Maps / H01

H01: one cubic millimetre of human cortex, mapped cell by cell

H01 is a piece of human temporal cortex, about one cubic millimetre, imaged with an electron microscope at nanometre resolution and reconstructed in 3D by Google Research and the Lichtman lab at Harvard (Shapson-Coe et al. 2024, Science). The tissue was removed during surgery for epilepsy, so it comes from a living adult human brain, and it is released openly under CC BY 4.0.

What is in it

Volumea slab about 3 × 2 mm wide and 0.17 mm thick ("just over one cubic millimetre")
Images5,019 sections, ~34 nm thick, 4 nm pixels — about 1.4 petabytes
Cells57,216 in total: 16,087 neurons, 32,315 glia, ~8,100 blood-vessel cells (paper, Table S5)
Synapses~150 million detected automatically
Layersall six cortical layers plus some white matter

How to look at it yourself

  • Neuroglancer (in the browser): the H01 team's gallery opens ready-made views — whorled axons, a dendrite through a cell body, glia glued to neurons, every blood vessel. Start at the H01 gallery.
  • Raw data: gs://h01-release/data/20210601/ (images, segmentation c3, synapses, cell bodies).
  • Two synapse releases. Only one (20210729) matches the paper's per-cell tables. We used the wrong one at first — here is how to tell them apart.

Things to keep in mind

  • One sample, one person. The tissue comes from epilepsy surgery; whether it is typical of healthy cortex is an open question the authors raise themselves.
  • Automated reconstruction. Cells can be split or merged by mistake; the "strange" structures in the gallery could be biology or glitches.
  • Edges. Cells are cut off at the sample boundary, so counts per cell are lower bounds.

