CA3 renderings

Hippocampal CA3 connectomics reveals a gradient of mossy fiber inputs and selective feedforward inhibition onto pyramidal cells Zheng, Park, Hammerschmith, Lu, Yu, Sorek, Silverman, Jordan, Sterling, Silversmith, Schlegel, Jefferis, Collman, Seung, Tank · Nature Neuroscience, 5 August 2026 · doi:10.1038/s41593-026-02388-9
The journal version sits behind a paywall, so the open access preprint is worth keeping to hand. The dataset is shared through Pyr, an online platform for hippocampal connectomics.

1,815

pyramidal cells

229

inhibitory cells

55,000+

mossy fibers

25,200

mossy fiber inputs mapped

1 × 1 × 0.1 mm

of tissue, imaged whole

Wireframe diagram of the reconstructed tissue block with one pyramidal cell inside it for scale — The block that was cut, imaged and reconstructed: **1 x 1 x 0.1 mm** of mouse CA3. One pyramidal cell is drawn inside it at true relative scale. Nearly two thousand of them are in there, threaded through by more than fifty five thousand mossy fibers. Turn it yourself: the same block, plate and cell, in 3D.

Overnight renders

Rendered on one workstation in a single night, 103 minutes of GPU time for the four pieces below. The two widescreen masters also replace their vertical cuts further down this page when you are on a wide screen.

304 CA3 pyramidal cells arranged along the curve of the cell layer, coloured from deep blue through violet to gold by how many mossy fiber inputs each one receives — **The convergence gradient.** 304 pyramidal cells coloured by how many mossy fibers reach them, deep blue for the fewest and gold for the most, on a log scale because the count runs from 0 to 235 with a median of 41. Every cell here sits in the deepest part of the block, where the reconstruction is most complete. Read along the curve of the cell layer and the gradient is **not** a simple ramp: both ends carry heavily innervated cells and the middle carries the fewest. In this sample the median runs about 50 at one end, dips to 33 mid curve, then climbs to 72 at the far end, where the busiest cells reach 160 and above.

Selective feedforward inhibition. Eighteen seconds. Seven mossy fibers converge on one interneuron, in emerald, and its axon then makes 333 synapses onto 87 thorny pyramidal cells against only 50 onto 21 sparsely thorny ones. Normalised by how many cells of each subtype exist here, that is 2.49 times more inhibition landing on each thorny cell. The wave travels the interneuron's own cable, and each target lights at its real path distance, from 216 to 734 µm out.

The scale ladder. Twenty seven seconds from the whole block to a single synapse, holding on each rung long enough to read it. The tissue fills with 1,064 cells, the view closes on one pyramidal neuron, then on the one thorny excrescence where a mossy fiber wraps around it and makes 53 synapses inside a box under seven micrometres on a side, and the spike arrives. Registration was checked independently: this cell's soma lands on the stratum pyramidale boundary in the real EM, with those boutons 39 um into the mossy fiber band.

The gradient, measured as it goes. A band of light travels the cell layer and the readout recomputes the median mossy fiber input count for the cells inside it, from the same measurements that set the colours, so the number and the cells lit under it cannot disagree. Watch the figure rather than the caption: it starts near 51, falls to 33 through the middle, and climbs past 64 at the far limb, where the busiest cells carry more than 170. The count of cells in the band is shown throughout, because the sample thins at the ends and a median over twenty cells deserves less weight than one over eighty.

Why the variation matters

The textbook picture of this connection is a detonator. Each CA3 pyramidal cell is supposed to hear from only a handful of granule cells, but so powerfully that one of them can nearly fire the cell by itself. Few inputs, each enormous, is exactly the arrangement you would design if you wanted two similar experiences to end up recruiting two different sets of CA3 cells — which is what pattern separation means, and why this synapse is thought to matter for laying down a memory that does not smear into yesterday's.

That picture assumes every CA3 cell is roughly alike. This gradient says otherwise. How much dentate input a cell receives depends substantially on where it sits along the layer, and the spread is not subtle: the busiest cells here carry more than five times the median of the quietest stretch. A cell's position is part of its job description, so CA3 is less a uniform pool of interchangeable units than the classical model treats it as.

