Functional connectomics spanning multiple areas of mouse visual cortex

Nature 640:435–447. Published 9 April 2025. DOI: 10.1038/s41586-025-08790-w. Peer-reviewed primary data/resource paper; open access, CC BY 4.0. Read 29 September 2026.

What was read

Read the entire final publisher HTML, archived at papers/consciousness_connectomics/microns2025_functional_connectomics.html: abstract, every main-text section, complete Methods, availability statements, references, acknowledgements, contributions, declarations and all captions. The normalized reading extraction contains 2,059 lines. Three display equations omitted by the paragraph extractor were separately checked in the HTML. All seven main figures, five Extended Data figures and three Extended Data tables were visually inspected. The complete three-page Reporting Summary was read.

This article lists no separate substantive supplementary methods document: its methods are online, and the separate supplement is the Reporting Summary. Peer-review correspondence, linked companion articles, code, full image volumes, annotation tables and recordings were not independently audited. No reanalysis or replication was performed. In particular, this reading does not independently validate the companion functional-model or like-to-like wiring results. References and search-result snippets are not counted as separately read papers.

Exact downloads, source URLs and SHA-256 hashes are recorded in papers/consciousness_connectomics/microns2025_functional_connectomics.provenance.json. Figures, tables and Reporting Summary are archived under that prefix. Reported counts describe the paper's release, including proofreading through 16 September 2024, rather than the latest live resource.

Question and main contribution

How can synapse-scale anatomy be paired with functional responses in the same individual neurons of the same animal, at a scale large enough to span cortical layers and visual areas?

The paper releases a cubic-millimetre-scale mouse visual-cortex dataset combining in-vivo calcium imaging, visual stimuli, movement/pupil records and subsequent destructive electron microscopy. It includes estimated functional responses for 75,909 unique excitatory neurons and an EM reconstruction containing over 200,000 cells and 524 million detected synaptic clefts. This is a partial cortical volume in one male mouse, not a whole mouse brain or population atlas.

Its key advance is the capacity to ask anatomical and physiological questions about matched cells. Its main result is a resource and validated measurement pipeline. Extensive functional-model and circuit findings are reported in companion studies and described here as examples.

Animal, spatial coverage and temporal scope

• One mouse expressing GCaMP6s in excitatory neurons through Slc17a7-Cre and Ai162. Thirty-one animals underwent surgery; eight were selected using surgical/optical/tissue criteria, then one was chosen for EM. This is quality-based specimen selection, not random sampling of a population.
• Awake, head-restrained imaging while the mouse could walk on a one-axis treadmill. Visual stimuli were shown to the left eye; pupil size, eye movement and locomotion were recorded.
• Functional target approximately 1,200 × 1,100 × 500 μm, spanning layers 2–5 in VISp and higher visual areas VISlm, VISrl and VISal. The EM volume is approximately 1.3 × .87 × .82 mm in in-vivo coordinates and reaches from near pia toward white matter. The functional and structural volumes/coverage differ.
• Fourteen selected scans from 19 completed scans across six days. Eleven used 6.3 Hz imaging, two 8.6 Hz, one 9.6 Hz. Planes were sampled sequentially; the 75,909 neurons were not all simultaneously recorded with single-spike resolution.
• The displayed natural/synthetic stimuli total approximately 84 minutes per scan, with spontaneous periods before/after. The resource measures responses in a bounded set of visual and behavioural conditions, not the mouse's lifetime learning or full behavioural repertoire.

Functional measurement and its limits

The authors obtain somatic regions of interest with constrained non-negative matrix factorization, extract fluorescence, and deconvolve activity estimates. Of 125,413 masks, 115,372 were classified as somatic. Since some cells appear in multiple scans, registered centroids were greedily grouped by proximity, yielding an estimate of 75,909 unique cells. This estimate differs from the number of cells with verified EM matches.

