The male fruit fly connectome: mapping a complete nervous system with AI

What is the male fruit fly connectome?

It is a complete wiring diagram of one adult male fruit fly's central nervous system: every neuron in the brain and in the ventral nerve cord, the insect counterpart of the spinal cord, plus the synapses that join them. HHMI's Janelia Research Campus led the project, with Google Research, the University of Cambridge and the MRC Laboratory of Molecular Biology among its partners. The main paper appeared in Cell.

The Google Research post by Michał Januszewski and Viren Jain counts more than 166,000 neurons and 125 million synaptic connections, and calls it the largest brain map by number of neurons so far. The Cell paper gives 166,700 neurons and 11,710 neuron types, all proofread and annotated, including which neurons express fruitless and doublesex, two transcription factors that mark sexually dimorphic neurons in flies.

The map complements the female fruit fly brain map from the FlyWire project and a more recent female brain and nerve cord map. With both sexes mapped at synaptic resolution, researchers can compare circuits directly and start to measure how much wiring varies from one animal to another.

Why map a fruit fly brain?

Because a whole human brain is still out of reach and a fly's is not. The post notes that a human brain holds about 86 billion neurons. The male fly's entire central nervous system holds about 166,700 neurons, yet it still produces navigation, courtship and aggression. Drosophila melanogaster is also a long-established model organism in genetics, so connectome work builds on a deep base of existing knowledge.

Sex differences make the male map especially useful. Male and female flies behave differently in courtship and aggression, and the authors argue that dimorphisms act as a natural experiment: when the same stimulus produces different behaviour in each sex, the circuits that differ point to how the brain produces that behaviour.

How do you map every neuron in a brain?

By slicing, imaging and reconstructing. According to the preprint, the nervous system was cut into 20 μm slabs with a heated-blade method, then imaged by seven enhanced focused ion beam scanning electron microscopes over 13 months. The result was a 160-teravoxel image volume at 8 nm resolution in every direction, which software then aligned into one continuous block.

Google Research's contribution centred on automated neuron reconstruction. Flood-filling networks, convolutional neural networks that start from a single pixel and grow outwards to claim every pixel of the same neuron, segmented the volume. Separate networks detected synapses: the preprint reports 46 million presynaptic sites linked to 312 million postsynaptic densities, found with an average precision of 0.82 and recall of 0.81. Further models detected cell nuclei and predicted each synapse's likely neurotransmitter.

Human experts then did what the models could not. The preprint describes 29 proofreaders working over three years, an estimated 44 person-years in total, correcting segmentation errors, alongside teams that annotated every neuron and assigned its cell type. The authors report that 98.9% of the 141,780 detected neuron nuclei belong to a proofread neuron.

How complete is the wiring map?

Very complete by the field's standards, though not perfect. The preprint reports that 94% of presynaptic sites and 42% of postsynaptic sites in neuropil are attached to proofread neurons, figures the authors describe as matching or exceeding earlier published connectomes.

They also propose a stricter measure: the share of synaptic connections where both partners are proofread neurons. For the male nervous system that figure is 40.1%. The resulting graph links 166,391 neurons through 25.6 million edges. A small number of cells could not be reconstructed because of sample artefacts at the edge of the volume or segmentation problems.

Cross-referencing makes the data more reusable. Using shape and connectivity similarity, the team matched 97.5% of neurons to cell types in at least one earlier Drosophila connectome dataset. Along the way they revised about 4% of FlyWire neurons, compared with more than 44% of cell types in an earlier effort to match the hemibrain to FlyWire, a sign that cell type definitions are settling.

What does the map reveal about sex differences in the brain?

Most neuron types are shared by both sexes, and the differences cluster in higher brain centres. Comparing the male map with the female brain connectome at synaptic resolution, the Cell paper reports 8,069 isomorphic neuron types, 138 dimorphic types, 289 male-specific types and 71 female-specific types.

Sensory and motor neurons at the periphery are largely the same in both sexes. Inside higher brain centres, male-specific connections form hotspots built around male-specific neurons or extra branches, and dimorphic neurons route the same information differently in males and females. In one example from the post, a male neuron carries two extra projections that its female counterpart lacks.

The preprint adds that fruitless and doublesex expression lines up strongly, but not perfectly, with dimorphic wiring. More than 1,500 neurons expressing these factors showed no wiring difference, which the authors suggest may mean the factors also regulate properties other than connections, although technical limits may play a part.

Limitations and open questions

The authors are explicit that a wiring diagram is a starting point. Their preprint lists several limits, and comparing one male brain with one female brain raises statistical problems of its own.

