Summary
Berg et al. published the first complete connectome of an adult Drosophila melanogaster male central nervous system — the entire brain, both optic lobes, and the ventral nerve cord (VNC, the insect analogue of the spinal cord) reconstructed as a single connected volume. With 166,691 neurons and ~25.6 million connections (edges) between them, it is the largest brain map by neuron count produced to date, and the first connectome to span brain and nerve cord in one animal rather than the brain alone. Published in Cell (online 3 September 2026; bioRxiv preprint, DOI 10.1101/2025.10.09.680999). It is also the first male connectome, enabling the first cellular-resolution male–female comparison of neural wiring in any animal.
Scale and method
Organism: adult male Drosophila melanogaster, complete central nervous system with intact neck connective joining brain and VNC.
Numbers:
- Neurons: 166,691 (including sensory axons)
- Connections (edges): ~25.6 million directed neuron→neuron edges
- Synapses: 46 million presynaptic sites; 312 million postsynaptic densities
- Cell types: 11,691 unique annotated types
- Graph density: extremely sparse — a fully connected graph over 166,700 nodes would have ~27.8 billion edges, versus ~25.6M actual (~0.02%); mean partners per neuron ≈ 37
Imaging: 7 enhanced focused-ion-beam scanning electron microscopy (eFIB-SEM) systems over 13 months, producing a 160-teravoxel volume at 8×8×8 nm isotropic resolution.
Reconstruction: automated flood-filling-network segmentation followed by ~44 person-years of human proofreading; automated synapse detection at ~0.82 precision / 0.81 recall. Completion: ~94% presynaptic and ~42% postsynaptic completion in neuropil; ~40% of edges with both partners proofread.
Key findings
Sparse, recurrent, small-world graph: The connectome is not layered or densely connected like a typical artificial neural network. It is a sparse directed graph — closer to a graph neural network in which each node exchanges signals only with its ~37 partners. Despite the sparsity, >90% of neurons are mutually reachable within one giant component (short path lengths), and nearly every neuron participates in a recurrent loop; feedforward-only flow is the exception.
Weights as fractional input share: Connection strength is given by synapse count, but functionally as a share of a target neuron’s total input, not an absolute value. A neuron rarely fires from a single input; crossing threshold requires coincident input from several partners — the circuit reads coincidence, not single signals. Inhibitory inputs enter the same summation as negative terms, so excitation and inhibition compete in one weighted sum.
Sexual dimorphism at synaptic resolution: comparing the male CNS to the earlier female brain, the study classifies cell types as 7,205 isomorphic, 114 dimorphic, 262 male-specific, and 69 female-specific, mapped against fruitless/doublesex expression. This is the first genome-linked, wiring-level catalogue of how a nervous system differs between sexes.
Limitations
- Some cells near volume edges or with segmentation artifacts remain unreconstructed (e.g., certain sensory/motor neurons, Johnston’s-organ auditory afferents).
- No fully proofread female VNC connectome existed, so female-side dimorphism labels for the nerve cord had to be extended from the literature rather than measured.
- Technical limits prevented exhaustive fruitless/doublesex annotation in the VNC.
- The connectome records wiring (synapse counts), not learned synaptic efficacy — it shows where learning can occur, not what an individual fly has learned. Two flies that learned opposite associations would yield near-identical maps.
Connection to AI and ALife
The male CNS connectome sharpens several structural contrasts with artificial neural networks:
- Architecture is “pre-trained” by evolution: who connects to whom, and with how many synapses, is largely genetically specified rather than randomly initialized — an inherited prior found by evolution’s outer loop, related in spirit to the genetically-fixed wiring already visible in C. elegans (white-1986-the-structure-of-the-nervous) and the larval brain (winding-2023-the-connectome-of-an-insect).
- Graph structure over layered structure: sparse, directed, recurrent connectivity resembles a graph neural network more than a dense MLP; the ~65% recurrence seen in the larva reappears here at 50× scale, reinforcing that biological computation is fundamentally recurrent (hochreiter-1997-long-short-term-memory).
- The perceptron re-read: mapping real neural wiring at scale invites re-examining Rosenblatt’s perceptron (rosenblatt-1958-the-perceptron-a-probabilistic-model) and the McCulloch–Pitts neuron (mcculloch-1943-a-logical-calculus-of-the) — thresholded weighted sums with competing excitation/inhibition are visible in the biology, but embedded in a sparse recurrent graph rather than dense layers.
- What the connectome does and does not give: it provides the immutable structure (nodes, edges, synapse counts) but not the learned efficacies — a reminder that in biology the “weight” is count × efficacy, and only the count is in the map.
Related connectomes
| Organism | Neurons | Synapses | Status |
|---|---|---|---|
| C. elegans | 302 | ~7,000 | Complete (1986) |
| Drosophila larva | 3,016 | 548,000 | Complete (2023) |
| Drosophila adult female brain | ~140,000 | ~50M | Complete (FlyWire, 2024) |
| Drosophila adult male CNS (brain + VNC) | 166,691 | ~358M sites | Complete (2026) |
| Mouse | ~71M | ~100B | Partial |
| Human | ~86B | ~100–500T | Not started |