AI Segmentation and Human Proofreading Complete the Male Fruit Fly’s Full Central Nervous System Connectome, Revealing 125 Million Synapses

Google Research, HHMI Janelia, the University of Cambridge, and the UK Medical Research Council Laboratory of Molecular Biology (MRC LMB) have formally published the first complete connectome of the adult male fruit fly central nervous system. Extending from the central brain and optic lobes to the ventral nerve cord—the insect equivalent of the spinal cord—the dataset contains more than 166,000 neurons and approximately 125 million synapses. Unlike previous atlases limited to the brain, the new connectome lets researchers trace, within a single map, how sensory inputs cross the neck connective and reach motor neurons that control movements such as those of the legs.

The reconstruction process began with nanoscale imaging using focused ion beam scanning electron microscopy. Google’s machine-learning system then stitched the image sections together and performed the initial 3D segmentation. Its core technical lineage includes flood-filling networks, in which a convolutional neural network starts from a single pixel and repeatedly predicts which neighboring pixels belong to the same cell. The automated output was not the final result: experts at Janelia and Cambridge still had to correct erroneous joins individually, annotate synapses, and classify cell types. The achievement therefore also exposes the current limits of AI for connectomics—the models dramatically narrow the search space but have not eliminated costly human proofreading.

The new data also enable the first comparison of complete adult male and female nervous systems at synaptic resolution. The team identified 262 sex-specific and 114 sexually dimorphic cell types. Together, they account for only about 4.8% of the central brain, but their connectivity differences can propagate across much broader networks. This provides testable candidate circuits for studying courtship, aggression, and sensorimotor transformation, rather than requiring researchers to infer local neuronal roles from behavior alone.

The dataset is licensed under CC BY. Researchers can query connectivity through NeuPrint, inspect imagery with Neuroglancer, or conduct programmatic analyses using Python and R packages. An important caveat is that a connectome describes structure: it does not capture time-varying synaptic strength, neuromodulation, or the system’s complete activity state. The data also come primarily from a single male specimen. Key next steps will include reconstructing additional individuals, aligning the connectomes with activity imaging and gene-expression data, and testing whether AI segmentation can further reduce human proofreading while preserving auditable accuracy.