The calcium indicator was expressed in excitatory neurons. Although inhibitory cells and glia are structurally reconstructed, the dataset does not provide equally comprehensive functional recordings for them. The present release extracts somatic rather than exhaustive dendritic-compartment signals. Deconvolved calcium activity is an indirect, temporally filtered estimate, not a direct exhaustive record of action potentials.

Stimuli comprise cinematic clips, sports clips and rendered first-person environments, plus two parametric patterns: global directional “Monet2” and locally varying “Trippy.” Six ten-second natural clips are repeated ten times per scan. Their oracle score is a leave-one-repeat-out response correlation: it measures repeatability of responses to those clips, not accuracy of a reconstructed brain model.

Depth-dependent photon scattering and out-of-plane fluorescence can contaminate signals. Apparent differences across layers should therefore be checked using matched-depth controls or another recording technique. Separate sessions also differ in physiological state and technical quality; locomotion and pupil measurements help but do not exhaustively recover state.

Structural acquisition and reconstruction

Following in-vivo recordings, the brain was fixed and tissue stained, embedded, physically sectioned and imaged. The process does not preserve a living functioning brain.

• 27,972 nominal 40-nm sections were collected; 26,652 imaged using five customized TEMs over approximately six months, producing about 2 petabytes of raw images at approximately 4-nm in-plane resolution.
• An approximately 800-μm sequence with around 0.1% section loss and no consecutive section loss was selected for reconstruction. Earlier imaged portions with more damaging losses were excluded.
• Knife changes/re-trimming led to separate 35% and 65% subvolumes. They were aligned into one coordinate frame but reconstructed and represented separately. All reported proofreading was in the larger subvolume, Minnie65.
• Images were aligned using coarse and fine methods; final reconstruction ran at 8 × 8 × 40 nm. Locations with insufficient alignment or missing/occluded tissue were masked out rather than guaranteed recovered.
• Dense segmentation used affinity-predicting neural networks and graph clustering. Additional networks detected nuclei, synaptic clefts and synaptic partners.

Accuracy and completeness

The detected clefts total approximately 186 million in Minnie35 and 337 million in Minnie65. Manual identification of 8,611 synapses in 70 small test volumes gave 96% precision, 89% recall and 92% F1 for detection. Extended Data Figure 1 places those test volumes in Minnie65. Separate held-out partner assignments (191 synapses) yielded 98% accuracy. Detection accuracy, partner-assignment accuracy and successful connection to a full neuronal soma are separate quantities.

The release incorporates 1,046,656 edits. Only 1,433 neurons have proofread axons, with variable extension and completeness. For a functional study, 85 excitatory cells had full axonal/dendritic extension within the subvolume; another sampling effort proofread dendrites of 1,188 excitatory cells. Additional proofreading strategies deliberately target local cylinders, inter-area projections, or at least 100 synapses. Such policies create non-random observation patterns.

Postsynaptic dendritic assignments that already reach a soma were approximately 99% correct, according to comparison with proofread dendrites. This is a precision statement. It does not mean 99% of every input/output synapse is mapped, nor that missing axons or detached spines disappear. Typical proofread axon output-to-soma mapping ranges from about 60–95%, depending on cell type and position. The worked example maps 1,053 of 1,412 outputs to single-nucleus objects (74.5%); 359 remain on orphan fragments. Only 365 of these outputs connect to functionally matched targets.

The separate result 84,035 individually segmented neurons follows correction of merged multi-soma objects in the larger subvolume. It should not be conflated with either the total cell count across both subvolumes or the functional neuron estimate.

Matching function to structure

Fiducial matching and registration align in-vivo optical data with fixed-tissue EM. The reported average leave-one-out registration residual is approximately 3.8 μm. Registration was then used to support manual and automated cell matching.

Rows are alternative/overlapping tables, not additive samples. Multiple ROIs can correspond to the same cell across scans. “Separation” measures how much farther away the nearest alternative cell lies than the selected match after transformation; small residuals and large separation improve confidence.

