We Haven’t Uploaded a Fly Brain — But We Can Finally Test How to Simulate One

In September 2026, researchers published a connectome of the adult male fruit fly containing 166,691 neurons across the brain and ventral nerve cord, linked by roughly 125 million synaptic connections. Google Research describes it as the largest brain map yet produced by number of neurons. Within days, the internet was talking about downloadable fly brains, virtual flies and the beginnings of biological mind uploading.

We have not uploaded a fly. What has happened may be more important scientifically: neuroscience is acquiring nervous systems detailed enough that researchers can begin testing, rather than merely speculating about, what a brain simulation actually requires. If we know the wiring in extraordinary detail, how much of the remaining biology must we reproduce before a computational model begins to predict what the real nervous system does?

Editorial image generated by the author.

A Wiring Diagram Is Not a Brain Upload

A connectome is essentially a map of neural wiring. It identifies neurons and records which neurons form synaptic connections with which others. The new male fly dataset also contains extensive information about neuron shape and type, and its comparison with existing female connectomes reveals where circuitry is shared between the sexes and where it differs. Because the new reconstruction includes the ventral nerve cord as well as the brain, researchers can trace pathways much further from sensory input towards the neural machinery controlling the body.

That is an extraordinary amount of information, but it is not a recording of a living fly in operation. A structural reconstruction does not directly contain the membrane voltage of every neuron at some moment in the animal's life, the concentrations of every neuromodulator, the state of every ion channel or the many temporary biochemical changes that alter how neurons and synapses behave. Nor does the connectome alone tell us how much of an animal's prior learning or internal state can be recovered from structure.

The analogy to an electronic circuit diagram is useful up to a point. We increasingly know which biological components connect to which others, but that does not automatically tell us their complete operating state. Biological neurons are also far less uniform than transistors: their responses depend on morphology, receptors, chemistry, recent activity and the surrounding physiological environment.

Until recently, neuroscience lacked even this kind of wiring diagram for an animal as complex as a fruit fly. Researchers can now follow potential paths between sensory neurons and motor systems, compare individual and sex-related differences at synaptic resolution and ask how changes in wiring alter the computation performed by a circuit. The more consequential experiment begins when that wiring diagram is made to run.

What a Crude Model Can Already Predict

In 2024, Philip Shiu and colleagues built a computational model of the adult fruit fly brain using an earlier FlyWire connectome containing more than 125,000 neurons and around 50 million synaptic connections. They did not attempt a molecularly faithful reconstruction of each cell. Instead, they represented the neurons using comparatively simple leaky integrate-and-fire models.

Inputs raise a simulated neuron's membrane potential, which decays over time. If it crosses a threshold, the neuron fires and influences neurons connected downstream. Connectivity and synapse counts supplied much of the network structure, while predicted neurotransmitter identity helped determine whether connections were excitatory or inhibitory.

As biology, this is drastically simplified. Real neurons contain different ion channels and receptors. Their branching dendrites influence how signals are integrated. Synapses change with experience and recent activity. Neuromodulators alter the operation of entire networks, while hunger, arousal, hormonal state and other physiological conditions can change how the same sensory stimulus is processed.

A reasonable expectation might therefore have been that crude mathematical neurons inserted into a connectome would produce complicated but biologically uninformative activity. Instead, the model recovered meaningful features of real sensorimotor circuits. When researchers stimulated simulated sugar- and water-sensing neurons, it predicted many of the neurons involved in feeding responses in biological flies. It also generated predictions about neurons whose manipulation should influence feeding, which could then be tested experimentally. Similar modelling recovered important features of the circuitry involved in antennal grooming.

Across 164 predictions for which experimental comparisons were available, 91 percent were consistent with the empirical results. Under a more restrictive calculation excluding one large optogenetic experiment, the figure was 84 percent.

Those numbers do not mean that the simulated fly brain is “91 percent accurate”, still less that scientists reproduced 91 percent of a fly's behaviour. The tests concerned particular predictions within particular feeding and grooming circuits, and the researchers themselves emphasised that connectivity constrains possible circuit mechanisms without uniquely determining them.

Even within those limits, a model built largely from connectivity, approximate connection strengths, neurotransmitter identity and simple neuron dynamics was able to make experimentally useful predictions about a biological nervous system. Something about what those circuits do is already encoded in their architecture. The open question is how far that remains true as the demands placed on the simulation increase.

The Fidelity Frontier

One intuition treats the nervous system somewhat like an engineered machine: identify the components, determine how they interact, supply the appropriate inputs and the resulting behaviour should follow. Another places much greater weight on the continuously changing biological state of the system. From that perspective, knowing the wiring may be insufficient unless we also know synaptic strengths, receptor distributions, neuromodulation, physiological state and recent history.

The advantage of the fly is that researchers can increasingly stop treating this as an abstract dispute. Start with connectivity and simple neurons, compare the simulation against recordings and interventions in real animals, then add more accurate neurotransmitter information, receptor dynamics, neuronal morphology, synaptic plasticity or internal-state signals. The relevant question is not simply whether each addition makes the simulation more realistic, but which predictions it improves.

There may be no single “minimum sufficient description” of a fly brain. A model capable of predicting whether a sugar stimulus initiates feeding may be inadequate for navigation. A model that predicts short-term sensorimotor responses may fail once learning is involved. Reproducing a behavioural choice may require less biological detail than reproducing the timing and activity of thousands of neurons leading to that choice.

What the connectome makes possible is therefore not a search for one magical resolution at which a brain suddenly becomes simulatable. It allows researchers to map a fidelity frontier: which biological details become necessary as we demand increasingly difficult predictions from the model.

