Researchers have unlocked new possibilities for artificial intelligence by digitizing the brain of a fruit fly, allowing it to trade cryptocurrency, play video games, and even brew beer. This advancement comes from making the connectome of the male Drosophila's central nervous system publicly available.
The most intriguing aspect of this project is that the simulated fly can behave like a real one without any prior training. While some of the experiments may seem whimsical, they hold significant implications for medical and technological advancements.
Let’s explore the purpose of this digital insect brain, its practical applications, and its relevance to AI learning.
Two Decades of Research on the Fly Brain
In June, the FlyEM international consortium released the connectome of the male fruit fly. This project involved researchers from Janelia Research Campus, the University of Cambridge, MRC Laboratory of Molecular Biology, and Google Research.
Source: FlyEM.The research spanned 20 years, during which the insect's nervous system was sliced into thousands of thin layers, each photographed using an electron microscope. The resulting images were then used to reconstruct the three-dimensional structure of the brain.
Due to the automatic neuron tracing's inherent inaccuracies, manual verification was essential. Researchers meticulously tracked the paths of individual neuron branches.
The final product was a comprehensive map of the male fruit fly’s nervous system, detailing 166,700 neurons, 124.2 million synapses, and 11,710 cell types. In comparison, the roundworm Caenorhabditis elegans, which was first described in the 1980s, consists of just 302 neurons.
Connectome of the male Caenorhabditis elegans. Source: N+1.This effort resulted in the first complete diagram of the male fruit fly’s central nervous system, consolidating the brain, visual lobes, and ventral nerve cord (analogous to the spinal cord in insects) into a single map. Previously, these components had been mapped separately, often with disconnected pathways. Now researchers can trace the entire journey of nerve signals, from sensory input to muscle commands.
A separate study focused on gender comparisons, revealing that most neurons in males and females are of the same type. Differences were primarily found in brain areas that integrate incoming information and regulate behavior, while the regions responsible for processing external signals and controlling movements were nearly identical.
In essence, identical input signals can be processed differently based on sex.
All data is available under a CC-BY license, allowing anyone to download the connectome and utilize it.
From an academic standpoint, this tool enables researchers to test hypotheses about neuron connections and their effects on behavior, paving the way for new experiments and discoveries.
“This fundamental map of the adult male fruit fly brain can accelerate our understanding of the brain and is a significant milestone in neurobiology,” stated researchers from Google Research.
A standout feature of the connectome is its capability for simulation, which has opened avenues for experiments beyond traditional lab settings.
Real-World Experiments
Enthusiasts quickly began placing the digital fly in various scenarios. Some simply had it play video games, while others succeeded in recreating genuinely significant scientific experiments.
From DOOM to Minecraft
One of the first projects involved programmers sending the fly into the classic 1993 shooter DOOM.
In September, Coinbase engineer Alex Wormuth launched the Doomfly project, connecting the fly's brain simulation to the game. Each game frame translates into sensory signals, with pixel brightness and color fed to the visual neurons, which then determine the fly's movement and shooting actions. Damage indication is managed by two dopamine neurons, PPL101, linked to hunger and aggression.
Source: X/Alex Wormuth.While the live stream of the fly’s gameplay is now offline, it previously showcased the fly's actions alongside neuron response logs for every move.
The virtual fly proved to be a poor player, but the underlying mechanism works—the connectome's neural activity controls the character in real time.
Researcher Lyra Bubbles tested the model in the rhythm game Beat Saber, reporting that the fly performed fairly well, although it required some initial training for the visual cortex's responsiveness.
the fly brain can play beat saber pic.twitter.com/AqKMqD6gzP
— lyra bubbles (@_lyraaaa_) September 9, 2026
Developer Jessica Paquet integrated the connectome into Super Mario 64, while YouTuber Ro0oney did so in Minecraft, using a model based on the female fruit fly from the FlyWire project.
Source: X/Jessica Paquet.All these experiments share a common thread: the fly's behavior was not programmed. Developers merely connected neural signals to game parameters, leaving the biological framework to make decisions.
Trading and Marketing
Another line of experiments focused on financial applications. Wormuth created the Stonkfly simulation, which analyzes Bitcoin charts as images and makes trading decisions—buy, sell, or hold.
The reinforcement logic mimics the actual fruit fly's brain: profits activate dopamine neurons, while losses trigger aversive cells.
Simulation of trading using the fly connectome. Source: Stonkfly.On its first day with a virtual account of $100, the system earned $1. The project’s author emphasized that the simulation is not evidence of trading capabilities but rather a test of the relationship between sensory input, neural activity, and responsive actions.
Meanwhile, Czech brewery Vaclav utilized the fly connectome to market its products. The company created three simulations where the fly brews different types of beer.
The fly brews Czech lager. Source: Vaclav.“This is a simplified brewery model where we brew our beer. […] The fruit fly faces our everyday challenges—choosing the direction of the wort flow and when to turn off the heat,” the site states.
Virtual Body
One of the most notable applications emerged in spring when Eon Systems connected the fruit fly connectome to a virtual body using the NeuroMechFly v2 simulator.
The virtual insect was equipped with compound eyes, antennae, and a physically modeled body. The result: without any training, it began walking, grooming its antennae, and responding to stimuli just like a real fruit fly.