• One animal per sex. Comparative connectomics still has very few samples. The authors recommend comparing cell type graphs rather than raw connectome graphs, and they found expert curation was needed to separate true biological differences from technical noise.
• Uneven comparison partners. Thorough cross-sex comparisons were possible only against the female brain from FAFB/FlyWire. Female nerve cord datasets were partial or still in progress, which limited comparisons there.
• Structure, not function. The connectome shows wiring, not activity. Predictions about information flow and behaviour still need functional experiments to confirm them.
• A snapshot. The map captures one animal at one point in its life. Developmental divergence may hide neurons that share an origin but differ too much in shape, and experience can modify even strongly innate behaviours.
• Coverage gaps. Some cells at the edge of the volume could not be reconstructed, and technical limits prevented exhaustive annotation of fruitless and doublesex expression in the nerve cord.

From flies to fish and mice

Vertebrates are next. The post describes work led by Columbia University and published in Nature that mapped part of the elephantnose fish hindbrain to study a circuit for sensory prediction, and an upcoming whole-brain larval zebrafish resource with Harvard that pairs neural structure with molecular cell types.

Methods keep improving. Google's reconstruction system PATHFINDER gained speed and accuracy after synthetic neurons were added to its training data. The goal is to cut the years of manual verification that a fly brain still requires, so that larger projects, including a portion of the mouse brain, fit within realistic budgets and timelines.

Three companion papers already use the male connectome to study vision, taste and social behaviour.

Why does this matter for applied and enterprise AI?

Analysis: the connectome is a case study in building a trustworthy dataset at a scale no team could label by hand. The pattern will be familiar to any organisation turning large volumes of raw data into a reference resource.

• Automation first, verification always. Flood-filling networks and synapse detectors made the project feasible, but 44 person-years of proofreading made it reliable. Plans for AI data pipelines should budget for the human review that turns model output into a trusted asset.
• Report quality honestly. The authors publish precision, recall and completeness figures, and propose a stricter completeness measure than the usual one. Clear, conservative quality metrics make a dataset easier to trust and reuse.
• Shared vocabularies multiply value. Matching cell types to earlier datasets, and recording each dataset's names side by side, lets others combine resources. Enterprise data teams face the same task when reconciling entities across systems.
• Better models shrink the review bill. Improvements such as synthetic training data aim to reduce manual correction, which still takes years of effort. The same lever applies to document, image and record processing: better models mean less human review per item.

How to access the data and tools

The data is public. A landing page at male-cns.janelia.org gathers images, segmentation, neuron reconstructions, the connectome graph, neurotransmitter predictions and annotations, with interactive tools such as neuPrint and Clio and bulk downloads. The site states that the dataset is licensed under CC BY.

The volume can also be explored in Neuroglancer, the open-source viewer for large multidimensional datasets that Google created and publishes under the Apache 2.0 licence. Derived data specific to the paper is on GitHub in the flyconnectome 2025malecns repository.

Questions and answers

References

1. Berg, S., Beckett, I. R., Costa, M., Schlegel, P., Januszewski, M., Marin, E. C., Nern, A., Preibisch, S., Qiu, W., Takemura, S. Y., Fragniere, A. M. C., Champion, A. S., Adjavon, D. Y., Cook, M., Gkantia, M., Hayworth, K. J., Huang, G. B., Katz, W. T., Kämpf, F., ... Jefferis, G. S. X. E. (2026). Sexual dimorphism in the complete Drosophila male central nervous system connectome. Cell, 189(18), 5504-5526.e15. https://doi.org/10.1016/j.cell.2026.08.015 (external site)
2. Berg, S., Beckett, I. R., Costa, M., Schlegel, P., Januszewski, M., Marin, E. C., Nern, A., Preibisch, S., Qiu, W., Takemura, S. Y., Fragniere, A. M. C., Champion, A. S., Adjavon, D. Y., Cook, M., Gkantia, M., Hayworth, K. J., Huang, G. B., Katz, W. T., Kämpf, F., ... Jefferis, G. S. X. E. (2025). Sexual dimorphism in the complete connectome of the Drosophila male central nervous system [Preprint]. bioRxiv. https://doi.org/10.1101/2025.10.09.680999 (external site)
3. Januszewski, M., Kornfeld, J., Li, P. H., Pope, A., Blakely, T., Lindsey, L., Maitin-Shepard, J., Tyka, M., Denk, W., & Jain, V. (2018). High-precision automated reconstruction of neurons with flood-filling networks. Nature Methods, 15(8), 605-610. https://doi.org/10.1038/s41592-018-0049-4 (external site)
4. FlyEM Project Team, HHMI Janelia Research Campus. (2026). Male CNS connectome (Version 1.0) [Data set]. https://male-cns.janelia.org/ (external site)
5. Google. (2016). Neuroglancer: Web-based volumetric data visualization [Computer software]. GitHub. https://github.com/google/neuroglancer (external site)