Numbers we measured from H01

99.4%of classified synapses in this human cortex sample land on dendrites; 0.394% on cell bodies and 0.197% on the axon initial segment
How: Paper main text (author manuscript p. 7): "Analyzing the post-synaptic component of 133,704,881 of these ... a high percentage of these synapses located on dendrites (99.4%), with far fewer innervating the AIS (0.197%) or somata (0.394%)". Supplementary materials p. 24: the postsynaptic structure is the closest classified skeleton node of the postsynaptic segment. Fig. 5 legend (p. 27): on a representative pyramidal neuron, synapse density is "highest along the AIS and strikingly sparse on the soma". Table S4 (SM p. 51): dendrites fill more of the tissue than cell bodies in every layer.
Caveat: Shares are of the 133.7 million synapses whose target could be classified; segments too small to classify (8.8%) and some others are not in the denominator. The video shows one neuron's input synapses as an illustration; the percentages are for the whole sample. One sample (human temporal cortex, epilepsy surgery).
0.17 mmthick: the "cubic millimetre" of human cortex (H01) is a thin slab, not a 1 mm cube
How: Paper main text (author manuscript p. 5): "170 micrometer thick slab of human cortex ... with a total volume of just over one cubic millimeter"; "approximately 3 millimeters from layer one to the white matter"; 5,019 sections, average 33.9 nm, "total sample thickness of 0.170 millimeters"; imaged volume 1.05 mm³ (uncompressed). The paper says the imaging throughput constrained the thickness. 1 / 0.17 ≈ 6 slabs to reach 1 mm (ours). The video shows the released EM volume at low resolution: one whole section (front) and one cut across all sections (side) at the same scale.
Caveat: The 3 mm is the paper's approximate span from layer 1 to white matter; the slab's outline is irregular, not a box. The video's images are not corrected for the 28% compression in the cutting direction. One sample (human temporal cortex, epilepsy surgery).
33.7%of one human neuron's 1,550 dendritic spines were not attached to it in the public automatic H01 map (about 1 in 3); the older map version was about the same (32.2%)
How: Paper values. Main text (author manuscript p. 6): "proofreading 1550 spines on the dendritic tree of a layer 2 pyramidal neuron ... 32.2% of its spines detached from dendritic shafts in the c2 segmentation, vs. 33.7% in the c3 segmentation". Supplementary materials p. 18: the 1,550 spines were identified by hand, and the share not part of the neuron's automatically agglomerated segment was computed.
Caveat: One neuron, checked by hand; the rate elsewhere in the volume may differ (the paper notes split errors are not uniform across H01). The public c3 map is the version analysts use. Manual proofreading of the hundreds of thousands of spines on the 104 proofread neurons was not done (main text p. 6), so per-neuron input counts from the automatic map miss inputs on detached spines.
3 → 7dense bands of neurons, found by a clustering algorithm from cell positions and cell-body sizes alone, gave the boundaries that split this human cortex sample into 7 regions: layers 1 to 6 and white matter
How: Paper (supplementary materials pp. 26–27, "Identification of cortical layers"): layers found "without any prior assumptions about the numbers of these layers, or their locations"; HDBSCAN clustered neurons by x–y distance and soma-volume difference and "identified three clusters"; concentric circular arcs fitted to the top and bottom edges of each cluster gave six boundaries that "divided the cortex into seven regions, corresponding to classical descriptions of layers 1 through 6 and white matter". Neurons per layer label in somas.csv (ours): L1 827, L2 4,656, L3 2,692, L4 3,440, L5 2,313, L6 1,077, white matter 648, none 434 (of 16,087).
Caveat: One sample (human temporal cortex, epilepsy surgery). The boundaries are smooth arcs fitted to clusters, so cells near a boundary may sit in the neighbouring layer by classical criteria. The paper's code is in its Zenodo analysis-scripts repository.
96.5%of axon-to-neuron connections in this human cortex sample are a single synapse; yet rare axons make 7 to about 50 synapses on one neuron, more often than chance predicts
How: Paper values (author manuscript, main text p. 12): synapses per axonal input, 1 = 96.49%, 2 = 2.99%, 3 = 0.35%, 4 or more = 0.092% (Table S8); 39% of the 2,743 neurons with at least 3,000 synapses had at least one input with 7 or more synapses. Structured abstract (p. 2): inputs of up to ~50 synapses. Main text p. 13: a null model of chance contacts is inconsistent with the observed strong pairs (p < 10^-10, 79,827,631 axons).
Caveat: One sample (human temporal cortex, epilepsy surgery). Based on the automated c3 segmentation, in which many axons are split into pieces, so per-axon counts may be too low. Chandelier axons on axon initial segments were excluded. The paper shows the strong pairs are not chance but does not say what they are for.
~27,000synapses for every astrocyte in this human cortex sample, on average (about 150 million synapses, 5,474 astrocytes)
How: Paper abstract: ~150 million synapses. Table S5: 5,474 astrocytes (counted by hand). 150,000,000 / 5,474 ≈ 27,400 (our division).
Caveat: An average from two sample totals, not a measured count of the synapses each astrocyte touches. The synapse total comes from automated detection. One sample (human temporal cortex, epilepsy surgery).
2.4 : 1neurons to glia in layer 2 of this human cortex sample, the one layer where neurons outnumber glia (overall, glia win about 2 to 1)
