First Complete Male Fruit Fly Connectome Mapped With AI

Google Research, HHMI Janelia, MRC LMB and Cambridge published the first full wiring diagram of a male fruit fly nervous system: 166,691 neurons, 11,691 types.

by HowAIWorks Team
On this page

Introduction

On September 3, 2026, a team led by the MRC Laboratory of Molecular Biology with HHMI Janelia, the University of Cambridge and Google Research published the first complete connectome of an adult male fruit fly — a synapse-resolution map of every neuron in its brain and nerve cord, and the connections between them. The paper, Berg et al., ran in Cell 189(18), alongside three companion studies that put the map to work on taste, vision and social behavior.

The dataset contains 166,691 neurons sorted into 11,691 cell types, which makes it the largest finished whole-nervous-system wiring diagram by neuron count. Google Research puts the number of synaptic connections at about 125 million. It is free to use and browsable in Neuroglancer.

What Is Actually New

Previous fly connectomes — the Janelia hemibrain, FlyWire's whole female brain — were female, and stopped at the neck. This one spans the central brain, both optic lobes and the ventral nerve cord (the fly's spinal-cord analog) as a single continuous volume, joined through an intact neck.

That continuity is the point. It lets a researcher trace an unbroken path from a photoreceptor in the eye or a taste sensillum on the leg, through the brain, to the motor neurons that move a wing or a leg — sensation to behavior, in one dataset. Courtship, which involves vision, chemosensation, song production and locomotion at once, becomes traceable end to end.

How AI Built It

The physical process is brute force: fix a fly, cut it into slices far too thin for a light microscope to resolve, and image every slice with focused ion beam scanning electron microscopy. That yields millions of 2D images and no neurons — reconstructing a neuron means deciding, across thousands of consecutive images, which pixels belong to the same cell.

Google Research does that with flood-filling networks: convolutional neural networks that start from a single seed pixel and iteratively grow a segmentation outward until the object is complete. It is a computer vision problem where a single mistaken merge can fuse two unrelated cells and corrupt every downstream circuit claim. Google says recent accuracy and speed gains came partly from adding synthetic neurons to the training data.

Automation does not finish the job. Human experts proofread the machine output and annotated cell types, including expression of the sex-determination genes fruitless and doublesex. The MRC LMB puts that effort at the equivalent of 44 years of human labor. The connectome is a machine-then-human artifact, and the proofreading is the expensive half.

The Sex Differences Sit in the Wiring

With a female connectome already available, the two could be compared type by type for the first time in an animal with complex social behavior. The counts: 7,205 isomorphic cell types, 262 male-specific, 69 female-specific and 114 dimorphic — present in both sexes but connected to different partners.

That last category is the finding worth pausing on. A dimorphic neuron is not a different cell in males and females; it is the same cell type wired into a different neighborhood, and therefore participating in different behavior. The AOTU012 neurons are the paper's worked example.

The numbers show why a small anatomical difference matters. Sex-specific and dimorphic neurons make up 4.8% of the male central brain and 2.4% of the female one, and they cluster in higher-order centers rather than at the sensory or motor periphery. But they are unusually well connected: by the MRC LMB's account, around 12% of connections in the male brain differ between the sexes, against roughly 4% in the female. Dimorphism propagates outward through connectivity, so a few percent of cells reshapes a much larger share of the circuit.

The Three Companion Studies

Published the same day, each uses the connectome rather than merely citing it:

  • Taste — a complete gustatory connectome, in Cell. Gustation had been one of the least mapped senses at connectome resolution. The study maps every feeding motor neuron and traces full sensory-to-motor pathways, finding separate circuit architectures for nutrient assessment, motor control, neuroendocrine regulation and courtship.
  • Vision — the organization of visual pathways across the brain, also in Cell. Projection neurons linking the optic lobes to the central brain sort into feedforward, feedback and bilateral groups: the feedforward set forms an information bottleneck that consolidates visual features before distributing them through modular pathways, while feedback neurons return contextual signals from the central brain through recurrent loops.
  • Social behavior — in Current Biology, describing an expansive and heavily interconnected male-specific network positioned to control male social behavior, including courtship and aggression.

What It Does Not Give You

A connectome is anatomy. It records that neuron A synapses onto neuron B, not how strongly, not whether the synapse is excitatory or inhibitory in a given state, and not what neuromodulators are doing to the circuit at the moment the fly decides to sing. Simulating fly behavior from the map remains an open research problem — the wiring diagram constrains the hypotheses rather than replacing the experiments.

The scaling gap is also worth stating plainly. This fly has 166,691 neurons; standard estimates put a mouse brain at roughly 70 million and a human brain at around 86 billion. Connectomics teams are now working toward zebrafish and mouse, but nothing in this result implies a human connectome is near.

Conclusion

The interesting part of this release is not that AI mapped a brain. It is the division of labor: a deep learning pipeline turned an intractable segmentation problem into a tractable one, and human experts then spent the equivalent of 44 years verifying it. That is what AI in science mostly looks like right now — automation that moves the bottleneck rather than removing it.

The scientific return arrives through the companion papers, and through whoever uses the map next. It is public, and the female connectome it can be compared against is public too.

Sources

Frequently Asked Questions

A connectome is a complete wiring diagram of a nervous system: every neuron and every synaptic connection between them, reconstructed at the resolution of individual synapses. It describes the circuit, not the activity running through it.
166,691 neurons sorted into 11,691 cell types, spanning the central brain, both optic lobes and the ventral nerve cord. Google Research puts the number of synaptic connections at about 125 million. It is the largest completed whole-nervous-system connectome by neuron count to date.
Millions of 2D electron-microscope images of thin tissue slices were segmented into 3D neuron shapes by flood-filling networks — convolutional neural networks that start from one pixel and grow the object outward. Human experts then proofread the result, work the MRC Laboratory of Molecular Biology puts at the equivalent of 44 years of human labor.
Of the cell types compared, 7,205 are isomorphic (the same in both sexes), 262 male-specific, 69 female-specific and 114 dimorphic — present in both sexes but wired to different partners. Those neurons make up 4.8% of the male central brain and 2.4% of the female one, yet account for a much larger share of the connections that differ between the sexes.
Taste, vision and social behavior. Two appeared in Cell alongside the connectome — a complete gustatory connectome tracing taste from sensory neurons to feeding, foraging and courtship circuits, and a map of how visual pathways are organized across the brain — and one in Current Biology, describing the male-specific network that governs social behavior.
Yes. The dataset is free for researchers worldwide and browsable in Neuroglancer, Google's open-source viewer for large volumetric data, via the Janelia FlyEM male CNS site.
No. A connectome is anatomy, not physiology. It gives the circuit diagram but not synaptic strengths, neuromodulation or dynamics, so predicting behavior from it still requires separate experiments and models.

Continue Your AI Journey

Explore our glossary and model catalog to deepen your understanding.