The most interesting part is the twist, and it cuts against the simple reading. The distal cells that collect the most inputs collect them through smaller terminals holding fewer vesicles. Count and per-contact strength run in opposite directions. So the gradient in the numbers above is not a gradient in raw drive, and reading it as one would overstate it — a cell with three times the contacts is not being shouted at three times as loudly. It looks more like the same total budget being spent two different ways: a few loud inputs at one end of the layer, many quiet ones at the other.

The wiring also turns out to be spatial. Cells here share far more mossy fiber inputs than chance allows, but that sharing is explained better by which cells sit near each other than by how many inputs each one has. Neighbours sample the same passing fibers. That is in real tension with pattern separation, which works best when nearby cells hear different things, and it is the kind of constraint you only see once you can trace every fiber rather than sample a few.

What this cannot tell you. Terminal size and vesicle count are anatomical proxies for synaptic strength, not recordings of it; nobody measured a current here. The gradient is traced across roughly 1.1 mm of one block from one animal, which is a slice of the proximodistal axis rather than the whole of it, and one animal cannot tell you what varies between animals. And the sample thins toward both ends, which is why the cell count sits in the readout the whole way along.

The population, widescreen. Twenty seconds. Populations arrive in circuit order: mossy fibers first, then the cells they drive, then the wider pyramidal population, and inhibitory interneurons last.

The synapse story, widescreen. Nineteen seconds. The same sequence as further down the page, rendered for a wide screen rather than cropped to one.

These two are tall subjects in a wide frame, so they carry black at the sides. The population is deeper than it is broad and nothing is cropped to hide that.

Watch it assemble

Populations arrive in circuit order: mossy fibers first as the incoming signal, then the CA3 cells they drive, then the wider pyramidal population, and inhibitory interneurons last.

What the study found

The hippocampus is where experience becomes memory, and CA3 is the part that binds the pieces of an event together. To understand how, you need to know not just which cells are present but exactly which ones talk to each other.

This study took a block of mouse CA3, imaged it with an electron microscope, and traced every neuron and every connection inside it. The result is a complete wiring diagram of a small piece of real brain: 1,815 pyramidal cells, the main excitatory neurons, 229 inhibitory cells, and more than 55,000 mossy fibers, the axons carrying signals in from the neighbouring dentate gyrus.

Not all pyramidal cells are alike

Some are covered in elaborate spiny structures called thorny excrescences, and these cells receive many mossy fiber inputs. Others, the sparsely thorny cells, receive almost none. The split is sharp rather than gradual, so these are genuinely two kinds of cell rather than two ends of one continuum.

Pyramidal cell bodies covered in thorny excrescences — Thorny excrescences on pyramidal cell bodies. Each balloon like cluster is a single elaborate spine, and a mossy fiber terminal wraps around it.

Input is unevenly spread across space

Cells further along the region receive substantially more mossy fiber contacts than cells nearer the beginning. Cells sharing the same incoming fibers also sit closer together than chance would predict, suggesting the wiring is organised rather than arbitrary. Curiously, the cells receiving the most inputs get them through smaller terminals holding fewer vesicles, so more connections does not simply mean more of everything.

Inhibition is targeted, not general

Mossy fibers also drive inhibitory neurons, which act as brakes on the circuit. But they drive only the inhibitory neurons that go on to target thorny cells. Those serving sparsely thorny cells receive no mossy fiber input at all. So the same signal that excites one type of cell also recruits the brake belonging specifically to that type, leaving the other type untouched.

182 thorny pyramidal cells

Inside a single cell

Thorny excrescences are the elaborate spines that give thorny pyramidal cells their name. A mossy fiber ends on one in a single enormous terminal, making dozens of contacts at one site rather than one contact in many places. This cell receives 165 mossy fiber synapses from just 6 fibers.

One thorny pyramidal cell in blue with the six mossy fibers that contact it in gold

this one cell receives

165

mossy fiber synapses, and they arrive from only 6 fibers. A single terminal makes dozens of contacts on one thorn.

reconstructed at

4.5M

triangles, with no simplification applied, which is why the thorny excrescences hold their shape this close up.

in the full volume

1,815

pyramidal cells, 229 inhibitory cells and more than 55,000 mossy fibers were reconstructed.

the dividing line

mossy fiber inputs separates thorny cells from sparsely thorny ones. Thorny cells reach 165. Sparsely thorny cells stop at 3.