How: Paper, Table S6 (cells per mm³, compression-corrected). Layer 2: neurons (spiny 21,467 + interneurons 11,655 + unclassified 623) = 33,745; glia (astrocytes 3,592 + oligodendrocytes 3,872 + microglia/OPC 6,490) = 13,954; ratio ours. White matter: glia 65,552 vs neurons 4,174, about 16 to 1. Overall: Table S5, 32,315 glia vs 16,087 neurons. Recounting somas.csv gives 0.42 glia per neuron in layer 2 and 18.8 in white matter.
Caveat: One sample (human temporal cortex, epilepsy surgery). The paper's layers are data-driven clusters. The paper links myelin to oligodendrocyte density (both lowest in layer 2); it does not test why layer 2 has fewer glia.
5,019slices, about 34 nanometres thick on average, were cut from this speck of human cortex and photographed; the imaging took 326 days
How: Supplementary materials p. 2: "In total, 5019 sections were cut, ranging in thickness between 30 and 80nm, with 92.5% between 30 and 33nm"; average thickness as cut 33.9 nm (Table S10, p. 58). Main text (author manuscript p. 6): total imaging time 326 days; 247 M tiles (1.8 PB) and a unified ~1.4 PB aligned image volume.
Caveat: Some sections were cut double-thick and are shown twice in the online viewer, which therefore has 5,293 digital sections. One sample (a 170 µm thick slab of human temporal cortex, epilepsy surgery).
1.2%of the neurons in this human temporal cortex sample are spiny stellate ("star") cells: 193 of 16,087
How: Paper, Table S5 (manual counts of all cells): spiny stellate 193 of 16,087 neurons = 1.2% (our division). The public somas.csv gives the same counts. Cell types were assigned by hand from EM appearance and 3D shape (supplementary methods).
Caveat: One sample, from the anterior middle temporal gyrus (association cortex, epilepsy surgery). Spiny stellate cells are typical of layer 4 in primary sensory areas, so a low count here is expected; it says nothing about visual or touch cortex. The paper's layers are data-driven, so we do not report where in the layers these cells sit.
1 in 7cells in this human cortex sample belongs to the blood vessels (about 8,000 of 57,000, counted by hand)
How: Paper, Table S5 (manual counts of all cells): all blood-vessel-related cells 8,136 of 57,216 cells = 14.2% (our division). Sub-rows: endothelial cells 4,604, pericytes 3,549, circulating immune cells 46, blood-vessel cells in the main cell list 285.
Caveat: The sub-rows do not add up exactly to 8,136, and the main text gives about 8,100 of 57,180 cells, so we round to about 8,000 and 1 in 7. One sample (human temporal cortex, epilepsy surgery).
55%of the neurons in this human cortex sample are pyramidal cells (8,803 of 16,087), each classified by people
How: Paper, Table S5 (manual counts of all cells): pyramidal 8,803; other excitatory (spiny 1,535, spiny stellate 193); interneurons 4,688; unclassified neurons 868; all neurons 16,087. 8,803 / 16,087 = 54.7% (our division). The public somas.csv gives the same counts. Cell types were assigned by hand from the EM images and 3D shapes (supplementary methods).
Caveat: One sample (human temporal cortex, epilepsy surgery). "Extrapyramidal" is not a cell type; it names motor pathways outside the pyramidal (corticospinal) tract.
~39%of synapses in this human cortex sample are inhibitory ("stop") once the automated count is corrected; the raw count said 26%
How: Supplementary methods: the automated pipeline classified 111,272,315 synapses as excitatory and 38,599,354 as inhibitory (inhibitory 25.75% of the raw count; our calculation). People annotated 5,554 true synapses on 42 axons (25 excitatory, 17 inhibitory; Table S3): the detector missed 11% of excitatory and 35% of inhibitory synapses, with false discovery rates of 3.2% and 2.7%. Corrected for missed, false and misclassified synapses, the main text estimates 111.6 M excitatory (61.2%) and 70.7 M inhibitory (38.8%) synapses.
Caveat: The 38.8% is the paper's estimate; we could not reproduce it exactly from the published formula and global counts (it was probably applied per 10 µm cube), and a global plug-in gives about 37%. Error rates come from 42 axons. One sample (human temporal cortex, epilepsy surgery).
~34,000/mm³neurons in layer 2, the most crowded layer of this human cortex sample; layer 4 is next (~29,000), layer 1 has ~5,000
How: Paper Table S6 (cells per mm³ by layer, corrected for 28% compression). Neurons = spiny + interneurons + unclassified: L1 5,272 · L2 33,745 · L3 14,314 · L4 29,463 · L5 14,118 · L6 11,456 · white matter 4,174. Layers II and IV are the classical external and internal granular layers.
Caveat: One sample (human temporal cortex, epilepsy surgery). Layer borders are the H01 team's, drawn from cell positions and sizes. Densities differ between cortical areas.
~73%of the mapped tissue in every cortical layer is wire (axons and dendrites); cell bodies take up about 10%
How: Paper Table S4 (volume occupancy by layer, % of segmented tissue). Wire = axon + dendrite + myelinated axon: 74.1 / 73.1 / 74.3 / 70.4 / 72.7 / 73.1% in layers 1–6. Cell bodies: 2.1 / 10.6 / 9.6 / 12.9 / 11.8 / 11.1%. Axons alone 40.0–44.4%, dendrites 23.1–32.3%. Sums are ours.
Caveat: Shares are of the tissue the paper could classify: 36% of voxels, including myelin sheaths, were not segmented, so these are not shares of every cubic micrometre. One sample (human temporal cortex, epilepsy surgery). The ~60% optimum is a theory (Chklovskii et al. 2002), not measured in human tissue.
~6,000/mm³microglia and OPCs in every cortical layer — the one cell class spread evenly, varying only about 10% from layer 1 to layer 6
