Neuronal wiring diagram of an adult brain

Nature 634:124–138. Published 2 October 2024. DOI: 10.1038/s41586-024-07558-y. Peer-reviewed primary reconstruction/resource paper; open access, CC BY 4.0. Read 29 September 2026.

What was read

The final publisher version, archived as papers/consciousness_connectomics/dorkenwald2024_adult_fly.html, was read in full: abstract, introduction, every results/discussion subsection, complete online Methods, data/code statements, references, acknowledgements, contributions and declarations. The normalized reading extraction has 2,620 lines. The display equation omitted by the paragraph extractor was checked separately against the HTML. All seven main figures and ten Extended Data figures were inspected as images, and all captions read.

The publisher's Supplementary Information DOCX was read completely: Supplementary Figures 1–10 (images inspected), Tables 1–3 (all rows), and the two video legends. The three-page Reporting Summary was scanned rather than text-based, so all three pages were rendered and read visually. The FlyWire and Eyewire supplementary author/affiliation notes were also read fully. Supplementary Videos 1–2 themselves were not watched; these are rotating renderings of neurons and neuropils, whose legends were read. Peer-review correspondence, linked companion papers, code, raw microscopy, and released connectivity matrices were not audited. No empirical analysis or replication was run. A paper appearing in the reference list or a search result was not treated as independently read evidence.

Exact source URLs, retrieval date and downloaded-file SHA-256 hashes are in papers/consciousness_connectomics/dorkenwald2024_adult_fly.provenance.json. Main and extended images and supplementary documents are stored under the same filename prefix. All quantitative claims below concern the paper's release, version 783, not the latest live database.

Question and contribution

Can one reconstruct the chemical wiring of an entire adult fly brain, and does whole-brain coverage reveal pathways unavailable in cropped reconstructions? The main accomplishment is a reusable anatomical measurement resource: 139,255 reconstructed neurons with 54.5 million chemical synapses between them, spanning both optic lobes and the central brain. Morphology, synapse locations, annotations and predicted neurotransmitter identities support circuit hypotheses.

This is a whole-brain reconstruction of one individual, not a population-average brain, not the fly's complete nervous system, and not an operating simulation. The ventral nerve cord, most sensory organs and muscles are outside the volume. Linking it to nerve-cord reconstructions from other flies would create a cross-individual composite.

Specimen and reconstruction

• One seven-day-old adult female, (iso) w1118 × (iso) Canton-S G1, selected from screened brains. The paper reuses the FAFB electron-microscopy acquisition published earlier rather than collecting longitudinal activity from this fly.
• Serial-section transmission EM at 4 × 4 × 40 nm; a re-aligned FAFB14.1 volume was automatically segmented using boundary-detecting convolutional networks and mean-affinity agglomeration.
• Centralized teams, collaborating laboratories and citizen scientists corrected segmentation. The effort is estimated at 33.1 person-years, based on 3,013,513 edits, approximately 79 seconds/edit and a 2,000-hour work year. This is a rough labour estimate, not the project's total cost.
• Proofreading emphasized microtubule-rich neuronal backbones. Small postsynaptic twigs were restored incidentally or for specific scientific interests. After cells with nuclei, remaining segments were prioritized by synapse counts and strong connections.
• Automated chemical-synapse predictions were filtered using cleft scores, assignment to segments and duplicate removal. Approximately 244 million original predictions became about 130 million filtered synapses. Only a subset have both ends attached to the proofread neuron set.
• The stored image orientation is left–right inverted relative to the biological fly; annotations correct the labels, while figures retain the mirrored underlying imagery. Comparisons must respect this convention.

The connectome is therefore the result of image acquisition, segmentation, synapse detection, partner assignment and annotation. These are separate measurement steps with different error profiles.

Completeness and accuracy: distinct quantities

Attachment is conditional on automated detections; it is not an independent estimate of all biological synapses. The authors explicitly state that synapses themselves were not manually proofread and that detector performance varies by region and cell type, particularly for sensory neurons. Detection training relied on a small calyx ground-truth sample, although additional regions were evaluated in the underlying work.

The five-synapse threshold is described in Methods as reasonable but arbitrary. There is no bimodal distribution identifying a natural boundary. It reduces false positive edges while discarding some genuine weak connections. The prediction that multisynaptic connections have higher reliability assumes sufficiently independent errors; region-specific imaging and detection problems can violate simple independence assumptions.

Increasing postsynaptic attachment from 44.7% to 50% would require proofreading over 700,000 further fragments. The authors expect better acquisition/alignment/segmentation, rather than simply more manual work, to solve this bottleneck. A partially severed left lamina is a known asymmetric artefact. Whole-brain coverage and complete synaptic recovery are different achievements.

Main anatomical results

Centralization and interfaces. Of the neurons, 118,501 are intrinsic to the brain, meaning their synapses lie within it. They make up 85% of reconstructed brain neurons. There are 1,303 descending and 2,362 ascending neurons, 106 motor neurons and 80 endocrine neurons. Inputs and outputs are anatomically constrained, although mixed pre/post synapses along fly neurites make simple feedforward labels imperfect.

