model barely propagates signals on its own (consistent with limitations Shiu noted regarding long-range circuits), so one cannot claim that "all behaviors are emergent."
5. Thrust is derived from the average DN level relative to an adaptive baseline (artificial sensory adaptation); the escape burst triggered by DNp01 spiking is emergent from the actual network (LC4/LPLC2→DNp01).
6. The fly model, flight physics, and stimuli (swatter/odor/wind) are illustrative representations. ## Measured Results (Empirical Data)

• Offline Whole-Brain Simulation (scripts/sim_full.py, numpy, 127,979 neurons / 3,794,615 synapses; ~20s real-time for 1s simulation time): 100 Hz gustatory + olfactory stimulation propagates selectively to the central brain and to the descending neuron DNb05 — demonstrating sensorimotor transformation; the network remains stable without runaway excitation (effective E/I balance). See results_full.json.
• Closed-loop Test (scripts/test_closedloop.js, node test_closedloop.js): Odor source approaches → OLF activity rises (13.5 → 44.4 Hz); swatter approaches laterally → ipsilateral visual input spikes (32 → 288 Hz); forcing DNp01 spike → escape response (thrust 7.5); silencing VIS → deactivation of 10,526 neurons.
• During tuning, numerical testing revealed two actual bugs: inverted closing-velocity sign and reversed left/right vectors — both fixed; tests now pass.

Suggested Viewer Experiments

• Click 🪰 Swat Now as the swatter approaches: observe VIS burst → DNp01 → ESCAPE! (mimicking the real giant-fiber reflex: LC4/LPLC2 → DNp01).
• Silence VIS then release the swatter: the fly is "blind" and fails to escape.
• Silence CX: loss of PFL3/E-PG compass input (observe the compass bump signal vanish).
• Disable sugar odor: OLF goes silent; the fly stops tracking the odor.
• "Brain View" Camera: 30k neurons color-coded by membrane potential (red = spiking); track the E-PG compass bump.

Structure

data/        pre-loaded real connectome (~34 MB / 100 MB limit) + results_full.json
scripts/     build_brain.py (brain construction) · sim_full.py (offline whole-brain sim) · test_closedloop.js
viewer/      index.html · sim.js (LIF) · env.js (environment + closed-loop) ·
fly.js (fly + physics) · ui.js (panel + graphs) · main.js ·
brain.json (9.2 MB brain) · brain_stats.json · lib/three.min.js (local, offline OK)

Citations (external works used here)

• Dorkenwald et al., Nature 2024 — Neuronal wiring diagram of an adult brain (FlyWire).
• Schlegel et al., Nature 2024 — Whole-brain annotation and cell typing.
• Eckstein, Bates et al., Cell 2024 — neurotransmitter predictions.
• Shiu et al., Nature 2024 — whole-brain LIF model (code: philshiu/Drosophila_brain_model).
• Murthy Lab — flywire-network-analysis (lightweight v630 data mirror). - Google Research + HHMI Janelia + Cambridge — male CNS connectome (166k neurons), Cell 2026 (Context: The data used in this project is the FlyWire FAFB dataset, as it fits within 100 MB).