Manual matches were cross-checked by asking whether ROIs independently matched to the same cell in different scans had more similar visual responses than nearby unmatched cells. They did, particularly for reliably responsive cells. Manual matching itself prioritizes strong responses, so representativeness and absolute error rates remain distinct from relative automatch agreement.

What the resource demonstrates scientifically

The images and annotations support analyses of neuronal morphology, cell classes, interlaminar connectivity, multi-hop inhibitory/excitatory circuits and structure–function relations. Figure 7 illustrates connectivity across cortical depths and a seven-layer-3-neuron example tracing direct excitatory outputs and indirect routes through inhibitory neurons. These are anatomical examples of pathway specificity, not a general theory of cortical computation or consciousness.

The authors make an especially useful distinction for individual modelling. In flies, many cell types correspond to very few stereotyped cells, so anatomy in one individual can often guide experiments in another. A mammalian cell type contains many neurons with different functional tuning. Cell-type identity alone does not specify an individual cortical neuron's functional role. This motivates recording the same cells that are reconstructed rather than assigning an atlas's typical dynamics to them.

The article also describes a “functional digital twin,” built in companion work from neural-response training data, including shared training across mice. In this paper the phrase means a model predicting visual cortical activity under visual stimulation. It does not denote a simulation of the mouse's entire brain, body, memory, identity or subjective experience. The paper does not itself supply the full model comparison needed to isolate the incremental predictive value of individual wiring.

Important limits and internal discrepancies

• One selected male, no experimental replication, randomization or blinding, as explicitly reported in the Reporting Summary. Millions of cells/synapses do not become independent mice.
• A cortical crop truncates long-range axons and inputs from outside the volume. All-layer local structure is not whole-area or whole-brain circuit closure.
• Proofreading depth varies by scientific priority; detached spines, optical artefacts, volume boundaries and functional-match confidence can create structured bias.
• Many molecular properties, synaptic efficacy/plasticity and ongoing neural states are not measured by the released static graph. Some organelle and glial morphology is available in images, but that is different from modelling their dynamics.
• There is no consciousness-state contrast or theory test. Awake visual responses are not an assay establishing which aspects cause subjective experience.
• The Methods timeline contains an unresolved inconsistency: it gives a structural stack on 21 March 2018 (P83) after perfusion on 16 March (P87), while a 19 December birth date is supplied. Both ordering and age/date labels cannot all be correct. The main text also describes P75–P81 whereas the timeline ends functional imaging at P80. Exact chronological claims beyond the general recording-then-fixation sequence need author/data clarification.
• The stated fiducial total, 2,934, differs from its listed components, 1,994 somata + 942 vessel points = 2,936. A later passage refers to 2,933 leave-one-out transforms. These inconsistencies do not erase the same-cell matching result but matter if reconstructing the pipeline exactly.
• The proofreading methods include a sentence cut off after “Any axon leaving the cylinder was cut and”. The exact continuation is not present in the archived publisher HTML. This reading reports the available method rather than silently inventing it.

Implication for Kurisutina

The strongest lesson is a measurement strategy: pair individual structure with individual function, preserve uncertainty at each transformation, and test the mapping empirically. Shared cell-type rules are a prior; they do not recover all individual response properties. Conversely, predicting responses from activity-trained models does not show that a wiring diagram alone was sufficient.

My inference is that a structural replica, neural-response predictor, adaptive behavioural replica and conscious continuation remain separate claims. This paper establishes a tractable bridge between the first two at a partial-cortex scale. It neither settles whether a sufficiently detailed connectome could support whole-brain emulation nor demonstrates that such emulation preserves a particular subject.

A next study could compare individual-connectome, cell-type/atlas-connectome and structure-free models given matched functional observations, then evaluate held-out stimuli, changed internal states and causal perturbations. Missing-edge and matching-confidence sensitivity analyses would be essential to interpreting that comparison. That is a research proposal motivated by this resource, not an experiment reported here.