Source: NeuroMechFly.The behavioral accuracy matched that of a living fly at 91–95%. When researchers intentionally “messed up” connections in the scheme, accuracy plummeted to 1%.
This demonstrated the hypothesis that behavior is determined by the connectome's architecture rather than random neural connections.
Major Limitations
All the mentioned simulations are far from practical applications. None fully reproduce the fly's brain: neurotransmitters, chemical modulation, and gene expression have not been taken into account.
Games and cryptocurrency trading primarily showcase the conceptual viability of the approach rather than its maturity. However, they demonstrate that an accurately replicated biological scheme can function.
Most importantly, it operates from the outset without any training, solely relying on the established neural connections shaped by evolution.
This begs the question of future potential. If the fruit fly's brain can be connected to a virtual body to produce meaningful behavior, what could be achieved with a mouse brain, and eventually, a human brain?
Future Prospects: Where the Digital Fly is Headed
Experiments with Drosophila are merely a starting point; the next frontier involves more complex neural connections. Such a breakthrough could spur entire scientific and technological fields, although challenges remain.
Digital Mouse
Researchers have chosen the mouse brain as the next target for digitization. A mouse has around 70 million neurons, which is approximately 560 times more than a fruit fly. This volume significantly complicates the mapping process, but work is already underway.
To tackle this, the Connectome-seq technology has been introduced, allowing researchers to quickly and accurately label neurons with molecular “barcodes” and create maps of their connections. This project paves the way for large-scale mapping of mammalian brains.
Additionally, a study published in Nature Neuroscience reported results from a large-scale reconstruction of the CA3 hippocampal region in mice using volumetric electron microscopy.
Digitized CA3 hippocampal region in mice. Source: Nature.Furthermore, Eon Systems, which created the simulation of the “real” fly, announced plans to construct a mouse connectome, followed by a human one.
The mouse brain is far more complex than that of an insect, making it a valuable model for studying memory, learning, and neurodegenerative diseases.
If successfully digitized and activated, it could lead to a new class of experiments—from drug testing to the development of brain-computer interfaces.
AI Learning
Traditional neural networks require vast datasets for training; to teach a model to recognize images or play games, millions of parameters must be adjusted.
The approach demonstrated by the fruit fly connectome is fundamentally different. It does not necessitate zero-based training, as the architecture shaped by evolution is already in place.
This theory has been validated. Researchers from the Advanced Concepts Team utilized the topology and weight distribution of the fruit fly connectome as a reservoir in Echo State Networks.
Scatter plot of data for training within the fruit fly connectome. Source: MDPI.The result was that the architecture based on a “living” CNS exhibited greater resistance to overfitting compared to standard implementations, achieving a normalized error of less than 2% when employing the complete connectome.
Another study on the same subject developed a generative approach to echo states based on the connectome for predicting time series. This method is seen as a way to create more efficient recurrent neural networks by incorporating principles from real brains.
Moreover, there are projects transferring the fruit fly connectome to neuromorphic chips and GPUs. The open repository Eon Systems on GitHub includes implementations of the model on Brian2, Brian2CUDA, PyTorch, and NEST GPU.
This represents a step towards creating energy-efficient computing systems that operate based on structural connections rather than data processing.
Medicine and Digital Twins
In February, researchers from Japan's National Center for Neurology and Psychiatry and Tohoku University published a study on the “digital twin of the brain.”
This system translates an individual’s connectome into predictions of multitasking behavior. In a sample of 228 participants, the accuracy of predicting behavioral signals exceeded 90%.
Model of the “digital twin.” Source: EurekAlert.Using gradient feedback, researchers conducted virtual interventions, selectively modeling target functions—such as the strength of the amygdala's reaction or the speed of information processing.
“This research represents a significant step towards mechanistic, individualized psychiatry, offering a powerful platform for understanding pathophysiology and developing personalized treatment methods,” the study states.
Additionally, there are already projects modeling schizophrenia using personalized virtual “brain twins” to test treatment hypotheses.
What’s Missing
However, all these prospects face a fundamental barrier. A review in NeuroAI for AI Safety states that the connectome describes how neurons are connected but does not explain how they function. There is a lack of data on neurotransmitters, chemical modulation, and synaptic dynamics.
Authors of the State of Brain Emulation Report 2025 highlight three key capabilities necessary for accurate brain emulation: recording neural activity, reconstructing its structure, and modeling.
Progress is being made in all these areas. Recording technologies now allow simultaneous monitoring of millions of neurons, and connectomics has advanced from mapping a small CNS of a worm to fully recreating the brain of an adult fruit fly.
Moreover, the cost of reconstructing a single neuron has decreased from approximately $16,500 (for the first connectome of a worm) to around $100 in recent projects involving fish larvae.
However, for the human brain, the challenge lies not only in data availability but also in computational power. Such emulation may require between 1018 and 1025 FLOPS, depending on the level of detail, while the most powerful current GPU, the NVIDIA H100, delivers about 3.9 × 1015 FLOPS.
Comparison of neuron counts across species. Source: Allen Institute.Nonetheless, the authors of the report remain optimistic, believing that recreating the functions of a small animal's brain may be achievable in the coming years, followed by efforts to emulate human brains.
***
The direction for future research has been set. However, how long it will take scientists to create our connectome remains a mystery. For now, we can only observe the extraordinary situations in which the fruit fly may be placed.
Follow ForkLog on social media
Telegram (main channel) Facebook X