Olfactory Pipeline — FlyWire Connectome-Based Chemical Classification

🔗 Try the live 3D brain viewer →

Click the animation to open the live demo. Real FlyWire neuron positions, real connectome weights, live predictions from 4 trained classifiers.

---

🧠 Live 3D brain viewer

An interactive visualization of a fruit fly's olfactory circuit, running entirely in the browser. Every neuron is plotted at its actual 3D position from the FlyWire FAFB reconstruction (139,255 neurons total). Circuit weights come from the real connectome — synapse counts from electron microscopy.

What you can do

• Real odorant mode — pick any of the 71 chemicals from the de Bruyne 2014 electrophysiology dataset and watch the circuit respond
• Gradient slider — smoothly interpolate between a wine prototype and an industrial prototype, watching the brain morph
• Hazard mode — select from explosives, drug precursors, toxic industrial chemicals, and radioactive compounds. Each fires a distinct ORN → PN → KC pattern through the real circuit
• Auto-play — press the play button (or spacebar) to cycle through all samples
• Four analysis tabs — Brain (live predictions), Perception (what the fly smells), State (neural state classifier), Analytics (model performance)

What you see in the 3D scene

Element	Description
Cyan points	915 olfactory receptor neurons (ORNs), clustered in the antennal lobe
Blue points	685 projection neurons (PNs), relaying to the mushroom body
Amber points	5,177 Kenyon cells (KCs), the mushroom body's sparse coding layer
Bright lines	Actual synaptic connections from the connectome, colored by signal strength

The viewer animates per-frame stochastic spiking. Firing probabilities come from the real activity of each odorant, but the exact spike timings are generated, not recorded.

---

What this project does

1. Extracts the olfactory circuit (ORN → PN → Kenyon Cell) from the real FlyWire connectome.
2. Loads real measured ORN responses (71 odorants × 20 sensilla) from the de Bruyne 2014 dataset.
3. Pushes ORN responses through the connectome to produce Kenyon Cell representations.
4. Trains 4 classifiers on both raw ORN features and connectome-derived KC features.
5. Reports honest cross-validated accuracy with shuffled-label controls.
6. Visualizes the entire pipeline in an interactive 3D viewer.

---

Headline result

We ran two evaluations. Both point the same direction.

1. Connectome comparison (5-fold CV on all 71 samples)

Feature set	Best model	Accuracy
Raw ORN (20-D)	SVM	83.0%
Connectome-derived KC (5,177-D)	RF	81.9%
Shuffled-label control	—	~50% (chance)

The connectome did not improve classification. The signal is already present at the receptor level, and the mushroom body's sparse expansion coding did not amplify it on this dataset. An honest negative result.

2. Held-out test (single 80/20 split, 15 samples never seen in training)

Model	Accuracy
LogReg	93.3% (14/15)
SVM	93.3% (14/15)
RF	86.7% (13/15)
MLP	60.0% (9/15)

See DATA.md for what every dataset, column, and number means.

---

Results figures

Model accuracy on the held-out test set

Behavior along the wine → industrial gradient

Synthetic samples generated by generate_dummy.py. Each model's probability of predicting "industrial" per sample — a smooth curve means the model interpolates sensibly between the two class prototypes.

Prediction heatmap (samples × models)

W = wine, I = industrial. Where models disagree, the sample is ambiguous.

Confusion matrices (held-out test set)

Random Forest	SVM	MLP	LogReg

---

How to use this project

Setup

git clone https://github.com/RaoMuhammadUmar/flywire-odor-classifier.git
cd flywire-odor-classifier
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Data (must be downloaded separately)

Raw datasets are not committed (too large). Download into data/ before running the pipeline.

Dataset	Source	Target folder
FlyWire FAFB v783	https://codex.flywire.ai/api/download?dataset=fafb	data/flywire_raw/
Nowotny/de Bruyne 2014	https://data.csiro.au/collection/csiro:10689	data/nowotny/
DoOR 2.0	http://neuro.uni-konstanz.de/DoOR/content/DoOR.php	data/door_raw/

Required FlyWire files: connections_princeton.csv, classification.csv, consolidated_cell_types.csv, coordinates.csv.

Train and evaluate

python train.py
python predict.py data/test_holdout.csv
python plot.py predictions/test_holdout_predictions.csv

Predict on synthetic gradient samples

python generate_dummy.py --out dummy.csv
python predict.py dummy.csv
python plot.py predictions/dummy_predictions.csv

Build and serve the 3D viewer locally

python viz/extra_scents.py
python viz/export_viz_data.py
cd viz && python3 -m http.server 8000

Then open http://localhost:8000.

Or run the whole pipeline end to end:

./demo.sh

---

Repository layout

generate_dummy.py     synthetic ORN samples along wine -> industrial gradient
train.py              train 4 classifiers, save to models/
predict.py            run inference on any CSV, save to predictions/
plot.py               visualize any predictions CSV
demo.sh               runs the whole pipeline end to end

viz/                  3D brain viewer (Three.js)
  export_viz_data.py    builds brain_data.json from outputs/ and models/
  extra_scents.py       literature-derived hazard compound vectors
  index.html            UI shell
  utils.js              color maps, math, textures
  brain.js              neurons, edges, spiking animation
  brain_only.js         no-op (external body removed for clean view)
  panels.js             tabs, prediction, perception, state, analytics
  app.js                scene, camera, orchestrator

DATA.md               what every dataset, column, and number means
WORKFLOW.md           one-shot command reference

data/                 raw data (not committed -- see Data section)
docs/images/          result figures + demo GIF
models/               trained model files (regeneratable)
predictions/          prediction outputs (regeneratable)
outputs/              intermediate pipeline artifacts
pipeline_artifacts/   one-time connectome extraction scripts

---

What is real vs. what is schematic

Real:

• Neuron 3D positions (FlyWire FAFB v783 soma coordinates, 139,255 neurons)
• Circuit weights (real connectome synaptic counts)
• ORN responses (de Bruyne 2014 electrophysiology)
• Model predictions (4 trained scikit-learn classifiers)

Approximated:

• Hazard compound ORN vectors — literature-derived from Marshall et al. 2010 (Chemical Senses). Representative magnitudes, not raw per-trial recordings.
• Radioactive compound vectors — speculative. Flies have no receptor for radiation. Included to demonstrate what the brain would do if such volatile byproducts existed.

Stochastic:

• Per-frame spiking animation. Firing probabilities are derived from mean activity per odorant. The exact spike timings are generated, not measured.

---

Citation

If you use this code, please cite the underlying datasets:

• Dorkenwald et al. (2024). Neuronal wiring diagram of an adult brain. Nature.
• de Bruyne, M., et al. (2014). CSIRO Data Access Portal, csiro:10689.
• Munch & Galizia (2016). DoOR 2.0. Scientific Reports.
• Marshall, B., Warr, C. G., & de Bruyne, M. (2010). Detection of volatile indicators of illicit substances by the olfactory receptors of Drosophila melanogaster. Chemical Senses.