Visual dominance and feedback. Among intrinsic neurons, 77,536 are confined to optic lobes/ocellar ganglia and 32,388 to the central brain. There are 8,053 visual projection neurons toward the central brain and 524 visual centrifugal neurons toward the optic lobes. Despite the lower centrifugal cell count, roughly half of optic-lobe neurons receive at least five centrifugal synapses. Early visual processing has extensive anatomical access to central feedback.

Size, degree and compartmentalization. Intrinsic neurons have a median reconstructed arbor length of 685 μm. Median in/out degrees under the five-synapse threshold are 11/13. Large neurons tend to contact more partners rather than merely strengthen a fixed partner set: synapse totals correlate strongly with degree. Yet a biological neuron need not be one computational node. CT1, with thousands of partners, has partly independent subcellular compartments; the locations of individual synapses permit models at that finer scale.

Hemispheres and regional projections. Most optic-lobe processing is ipsilateral; only 139 optic-lobe neurons cross hemispheres without making central-brain synapses. Around 40% of central-brain neurons are non-ipsilateral. A projectome over 78 neuropils aggregates each neuron's normalized input/output distributions, giving each neuron total weight one. This is a structural compression under an input/output independence assumption; it is not a directly measured communication capacity or physiological effective-connectivity matrix.

Traversal analysis. A probabilistic graph traversal estimates how near each neuron is to sensory or ascending input classes. Recruitment probability increases with the fraction of inputs from already reached neurons and reaches one at 30%. Ten thousand runs per seed class are averaged; resulting ranks are embedded with UMAP. Nearly all central neurons can be reached from each modality, different modalities reach output classes in different orders, and the central complex tends to be structurally distant from direct sensory input. The method ignores connection sign and biophysics. Its early/late ranks are not response latencies, and its use of “information flow” does not establish actual transmission during behaviour.

Ocellar circuit example. The 273 ocellar photoreceptors feed distinct ganglion compartments. OCG01 neurons supply putatively inhibitory ipsilateral and excitatory contralateral inputs to downstream neurons; input amounts covary (R = .78). Some descending neurons also receive strong optic-lobe input. This suggests a light-difference circuit supporting head/body righting and visual stabilization. It is an explicitly testable functional hypothesis from anatomy, not a causal demonstration of the proposed behaviour in this paper. Twenty-five centrifugal neurons provide additional feedback whose function remains undetermined.

What neurotransmitter labels do and do not provide

Six transmitter identities are predicted from EM appearance: acetylcholine, GABA, glutamate, serotonin, dopamine and octopamine. Majority aggregation across a neuron's synapses assumes one small-molecule transmitter per neuron. Reported validation from the cited classifier is 87% per-synapse and 94% per-neuron accuracy; these were not independently reproduced here.

The analysis provisionally treats acetylcholine as excitatory and GABA/glutamate as inhibitory. These are useful assumptions for fly circuits, but co-transmission exists and a transmitter label is not a full measurement of postsynaptic receptor composition, conductance or state-dependent effect. The discussion itself notes that inhibition can be shunting and that inhibitory conductance can greatly differ from excitatory conductance. A synapse count consequently constrains, but does not directly measure, its physiological weight.

Limits relevant to consciousness and person reconstruction

1. No electrical-synapse map. Gap junction identification requires higher-resolution data than this reconstruction provides. Chemical edges are not the whole interaction network.
2. No living dynamic state. The graph contains neither recorded spike trains nor membrane potentials, transient internal states, ongoing learning or moment-to-moment neuromodulation from this individual.
3. Incomplete non-neuronal description. Around 13% of cell bodies are estimated non-neuronal/glial; only a few glia were proofread. The images contain further biological information that is not represented by the neuron/synapse graph.
4. One selected female at one age. Broad stereotypy makes the resource useful across flies, but the paper explicitly highlights sexual dimorphism and variable mushroom-body connectivity. Agreement in cell types or strong connections does not establish recovery of an individual's learned memories.
5. No consciousness experiment. There is no comparison of conscious/unconscious states, report, perturbational consciousness assay or test of a theory of subjective experience. Neither the fly's experience nor a simulation's consciousness is measured.
6. No emulation sufficiency theorem. Companion simulations use additional assumptions about weights, signs and neuron dynamics. The existence of those models does not show this measurement alone is sufficient to reproduce all functions or the identity of the scanned animal.

These limits do not prove that every omitted molecular detail is necessary for useful emulation. They establish that sufficiency must be tested rather than inferred from the word “connectome.”

Implication for Kurisutina

The paper strengthens a structural route to constrained mechanistic models. A connectome can sharply reduce the space of plausible circuits and reveal recurrent, inter-regional and sensorimotor organization that behaviour alone may leave ambiguous. It supplies anatomy that an ordinary personal interview cannot supply.

My inference is that the project's existing distinction between an archive, a response predictor and an individual dynamics model must be extended to include structural fidelity as another separately measured property. Recovering edges, predicting neural responses, reproducing adaptive behaviour, retaining particular memories, and establishing consciousness require different evidence. This paper establishes a major advance in the first and helps formulate experiments for the second and third. It establishes neither personal continuity nor consciousness preservation.

A useful next research test would compare a structure-constrained model with an equal-data model without individual wiring on held-out activity and perturbational responses, while auditing edge uncertainty and what physiological parameters were fitted. This is a proposed design, not a result of this paper.