Scientists Are Simulating a Fruit Fly’s Brain. What Does That Actually Mean?

 AI-generated conceptual illustration of a fruit fly’s brain and neural network. It is an artistic representation, not an actual brain scan.

A complete wiring diagram, a few viral demos and a lot of big words. Here is what has actually been done, and what has not.

When I first read that scientists were simulating a fruit fly's brain on a computer, I felt a small jolt. Then a more careful question arrived: simulating what, exactly? A picture of the brain? A working copy? Something that, in its tiny way, thinks?

The honest answer sits somewhere in the middle, and the middle turns out to be more interesting than either “nothing happened” or “we copied a mind.”

First, what is a connectome?

A connectome is a wiring diagram: a list of nerve cells in a nervous system and the connections between them. Think of a road map of a city. It shows streets and junctions, but not the traffic moving through them.

In 2024, the FlyWire project published a detailed wiring diagram of an adult fruit fly brain, covering more than 125,000 neurons and about 50 million connections. Researchers produced it by slicing the brain into extremely thin layers, photographing them with an electron microscope and tracing the cells with the help of AI and extensive human checking.

This year, a second, larger map arrived. A team led by researchers at the Howard Hughes Medical Institute's Janelia Research Campus, working with Google Research and others, released a complete central nervous system connectome for a male fruit fly. It contains more than 166,000 neurons and about 125 million connections, and was published in Cell on September 3, 2026.

Unlike the earlier brain map, this one also includes the ventral nerve cord, which carries signals involved in controlling the fly's legs and wings. That matters because researchers can use the additional wiring information when investigating how neural activity might control movement.

But a bigger map is still a map.

What a map leaves out

A road map does not tell you how fast the cars drive, which lights are red or how many people are in each building. A connectome has similar gaps. It records that one cell connects to another, but does not by itself fully describe how strongly or quickly signals travel, how each cell behaves, or how connections and activity change with experience.

There is an old test case. The tiny worm C. elegans has 302 neurons, and its wiring diagram was published decades ago. That map has been useful for designing experiments, but a wiring diagram alone does not explain every aspect of how an organism behaves.

If a map of 302 neurons does not settle every question, a map of more than 139,000 neurons does not automatically settle them either.

From a map to a model

Mapping is not the same as simulating. To simulate a nervous system, researchers need rules for how its neurons behave and how signals pass through the network. 

In 2024, a team led by Philip Shiu and Kristin Scott, with collaborators, published a computational model based on the fruit fly connectome. They used simplified, switch-like neuron models and network connections derived from the wiring diagram.

The model produced useful predictions. When researchers activated sugar-sensing neurons in the simulation, it predicted activity in downstream neurons involved in the fly extending its straw-like mouthpart, called the proboscis, to feed. Experiments on real flies confirmed several of the model's predictions. The researchers also investigated circuits involved in grooming the antennae.

That is a meaningful scientific result: a model based on a wiring diagram helped generate predictions that could be tested against biological experiments.

But it was a limited result. The model represented the wider network, while the demonstrated behaviours involved specific circuits. As neuroscientist Patrick Mineault noted in his review of later demonstrations, only a few hundred neurons were activated in the feeding example he discussed. This was not a complete, adaptive simulation of everything a living fly can do.

Giving the model a body

In March 2026, a start-up called Eon Systems demonstrated a virtual fly by connecting a brain model to a simulated body. The demonstration showed the virtual animal performing actions such as walking, grooming and feeding. The company described its work in terms of uploading a fly, a framing that drew criticism from researchers.

The technical idea was more limited than the phrase “uploaded fly” might suggest. The demonstration combined existing components, including the FlyWire connectome, the 2024 computational model and a physics-based virtual fly body related to the NeuroMechFly project.

In his review, Mineault argued that selected neural outputs were mapped to pre-built body controllers, allowing the virtual body to perform particular actions. He also questioned how much the demonstration's visual system contributed to its behaviour.

These criticisms do not mean the work was worthless. Connecting a neural model to a body model is a useful experiment. But a system that triggers a pre-built movement is not necessarily demonstrating that a complete simulated brain independently sensed its environment, processed the information and generated the appropriate action.

Both things can be true: the demonstration can be technically interesting without proving that a fly has been uploaded into a computer.

The September wave

When the male fruit fly connectome became available, developers began connecting fly-brain models to games and other virtual environments. Online demonstrations included projects involving Beat Saber, Doom, Minecraft and Super Mario 64, alongside experiments with robot control.

These projects were not all doing the same thing.

