Nine working projects built on officially released Drosophila melanogaster brain data. They go from network science on a 3.5-million-synapse connectome, through machine-learning models trained on its wiring, to FlyLab: a physics-simulated fruit fly whose brains keep training continuously and perform tasks you give them in plain English.
Everything uses reputable primary sources (Janelia HHMI, the Princeton/MRC
FlyWire consortium, EPFL's NeuroMechFly), downloads without a login, and has
been run end-to-end. Every animation below was generated by the code in this
repo (make_gifs.py).
FlyLab: a brain that only has its two antennae steers NeuroMechFly (a micro-CT-based fruit-fly
body) to food in MuJoCo physics. The HUD shows what it smells and the descending commands it sends to its legs.
👉 Watch the brains train live: training dashboard 👉 Run the trained brains yourself: brain playground, which lets you drag the fly and its target, silence neurons or an antenna, test 200 arenas, and download the weights.
- FlyLab: teach a physics-simulated fruit fly (flagship)
- The data
- Quick start
- Connectome analysis projects 01–05
- Machine-learning projects 06–08
- Results at a glance
- Repository layout · Reproducing the animations · Citations & license
🎮 FlyLab — teach a physics-simulated fruit fly (Project 09)
Type "find the food", "go to the light" or "turn left 90". The fly's brain for that skill trains, then drives NeuroMechFly v2 (EPFL; a body built from a micro-CT scan of a real adult Drosophila, with 42 actuated leg joints, leg adhesion and ground contact) in MuJoCo physics through the official FlyGym walking controller.
- 7 skills, 7 brains. Each skill has its own neural network and its own training process, so learning one skill never overwrites another. An earlier single shared brain forgot skills: it fell from 86% to 57% physics success after a long curriculum.
- Continuous training with physics in the loop. Each round runs 100 Evolution-Strategies generations in a surrogate body calibrated against physics, then validates in real MuJoCo. A brain is only replaced when it does better in physics, and every new best is scored on held-out physics arenas never used for selection. Mastery is progressive over three levels (6 → 12 → 24 validation arenas).
- Runs itself. It restarts after reboots, spare CPU polishes optimal brains (a result is kept only if it stays optimal), and every day each brain is re-tested on new random physics arenas. The status table below rewrites itself daily and results are pushed hourly. See autonomous operation.
- Realistic senses. Odor is sampled at the model's measured aristae (0.52 mm ahead of the thorax, 0.38 mm apart), there are two compound eyes, and the gait wobble (0.15 mm and 3.6° per step), measured in physics, is fed back into training.
Each film is the current best brain on an arena it has never seen, rendered automatically as training improves it.
| Walk forward & stop | Turn in place | Go to a location |
|---|---|---|
 |
 |
 |
| Find food by smell | Escape a bad smell | Walk toward light |
 |
 |
 |
| Hide from light | ||
 |
Status on 2026-10-01 (rewritten daily by daily.py; the dashboard is live):
| Skill brain | Rounds | Physics validation | Held-out physics test (95% CI) | Fresh-arena check | Status |
|---|---|---|---|---|---|
| Walk forward | 34 | 100% of 24 | 100% of 32 (89%–100%) | 100% of 16 (2026-10-01) | ✅ optimal |
| Turn in place | 18 | 100% of 24 | 94% of 32 (80%–98%) | 94% of 16 (2026-10-01) | ✅ optimal |
| Go to a location | 36 | 100% of 24 | 100% of 32 (89%–100%) | 100% of 16 (2026-10-01) | ✅ optimal |
| Find food by smell | 65 | 100% of 24 | 100% of 32 (89%–100%) | 100% of 16 (2026-10-01) | ✅ optimal |
| Escape a bad smell | 215 | 96% of 24 | 94% of 32 (80%–98%) | 88% of 16 (2026-10-01) | ✅ optimal |
| Walk toward light | 205 | 100% of 24 | 97% of 32 (84%–99%) | 94% of 16 (2026-10-01) | ✅ optimal |
| Hide from light | 30 | 100% of 24 | 100% of 32 (89%–100%) | 100% of 16 (2026-10-01) | ✅ optimal |
How it got here (full engineering log): each failure was measured in physics and then fixed.