Eight seconds, thorns to whole cell. The mossy fibers arrive nearest first, so the one already in frame is the one that appears.

Watch it fire

Six mossy fibers reach this cell, and this is the honest version of what they do. One fires, lands hard on the thorn, and the cell does not answer. The other five arrive. All six fire together, and only then does the cell send a signal of its own. The pulse is not a sweep across the screen: every point on these meshes carries its own distance measured along the cell's own skeleton, so the wavefront follows the real branching and splits wherever the cable splits.

Eighteen seconds. All 165 mossy fiber synapses onto this one cell, from all six fibers that reach it.

A single mossy fiber makes one of the strongest connections in the brain, touching a CA3 pyramidal cell at dozens of release sites at once. Even so, one incoming spike is usually not enough to make the cell fire. When the granule cell fires a rapid burst of a few spikes, the connection strengthens within milliseconds and then reliably triggers a spike. That is why these are called conditional detonator synapses.

Two limits worth stating. The reconstruction is cut off at the edge of the tissue block, so this arbor is incomplete. And the descending branches carrying the signal are mostly basal dendrite with the axon embedded in them; at this resolution the two could not be cleanly separated. Three other cuts of the same event are here.

Every synapse these cells make

Sixteen CA3 pyramidal cells were traced. Fifteen of them have their axon followed through the volume, and those axons make 25,723 synapses onto 6,348 other cells in the same block of tissue. Every one sits at its real measured coordinate, not an estimated position.

That works out to a median of 1,949 outgoing synapses per cell, which sounds impossibly high until you remember what CA3 is for. Its pyramidal cells are the most heavily interconnected population in the hippocampus, wiring back onto each other through vast recurrent axon collaterals, and this is the network that theory says lets the hippocampus complete a whole memory from a fragment of one. The gold beads string along each axon because that is physically where the contacts are made, one after another as the axon passes its targets. Most partners are touched just once: 58% receive a single synapse.

The presynaptic cells fade up, their synapses bloom across the volume, and then the postsynaptic partners arrive around them.

A sea of mossy fibers

Seen edge on, the pyramidal cell bodies hang in a single layer while the mossy fibers thread horizontally through them. Every gold strand is an axon arriving from the dentate gyrus.

Pyramidal cell bodies in blue with mossy fibers threading through in gold

The classification, reproduced

The study separates thorny from sparsely thorny cells by mossy fiber convergence, drawing the line at more than 10 inputs. That split reproduces cleanly in this subset.

population	with MF input	median inputs	above 10
Thorny pyramidal	94 / 182	22	71%
Sparsely thorny	8 / 68	1	0%
Inhibitory	9 / 28	2	22%

Thorny cells reach 165 mossy fiber inputs at the top end. Sparsely thorny cells max out at 3, and not one crosses the threshold. Only 9 of 28 interneurons receive mossy fiber input at all, which is the selectivity the feedforward inhibition result describes.

Gallery

Stills from the same reconstruction, at full size, and one slow pass across the convergence gradient. Tap any still to open the large version.

A slow pass across the gradient. Twelve seconds, a sixth of a turn, eased at both ends so the arch is read rather than spun. Colour is how many mossy fibers reach each cell, the same measurement as the still further up. What the movement adds is depth: these are 304 cells occupying a curved sheet of tissue, not points on a scatter plot.

The CA3 population as a broad V shaped band with a circular inset on one mossy fiber bouton — The reconstructed population seen face on, gold mossy fibers threading through crowds of cells. The inset follows one mossy fiber axon down to its boutons, with a partner cell process pressed against it.

A row of CA3 pyramidal cell bodies, each in a different colour — Pyramidal cell bodies lined up as they sit in the layer, one colour each. The apical dendrites climb out of the top carrying their thorny excrescences, and the basal dendrites fan out below.

These six mossy fibers are not private to this cell. Each carries on through the tissue and lands on 56 further segments, another 126 synapses. Fourteen of those are complete cells with a nucleus in the volume, split evenly between the two pyramidal subtypes; the rest are traced fragments. The purple is not senders, it is fellow listeners hearing the same signal at the same moment.