How: Paper Table S6 (cells per mm³ by layer, corrected for 28% compression): microglia/OPC 6,329 / 6,490 / 6,519 / 5,922 / 5,854 / 6,004 in layers 1–6 (max/min 1.11). Neurons vary 6.4-fold across the same layers (5,272 to 33,745), oligodendrocytes 5.1-fold, astrocytes 3.4-fold.
Caveat: The dataset counts microglia and OPCs as one class. The reason given (cells keep their distance; a neighbour moves in when one is lost) was shown for OPCs in adult mice (Hughes et al. 2013), not for microglia or for human tissue. One sample.
31.5%of the mapped white-matter volume is the inside of myelinated axons (sheaths not counted) — versus only 1.0% in layer 1, the cortex's top layer
How: Paper Table S4 (volume occupancy by layer): myelinated axon 1.0 / 0.8 / 2.5 / 4.8 / 5.6 / 7.7 / 31.5% in layers 1–6 and white matter. The video pairs two raw electron-microscope tiles (layer 1 and white matter) from the same sample.
Caveat: One sample (human temporal cortex, epilepsy surgery). The paper's class is myelinated axons excluding their myelin sheaths, and the shares are of the segmented tissue: 36% of voxels (myelin sheaths among them) were not segmented. So 31.5% is not the share of every cubic micrometre of white matter. These are the paper's values, not our own count.
31%of the neurons in this human cortex sample are inhibitory — almost 1 in 3, versus about 12% in a mouse cortex map
How: Neurons by the dataset's own cell-type labels (somas.csv): 4,688 interneurons of 15,219 classified neurons (pyramidal 8,803, spiny atypical 1,535, spiny stellate 193, interneuron 4,688). Mouse: MICrONS cell table (mouse visual cortex), 9,017 inhibitory of 75,405 classified neurons = 12.0%. Loomba et al. 2022 (Science) report a 2.5-fold larger interneuron pool in human than mouse cortex. Script: github.com/shoo99/kbrain-map-numbers/blob/main/scripts/inhibitory-third.py
Caveat: One human sample (temporal cortex, epilepsy surgery). 868 unclassified neurons left out (with them: 29.1%). The share varies by layer (L2–3 about 35–37%, L6 12%). Human and mouse maps cover different brain areas and were classified by different methods.
37%of the cells in layer 1 are astrocytes — the most common cell in the cortex's top layer
How: Cells by the dataset's layer labels (somas.csv): layer 1 — 1,688 astrocytes of 4,595 cells; 1,688 of the sample's 5,474 astrocytes (30.8%) sit in layer 1. Consistent with the paper's Table S4 (astrocyte volume fraction 21.6% in layer 1 vs 14–15% in layers 2–6). Script: github.com/shoo99/kbrain-map-numbers/blob/main/scripts/layer1-astrocytes.py
Caveat: One sample; layer labels are the dataset's own. The map does not tell astrocyte subtypes apart (e.g. primate-specific interlaminar astrocytes, Oberheim et al. 2009).
18% vs 70%of the cells are neurons in layer 1, the cortex's top layer, versus layer 2 just below
How: Cells by the dataset's layer labels (somas.csv): layer 1 — 827 neurons of 4,595 cells (586 interneurons, 72 pyramidal); layer 2 — 4,656 of 6,651. Script: github.com/shoo99/kbrain-map-numbers/blob/main/scripts/layer1-few-neurons.py
Caveat: Layer labels are the dataset's own; one sample. Layer 1 is mostly wiring: pyramidal-cell branch tips meeting long-range inputs (Schuman et al. 2021).
20,139 vs 16,087oligodendrocytes outnumber neurons in this human cortex sample
How: Cell types in H01 somas.csv; matches the paper's Table S5 exactly.
Caveat: Driven by the white matter in the sample: in the grey-matter layers alone neurons outnumber oligodendrocytes (15,005 vs 9,350).
77%of the cells in the sample's white matter are oligodendrocytes
How: 10,110 of 13,063 white-matter cells in somas.csv (half of all oligodendrocytes in the sample sit there). Script: github.com/shoo99/kbrain-map-numbers/blob/main/scripts/white-matter-oligodendrocytes.py
Caveat: Layer labels are the dataset's own.
98%of neurons sit within 50 µm of a blood vessel
How: 10,441 neuron cell-body centres (somas.csv, away from the sample edges) to the nearest vertex of the finest blood-vessel meshes (1,477 segments). Median 22.9 µm; max 141 µm.
Caveat: Distance to mesh vertices (spacing ~1.6 µm); the vessel segmentation may miss some capillaries, so true distances may be shorter. One ~1 mm³ sample of human temporal cortex.
245 µm²of a glial cell's surface pressed against a neuron's cell body
How: Finest meshes of the gallery's 'parasitic_glial' cells: glial surface within 0.25 µm of the neuron (8% of the glial surface; ~5% of the neuron's cell-body surface).
Caveat: One pair of cells; the glial cell is labelled microglia/OPC.
≈7 µmacross: a flat, whorled axon coiled on a neuron's cell body
How: Finest public meshes of the gallery's 'axon_whorl' (97 segments); disc ≈6 × 7 µm, ~1.6 µm thick; checked against the raw electron-microscopy images.
Caveat: Automated segmentation; which pieces form the axon was chosen by the H01 team's gallery.
almost allinputs within 20 µm of one pyramidal neuron's cell body are inhibitory
How: All input synapses of cell 6644185383 in both public synapse releases; within 20 µm of the soma centre: 97.7% (20210601) and 96.5% (20210729, the release that matches the paper).
Caveat: One cell. Labels partly depend on whether a synapse sits on a spine.
~150 millionsynapses in just over one cubic millimetre of human cortex
How: Paper values. Structured abstract (author manuscript p. 2): ~150 million synapses. Supplementary materials p. 26: the trained model classified 111,272,315 synapses as excitatory and 38,599,354 as inhibitory, 149,871,669 in total (sum ours).
Caveat: Automated detection; the paper reports missed and false detections.

Stories that use this map