- Physics initially ran at 0.02× real time. Updating the controller at 1 kHz made it ~20× faster with the gait verified unchanged.
- The fly circled food instead of stopping, so the task now rewards dwell time.
- Escaping a smell scored 100% in the surrogate but 0% in physics. The cause was measured gait wobble; the fix was domain randomization plus sensory integration.
- The fly "stood still" by rocking one side backward, so the surrogate now models leg-activity sway.
- Walking toward light fidgeted at an eye-level lamp, so the lamp moved overhead.
- A shared brain forgot skills, so every skill now has its own brain.
python -m venv .venv-sim && .venv-sim\Scripts\pip install -r requirements-sim.txt
cd projects\09_flylab
..\..\.venv-sim\Scripts\python flylab.py # then type: find the food
..\..\.venv-sim\Scripts\python supervise.py # keep every brain training| Dataset | What it is | Scale | Source |
|---|---|---|---|
| hemibrain v1.2 | dense reconstruction of the central brain: every traced neuron + synapse | 21,739 neurons · 3.55 M connections · 14.3 M synapses | Janelia FlyEM (CC BY 4.0) |
| FlyWire FAFB | annotations for a complete adult brain | 139,248 neurons · 8,840 cell types | FlyWire consortium (CC BY 4.0) |
| NeuroMechFly v2 | biomechanical fly body + locomotion controllers (Project 09) | 42 actuated joints, MuJoCo | EPFL NeLy lab, FlyGym (Apache-2.0) |
Provenance, licenses, SHA-256 checksums and citations are in
datasets/README.md.
# from the FruitFlyBrain/ directory: projects 01-08
python -m venv .venv
.venv\Scripts\pip install -r requirements.txt # Windows (macOS/Linux: source .venv/bin/activate)
python download_datasets.py # connectome data, ~275 MB, first run only
.venv\Scripts\python projects\01_connectome_graph_analysis\analyze.py
.venv\Scripts\python projects\02_neurotransmitter_atlas\atlas.py
.venv\Scripts\python projects\03_shortest_path_circuits\trace.py
.venv\Scripts\python projects\04_connectome_neural_network\simulate.py
.venv\Scripts\python projects\05_navis_3d_viewer\view_neurons.py
.venv\Scripts\python projects\06_celltype_classifier\train_classifier.py
.venv\Scripts\python projects\07_neuron_embeddings\embed_and_cluster.py
.venv\Scripts\python projects\08_synapse_link_prediction\link_prediction.py
# project 09 (FlyLab) has its own environment: see the FlyLab section aboveEach script prints a summary and writes its figures and tables to its own
outputs/ folder. Each project's README.md has usage, options and a sanity
check against known fly neuroscience.
01 · Connectome graph analysis · hemibrain
The brain as a directed, weighted network: degree distributions, hubs and connected components. The biggest hubs come out as APL and DPM, the giant mushroom-body neurons known from the literature.
02 · Neurotransmitter & cell-type atlas · FlyWire
A census of a complete brain: 139,248 neurons by super-class and predicted neurotransmitter. The brain is ~62% cholinergic, 18% glutamatergic and 14% GABAergic, matching the published figures.
03 · Synaptic pathway tracer · hemibrain
The strongest multi-synapse routes between cell types (edge cost = 1 / synapses). Kenyon cells → MBONs recovers the mushroom body's learning connection directly, plus routes through the modulatory APL/DPM neurons.
04 · Connectome-constrained neural network · hemibrain
A recurrent network whose weights are the real central-complex wiring (1,968 neurons, 243,743 connections). Stimulating ER ring neurons drives the EL/EPG "compass" neurons, the known navigation pathway.
05 · 3-D neuron morphology viewer · NAVis
Real neuron shapes as a static render and an interactive 3-D HTML (drag to
rotate). --neuprint fetches the top hubs found by Project 01.
06 · Connectivity-based cell-type classifier · supervised
Predicts a neuron's cell-type family from who it connects to alone. RandomForest 88.2% / neural net 87.6% accuracy across 53 families (18,590 neurons), against an 11.3% majority baseline. The animation shows the neural net training epoch by epoch.
07 · Neuron embeddings & unsupervised cell typing · unsupervised
128-d embeddings learned from wiring alone recover cell types without any labels: NMI 0.59, 63% mean cluster purity. The animation shows t-SNE separating the families.
