NeuroFly

A research sandbox for testing how AI could detect, predict and compensate for functional decline in Drosophila neurodegeneration models.

Current public demo uses a deterministic synthetic model. NF-001 is an animated agent driven by experiment state; it is not a living organism or a biological brain.
Run the sandbox → Experiment registry
NF-001 LIVE
DROSOPHILA MELANOGASTER · 139,255 neurons · 15.1M synapses
Waiting for a stimulus
day 0days of life (simulated)
0%neurons lost (structural damage)
—reach of the last stimulus vs a healthy brain
—descending signal vs a healthy brain

Loading the real FlyWire map in the background…

9,000 real FlyWire neurons at their positions, seen from the front. Each stimulus lights them up hop by hop; ringed points are the descending neurons that move the fly (DNa01 and DNa02 steer, MDN walks backward).

NF-001 / LIVE CHAMBER
neurofly@lab:~$ boot --agent NF-001
✓ connectome engine nf001.1 (FlyWire v783) · random walk when idle
✓ chamber boundaries active
state: wandering
input: none
signal: baseline
age: day 0.0
brain: no neurons lost
engine: checking the connectome engine…
CHAMBER CONTROLS i 0/14
SENSOR FIELD
BOUNDARY
runtime active 00:00:00
3research questions
100%reproducible synthetic runs
0clinical claims
APIexperiment-first architecture
BRAIN DAMAGE · NF-001.2

Watch the wiring wear out

Everything the fly senses in the chamber (a wall it bumps, food it smells, a lamp it sees, or the auto stimuli) seeds real sensory neurons in the FlyWire adult brain connectome (139,255 neurons, 15.1M connections). The brain map at the top of the page shows where that activity goes, with the damage counters under it. As the fly ages it loses neurons, and the same stimulus reaches less of its brain.

Simulated ageing (an assumption, not a measurement): while the page is open the fly ages (1 day per minute by default) and loses neurons along the curve lost = 60% × (1 − exp(−(day/60)³)), in a fixed pseudo-random order. Lost neurons stop firing and relaying, so the same stimulus reaches less of the brain and the fly moves less. The fly dies when 55% of its neurons are lost (about day 82). Brain damage is reported two ways: neurons lost, and how much of a healthy brain's response the same stimulus still produces. This is a toy model of neuron loss, not of any specific disease.

LIFE LOG

The life of NF-001

Every connectome-driven move is logged in this browser (up to 300 events) and charted below. Nothing is sent anywhere; export the log as JSON or reset it at any time.

—days of life
0stimuli received
0body lengths travelled
0 / 0left / right turns
0backward retreats
0 / 0meals: food / dung
0peak neurons active in one hop
Ageing and brain damageNeurons lost and reach of each stimulus vs a healthy brain, by day of life
Descending-neuron activity per stimulusMean accumulated activity per neuron, log scale
TrajectoryPosition in the chamber after each move
Propagation of the last stimulusNeurons active at each hop
Where the activity wentLast stimulus, by FlyWire super class (share of total)
Behaviour repertoireActions over the whole life
Latest events
TimeSourceStimulusStrengthActionTurnBackwardSeededPeak active
No connectome events yet.

Model: linear-threshold propagation over structural wiring with assumed transmitter signs (GABA and glutamate inhibitory), not a spiking simulation, a neural recording or a living fly. The motor mapping is a documented heuristic (nf001-motor-map-0.1). Data: FlyWire Consortium, adult Drosophila connectome release 783 (DOI 10.5281/zenodo.10676866), CC BY-NC 4.0; synapses summed across neuropils. Non-commercial research use.

EXPERIMENT REGISTRY

Questions before claims

simulation early-detection

Early Neurodegeneration Detection

Can longitudinal behavioral signals reveal a synthetic disease trajectory before a conventional threshold is crossed?

locomotionturning_entropysleep_fragmentationstimulus_latency
simulation prospective-risk

Prospective Progression Prediction

Can an early behavioral window predict later synthetic functional decline?

baseline_deviationfuture_decline_scoretime_to_threshold
simulation adaptive-compensation

AI Adaptive Compensation

Can a closed-loop controller reduce task errors in a synthetic impaired agent without changing the disease model?

task_successpath_efficiencydecision_errorsassistance_cost
LIVE SYNTHETIC SANDBOX

Run a reproducible trajectory

RESEARCH ROADMAP

From sandbox to defensible science

  1. Phase 0 — Synthetic validation

    Exercise the architecture, provenance, APIs and visualization without pretending the values are biological.

  2. Phase 1 — Public datasets

    Train and evaluate behavioral models on published Drosophila tracking datasets with explicit provenance.

  3. Phase 2 — External lab collaboration

    Prospectively test early-detection hypotheses in an established neurodegeneration model under appropriate institutional oversight.

  4. Phase 3 — Closed-loop assistance

    Evaluate whether adaptive sensory feedback changes task performance, clearly separating compensation from treatment.

Build something great.

Start deploying Contactar a ventas