Neuroscientist Patrick Mineault examined the code and descriptions of several viral demonstrations in his September 14 review. He argued that many did not demonstrate a complete, meaningful loop from sensory input to neural processing to motor output. In some cases, neural activity or random signals were mapped to game controls; in others, additional training or flexible control systems played a major role.

For example, a system that maps selected neuron outputs to joystick buttons may appear to make a virtual fly play a game. But if the game environment does not meaningfully influence the model's activity, the demonstration does not establish that the simulated brain is seeing the screen and choosing actions as a living fly would.

That distinction matters. A demonstration can be entertaining and useful for public engagement without proving that the brain model is doing the work people assume it is doing.

The criticism should also be applied carefully. Mineault's assessment concerns the particular demonstrations he examined, not every possible experiment involving a fly connectome. Some developers are actively working on more realistic sensory-driven models, and these experiments may help advance the field.

Behaviour is not understanding

So what should we make of all this? Copying the connections is not the same as copying every function of a brain. It is reproducing one important layer of its structure.

And a model that produces fly-like movement does not prove that the real fly produces that movement through the same process. Different systems can sometimes generate similar behaviour.

A separate 2025 project led by researchers associated with DeepMind and Janelia developed a whole-body physics simulation of fruit-fly locomotion using reinforcement learning and real motion data. That work modelled walking and flying without using a complete connectome as the driver of the simulated body.

It illustrates an important point: realistic-looking movement can be produced through different modelling approaches. Appearance alone does not tell us which biological mechanisms have been reproduced.

What would stronger evidence of a working whole-brain simulation look like? Mineault proposed several tests: complex, natural behaviour guided by what the model senses; neural activity resembling recordings from real flies; and the ability to reproduce behaviours specific to an individual fly.

These are demanding goals. The demonstrations discussed here do not establish that all of them have been achieved. Nor does any of this work establish that a simulated fly is conscious. Reproducing selected neural activity or behaviour is not, by itself, evidence of subjective experience.

So why should anyone care?

Because it is a test of how much we understand. If we say we know how a fruit fly's brain works, we should eventually be able to build a model that reproduces increasingly complex aspects of its behaviour. The gap between what we can model and what a real fly does can help reveal what we still do not understand.

It also gives biologists a new tool. The 2024 model produced predictions that researchers tested on real flies, showing how computational models can guide experiments rather than merely imitate behaviour. If these simulations improve, they may offer ideas for neuroscience and robotics. That remains a possibility to investigate, not a result already guaranteed.

Finally, it is a useful lesson in reading announcements. The same wiring map can sit behind a careful scientific paper and a viral claim of an upload, and from the outside they can look alike. The difference is in the details: what was modelled, what was trained, what was borrowed from existing systems, and what was actually measured.

What stays unanswered

I started by asking whether a computer can copy a brain. I now think the better question is a quieter one: how much of how a fly's brain works can we describe well enough to simulate?

Today, the honest answer is more than we could a decade ago, and less than the videos suggest.

This article is general information and analysis, not scientific or professional advice. Descriptions of disputed demonstrations reflect the sources listed below; interpretations of individual projects may evolve as research continues.

Sources and Further Reading

1. Shiu et al. — “A Drosophila computational brain model reveals sensorimotor processing” (Nature, 2024).
   https://www.nature.com/articles/s41586-024-07763-9

2. Google Research — “A connectomics milestone: Mapping the complete male fruit fly brain” (September 3, 2026).
   https://research.google/blog/a-connectomics-milestone-mapping-the-complete-male-fruit-fly-brain/

3. HHMI Janelia / collaborators — “Sexual dimorphism in the complete Drosophila male central nervous system connectome” (Cell, 2026).
   https://www.cell.com/cell/fulltext/S0092-8674(26)00942-6

4. Patrick Mineault — “Deconstructing viral fly sims” (September 14, 2026).
   https://www.neuroai.science/p/are-flies-playing-beat-saber

5. Eon Systems — “How the Eon Team Produced a Virtual Embodied Fly” (March 10, 2026).
   https://eon.systems/updates/embodied-brain-emulation

6. Eon Systems — “We've Uploaded a Fruit Fly” (March 8, 2026).
   https://eon.systems/updates/weve-uploaded-a-fruit-fly

7. Vaxenburg et al. — “Whole-body physics simulation of fruit fly locomotion” (Nature, 2025).
   https://www.nature.com/articles/s41586-025-09029-4

8. Dorkenwald et al. — “Neuronal wiring diagram of an adult brain” (Nature, 2024).
   https://www.nature.com/articles/s41586-024-07558-y