08 · Synapse link prediction · graph ML
Does neuron A synapse onto neuron B? The split is leakage-free (embeddings are learned from training edges only). Gradient boosting reaches 0.971 ROC-AUC / 0.968 PR-AUC on held-out connections. The animation shows the ROC curve sharpening as trees are added.
| # | Project | Headline result |
|---|---|---|
| 01 | Connectome graph analysis | heavy-tailed wiring; top hubs APL (227k synapses) and DPM |
| 02 | Neurotransmitter atlas | 62% cholinergic, 18% glutamate, 14% GABA across 139k neurons |
| 03 | Pathway tracer | strongest KC → MBON learning route recovered |
| 04 | Connectome-constrained dynamics | ER → EPG compass pathway emerges from raw wiring |
| 05 | 3-D morphology | interactive neuron viewer, linked to Project 01's hubs |
| 06 | Cell-type classifier | 88% accuracy, 53 classes (baseline 11%) |
| 07 | Unsupervised cell typing | NMI 0.59 with no labels |
| 08 | Synapse link prediction | ROC-AUC 0.971 on unseen connections |
| 09 | FlyLab | 7 of 7 brains optimal on held-out physics tests; the rest keep training (table above) |
📄 RESUME.md has paste-ready, quantified résumé bullets for all of this.
FruitFlyBrain/
├── README.md · RESUME.md · CITATION.cff · LICENSE
├── requirements.txt # projects 01-08
├── requirements-sim.txt # project 09 (FlyGym / MuJoCo)
├── download_datasets.py # fetches the connectome data
├── make_gifs.py # builds every animation in this README
├── datasets/ # provenance docs (data itself is downloaded)
├── docs/gifs/ # the animations
└── projects/
├── 01_connectome_graph_analysis/ 05_navis_3d_viewer/
├── 02_neurotransmitter_atlas/ 06_celltype_classifier/
├── 03_shortest_path_circuits/ 07_neuron_embeddings/
├── 04_connectome_neural_network/ 08_synapse_link_prediction/
└── 09_flylab/
├── flylab.py # talk to the fly
├── train_brains.py # continuous per-brain training, physics in the loop
├── supervise.py # keeps every brain training, re-films and publishes
├── training/ # per-brain logs, curves and reports (TRAINING.md)
├── state/brains/ # the trained brains (~5 KB each)
└── dashboard/ # the live GitHub Pages dashboard
.venv\Scripts\python make_gifs.py # projects 01-08 + FlyLab training curves
.venv-sim\Scripts\python make_gifs.py flylab_videos # FlyLab brain filmsEvery animation is computed from the real data by each project's own code. Projects 06 and 08 re-run their pipelines and record the model while it trains.
Python 3.12 · pandas 3.0 · numpy 2.5 · networkx 3.6 · matplotlib 3.11 · scipy 1.18 · scikit-learn 1.9 · navis 1.12 · plotly 7.1 · FlyGym 2.1 · MuJoCo 3.9
- FlyWire synapse-level edges: free account on Codex,
or programmatic access via
CAVEclient/fafbseg. - neuPrint queries:
neuprint-python(free token). - FlyLab realism roadmap:
REALISM_AUDIT.md, a fact-checked list of what would make the simulated fly more lifelike.
Please cite the original works when using them:
- Scheffer et al. (2020). A connectome and analysis of the adult Drosophila central brain. eLife 9:e57443.
- Dorkenwald et al. (2024). Neuronal wiring diagram of an adult brain. Nature 634, 124–138.
- Schlegel et al. (2024). Whole-brain annotation and multi-connectome cell typing of Drosophila. Nature 634, 139–152.
- Wang-Chen et al. (2024). NeuroMechFly v2: simulating embodied sensorimotor control in adult Drosophila. Nature Methods 21, 2353–2362.
Full references are in datasets/README.md and CITATION.cff.
- Code & docs in this repo: MIT © 2026 Rohit (ROHITCRAFTSYT).
- Connectome datasets: owned by their creators and used under CC BY 4.0.
- NeuroMechFly / FlyGym: EPFL, Apache-2.0 (installed as a dependency, not redistributed).