How a fruit fly brain simulation works: the FlyWire connectome, neuron models and bodies

The sense → model → act loop

• MeasuredTaken from the real animal. In most projects this is only the wiring: which neurons connect, and through how many synapses.
• ModelledA mathematical rule chosen by scientists, such as how a neuron adds up its inputs and when it fires. It is a simplification of real biology.
• Hand-madeChosen by the project author: which neurons receive the input, which neurons are read out, and what counts as an action.
• TrainedLearned by a machine learning method, for example a readout layer or a controller trained to reach a goal.
• ScriptedOrdinary program code or animation that does not depend on the simulated neurons.

A project can use a real connectome and still get most of its visible behaviour from hand-made or scripted parts. That is why we grade each project separately. See Is it real? Whether the measured wiring itself makes a difference is a separate question, tested by scrambling it: Does fly wiring help? compares 31 such studies.

What a connectome is

A connectome is a map of the connections in a nervous system: which neurons exist, which neurons connect to which, and how many synapses join them. Fly connectomes are traced from electron microscope images of one animal, cut into thousands of very thin slices. Computers and human proofreaders follow each neuron through the images.

A connectome is a wiring diagram, not a working brain. It does not say directly how strong each connection is, how neurons change with learning, or how chemicals such as hormones change activity. A simulation has to add those parts, usually with simple rules.

Map or brain? Run with such a model, the map still does something specific to its wiring. In our own looming test (5 Oct 2026) a shadow on one eye made the real map's turning neurons fire on the far side and turned a simulated fly away in 4 of 4 runs; scrambled maps with the same number of connections per neuron stayed silent (0 of 3). One model, open loop, our hand-made mapping, a constant external walking drive, 1 s; not a real fly.

What the whole-brain models add, and what they leave out

Most projects reuse the leaky integrate-and-fire (LIF) model of Shiu et al. (Nature, 2024). In that model, and in most copies of it:

• Connection strength comes from synapse counts. Each connection's weight is its FlyWire synapse count multiplied by one free number. The authors chose that number by hand so that sugar-sensing neurons firing at 100 Hz drive the proboscis motor neuron MN9 to about 80% of its maximum.
• Excitation or inhibition comes from predicted transmitters. The sign of each connection comes from a machine-learning prediction of the neurotransmitter, not from a measurement of each synapse.
• Many parts of biology are missing. In the authors' words, the model "does not account for gap junctions, non-spiking neurons, internal state or long-range neuropeptides, and assumes that the basal firing of each neuron is zero". Neuromodulation and learning (plasticity) are not in the base model.
• It is slow on a normal computer. The paper reports about 5 minutes per simulated second per CPU thread for the Brian2 version.

Source: Shiu et al. 2024, Methods and Discussion (full text on PubMed Central).

Datasets compared: FlyWire connectome, MaleCNS, BANC and others

Different teams have mapped different flies and different parts of the nervous system. Always read a neuron count together with its dataset, version and scope. Never add a brain count to a nerve-cord count from a different animal.

Codex also lists MAOL v1.1 (male right optic lobe, 52,445 neurons). It comes from the same male fly as MaleCNS, so its neurons are part of MaleCNS, not extra.

Sources: Codex, FlyWire guidelines, male-cns.janelia.org, BANC project, Virtual Fly Brain neuron counts, Winding et al. 2023, Google Research blog.

Why the neuron counts differ

News stories quote many different neuron counts. Each one belongs to a specific dataset, version and scope:

What you have to download, and how big it is

On 6 Oct 2026 we read the size, checksum, licence and access rule of every file in 8 releases (403 files) from each host's own storage metadata, without downloading the data. To run the Shiu et al. model you need two files, 104.1 MB; a full release is gigabytes to hundreds of gigabytes. How to get the two files: Build your own.

Neuron models and engines

The neuron model decides how activity moves through the wiring. Most projects use the Shiu et al. LIF model or a copy of it on faster hardware. Numbers below come from the code and from the authors; we did not re-run them.

Bodies: NeuroMechFly (FlyGym) and flybody

A brain model alone does not move. To walk, groom or fly, a project needs a simulated body with legs, joints and physics, and a way to connect neurons to muscles.

Every embodied connectome demo we inspected needs extra machinery between brain and body: pre-trained or pre-programmed walking controllers, fitted rhythm generators, or trained gains. None of the brain datasets alone makes a fly walk.

Measured: who does the walking when a brain drives a body

On 2 Oct 2026 we measured this split ourselves. We connected the Shiu et al. whole-brain model to FlyGym's walking fly, stimulated the brain's forward-walking neurons (P9) and turned its descending neurons' rates into FlyGym's two-number walking drive with a rule we wrote. Each stage of the loop above did a different part of the work:

The result: the fly walked 13.9 mm in 1 s with the real brain, 14.1 mm with a scrambled brain and 14.3 mm with no brain. The legs, rhythm and balance were FlyGym's; the brain added turning commands. Chart, controls and limits: Who does the walking?; build it: Add a body.

Measured: who does the turning when a looming shadow appears

On 5 Oct 2026 we added a sense. We showed a looming shadow to one eye of the same brain model (every LC4 and LPLC2 neuron of that eye, the fly's looming detectors), held the walking drive constant ourselves, and let the brain's turning neurons steer. This time the brain's wiring decided which side's turning neurons fired; the walking was still FlyGym's, and the rule from rates to legs, which turns that side into a direction, was still ours:

The result: the real map turned the fly away in 4 of 4 runs, scrambled maps in 0 of 3 (their drives were the same as no brain). Small n, much quieter scrambled brains, and a direction that rests on our mapping (our caption, fixed before the test: "fly67's claim; the direction in real flies was not verified by us.") Chart and caveats: We ran the map; build it: Add a sense: looming.

Where in the map the turn runs

On 6 Oct 2026 we looked inside the brain model (no body). The looming detectors have no direct synapses onto the turning neurons DNa02: their signal reaches the far-side DNa02 through relay neurons, the strongest being PVLP141 on the eye's side. They do synapse directly onto the giant fibre, the escape neuron, of their own side. We then scrambled only the inside of the map, keeping every connection onto the output neurons. In the words we fixed in advance: "Scrambling only the inside of the map kept about -9% of the turn command and 69% of the escape signal." (−9% means none: the far-side command was absent in all 6 trials.) Silencing the giant fibre left the turn command alone, as the map predicts: it has no giant fibre → DNa02 connections. Scrambled maps turned up to the real map's activity still gave no turn command. These are counts and spike rates in one model, not a moving fly. Chart and caveats: Where in the wiring is the turn?