This changes the nature of the problem. Researchers can deliberately construct incomplete brains, compare them with real ones and observe where the approximation stops working. If realistic behaviour eventually requires detailed models of every dendritic compartment, receptor system, neuromodulator and changing synaptic state, we will have learned that biological implementation resists aggressive compression. If comparatively simple models continue performing unexpectedly well as the tests become more demanding, the architecture may carry more of the relevant computation than expected.

The existing experiments establish neither conclusion. Their significance is that we finally have a practical way to distinguish between them.

A Virtual Body Does Not Make an Uploaded Fly

The picture becomes more seductive once the neural model is attached to something that looks and moves like an animal. In March 2026, Eon Systems combined a model based on the earlier female FlyWire brain connectome with NeuroMechFly, a detailed biomechanical simulation of a fruit fly running in the MuJoCo physics engine. The virtual animal could receive selected sensory inputs, process them through the connectome-constrained neural model and use neural outputs to influence its simulated body.

This created something the isolated brain model did not have: a closed sensory-motor loop. Movement changes what the fly perceives; the changed sensory input alters neural activity; neural output produces another movement. Eon's system could demonstrate behaviours including food-seeking and grooming.

Eon's own technical account is considerably more cautious than the phrase “uploaded fly” suggests. Only a subset of sensory inputs and behaviours has been implemented. The neurons remain simplified, important forms of plasticity and internal state are missing, and mappings between selected neural outputs and the virtual body have been engineered by hand. Walking relies substantially on locomotion controllers already built into the body simulation rather than emerging from a reconstructed chain running all the way through the fly's biological motor hierarchy.

The system is therefore an interesting integration of a connectome-derived brain model and a sophisticated virtual body, not evidence that a biological fly has been reproduced computationally. Embodiment also exposes another dependency: a nervous system does not normally operate as an isolated information processor plugged into generic inputs and outputs. It develops inside an organism and receives continuous information about movement, energy state, damage, temperature and other physiological conditions.

The same issue becomes much harder in proposals for human whole-brain emulation, where how much of the body a functioning mind actually requires may turn out to be part of the brain-simulation problem rather than an optional complication added afterwards.

A more convincing test of the fly model would consequently be less visually impressive than watching a digital insect walk. Present a biological fly with stimuli that were not used to calibrate the model and record as much neural activity as experimental techniques allow. Give the computational model equivalent inputs, then ask whether corresponding populations activate, whether the same neurons are inhibited, whether activity propagates with similar timing and whether the model predicts the downstream behavioural response.

The experiments can then become progressively less forgiving. Silence particular neurons in both systems. Combine sensory cues. Change hunger or arousal. Repeat stimuli and look for learning. Choose situations specifically because the model was not tuned to reproduce them.

A useful whole-brain simulation would earn credibility in the same way other scientific models do: by making predictions and surviving attempts to make them fail. There need not be a single moment when the answer to “have we simulated the fly?” changes cleanly from no to yes. A model might first predict simple sensorimotor transformations, then more complicated behaviour, internal neural dynamics, learning and responses to unfamiliar situations. Each step would expose where the remaining biological approximation still matters.

Why This Matters Beyond the Fly

The temptation is to draw a straight line from 166,691 fly neurons to the roughly 86 billion neurons in the human brain. That line would be misleading. The difference is not simply one of computing power. Human nervous systems contain enormously more neurons and connections, while development, plasticity, cellular diversity, interaction with the body and dependence on lifelong experience introduce problems that cannot be solved merely by scaling up a Drosophila model.

Nor does successful behavioural simulation automatically tell us anything about subjective experience. Predicting what a nervous system will do is not the same as demonstrating that a simulation feels anything. Current fly research offers little basis for resolving questions about consciousness, personal identity or whether a computational copy of a particular organism would constitute a continuation of that organism.

Those harder questions sit behind speculative ideas such as whole-brain emulation and copied human minds. The fruit fly does not show that such futures are achievable. It helps expose one of the assumptions they depend upon: that the information required to reproduce a nervous system can be identified, captured and represented computationally.

Before asking whether a human brain can be emulated, it would be useful to know what fidelity is required even for a fly. Do accurate simulations need precise dendritic geometry? Complete receptor maps? Glial interactions? Hormonal signals? Initial electrical states? Detailed histories of synaptic plasticity? Or can some of these processes be compressed into simpler variables without losing the behaviour or neural activity we want to reproduce?

Drosophila occupies an unusually useful point in that investigation. Its nervous system is complex enough to integrate several senses, navigate, choose between competing behaviours and learn from experience, yet small enough that scientists can now reconstruct essentially its entire central nervous system at synaptic resolution. The fly is therefore less a miniature human waiting to be uploaded than an experimental test bench for the idea of brain simulation.

Prediction Before Explanation

There is one further possibility. We tend to imagine that neuroscience must first understand a brain and only then simulate it, but those achievements need not arrive in that order.

A future fly model might correctly predict neural responses across thousands of interventions while remaining difficult to compress into a satisfying human explanation of why the entire network behaves as it does. Researchers could know what will happen when a neuron is removed, a connection weakened or a sensory signal changed without reducing the behaviour of more than 160,000 interacting cells to a handful of elegant principles.

That would still constitute a substantial kind of understanding: a causal model capable of surviving experimental tests. It simply would not be the compact, intuitive explanation we often hope science will eventually provide.

We have not digitised a fly mind. We have something scientifically more useful: a nervous system detailed enough that researchers can deliberately leave biology out and measure what stops working. Different behaviours and different kinds of prediction may demand different levels of fidelity, but the fly gives us a way to discover those requirements experimentally — to learn which details can be discarded, which cannot, and how the answer changes as we ask the simulation to do more.

Brain emulation remains a distant ambition. At least in the fruit fly, however, part of the problem has crossed an important threshold: from wondering whether a wiring diagram is enough to finding out exactly how far it gets us.