How neurons communicate at synapses

In BCI-1.2 you watched a spike, an action potential, race down the axon as a self-renewing wave of voltage. Now it reaches the end of the line, the axon terminal, and hits a problem the whole nervous system is built to solve: the next neuron is not attached. There is a gap. The electrical signal you just spent a lesson deriving cannot simply keep going, because a voltage wave rides on a membrane and across that gap there is no membrane to ride on. This lesson is about the trick evolution uses to jump the gap, and about the quiet consequence of that trick, which is that a single neuron ends up summing thousands of tiny votes and firing only when they add up. That summing is exactly the picture that makes people reach for the artificial-neuron analogy, so we will build the real thing first and then say plainly where the analogy breaks.

The gap a spike cannot jump

Two neurons that talk to each other do not fuse. They come close and stop. The connection is called a synapse, and at a chemical synapse the two cells are separated by a gap of roughly 20 nanometers called the synaptic cleft. The neuron sending the signal is the presynaptic cell, the one receiving it is the postsynaptic cell, and between them is that sliver of salty fluid.

Here is why the spike cannot cross it on its own. Recall from BCI-1.2 that an action potential is a moving disturbance in the membrane's voltage, ions rushing through channels in one patch and triggering the next patch to do the same. It is a wave that needs a surface to propagate along. The cleft has no such surface. So the presynaptic terminal does something clever: it converts the electrical signal into a chemical one, squirts a chemical messenger across the tiny gap, and lets the postsynaptic cell turn that chemical back into an electrical signal on its own membrane. Electrical, then chemical, then electrical again. The gap is jumped by a molecule, not a voltage.

From voltage to calcium: the arriving spike pulls a trigger

The conversion starts the instant the spike's depolarization reaches the terminal. Sitting in the terminal membrane are voltage-gated calcium channels, the same family of voltage-sensing gates you met in BCI-1.2 for sodium, but tuned to pass calcium ions, Ca2+. They are shut at rest. The arriving depolarization is their trigger: the membrane swings positive, the gates snap open, and calcium flows in.

Why does calcium flow in and not out? For the same reason sodium rushes in during a spike. Calcium is concentrated outside the cell and kept extremely scarce inside, a gradient the cell spends energy to maintain, so the instant a channel opens, Ca2+ pours down its electrochemical gradient into the terminal. The gradient is the stored charge (S9.1), the channel is just the tap.

Calcium is a deliberate choice of messenger, not an accident. The resting neuron holds its internal calcium astonishingly low, roughly ten-thousand-fold below the outside. Against a near-zero background, even a small influx is an enormous fractional jump, which makes calcium a clean, high-contrast internal signal that the cell can read as "a spike just arrived here." That is a recurring trick in biology: keep a messenger rare so that a little of it means a lot.

Calcium dumps the neurotransmitter

Waiting inside the terminal are hundreds of synaptic vesicles, a dozen or so of them docked right at the membrane and primed, tiny membrane-wrapped bubbles each pre-loaded with thousands of molecules of a chemical messenger called a neurotransmitter. The calcium that just flooded in binds a calcium-sensor protein on those docked vesicles, and that binding is the "go" signal for the vesicle to fuse with the terminal membrane and spill its contents into the cleft. The fusion-and-release step is called exocytosis.

Two features matter for a BCI person. First, release is fast, the whole voltage-to-calcium-to-release chain takes well under a millisecond, which is why a synapse can keep up with a train of spikes. Second, release is quantal, meaning transmitter leaves in vesicle-sized packets rather than a smooth stream, so the signal crossing the cleft is inherently a little grainy and probabilistic. A given spike does not always release the same amount, or sometimes any at all. Real synapses are noisy channels, and any honest decoding story downstream inherits that noise.

Once dumped into the cleft, the neurotransmitter simply diffuses across. Diffusion sounds slow, but over 20 nanometers it is nearly instant, because diffusion time grows with the square of distance and this distance is minuscule. The messenger reaches the far wall in microseconds and binds the receptors waiting there.

Reading the message: two kinds of receptor

On the postsynaptic membrane, usually on a branch of the receiving neuron's dendrites, sit receptors shaped to catch that specific neurotransmitter. They come in two broad kinds, and the difference is about speed and mechanism.

An ionotropic receptor is a receptor and an ion channel fused into one protein, a ligand-gated channel. When the neurotransmitter binds, the channel changes shape and opens directly, and ions flow. This is fast, on the order of a millisecond, because the message and the pore are the same molecule. There is nothing in between.

A metabotropic receptor is a pure sensor with no pore of its own. When the neurotransmitter binds, the receptor does not open anything directly. Instead it kicks off an internal signaling cascade inside the cell, the exact machinery S9.3 covers, which then goes on to open or close channels elsewhere, adjust the cell's excitability, or change what genes are expressed. This is slower, tens of milliseconds to seconds and beyond, and its effect is more like turning a knob than flipping a switch. Hold onto the metabotropic route, because it is how neuromodulators reshape a whole circuit's behavior, and it is one of the reasons the tidy analogy at the end of this lesson leaks.

Excitatory versus inhibitory: the sign lives in the channel

Now the payoff. When a receptor opens a channel, ions flow, and that flow nudges the postsynaptic membrane voltage. Which way it nudges depends entirely on which ions the channel passes.

Take glutamate, the brain's main excitatory transmitter. It binds ionotropic receptors whose channel passes positive ions, mainly sodium, into the cell. Positive charge entering makes the inside less negative, a small depolarization toward the firing threshold. That little upward bump is an excitatory postsynaptic potential, an EPSP. It does not fire the neuron. A single EPSP is tiny, often under a millivolt, while the cell sits around -70 millivolts and needs to climb to roughly -55 millivolts to fire. One EPSP is one small vote for "fire."

Now take GABA, the main inhibitory transmitter. It binds ionotropic receptors whose channel passes chloride, Cl-, which flows in and makes the inside more negative, or holds it clamped near rest. Either way the membrane moves away from threshold, or resists moving toward it. That is an inhibitory postsynaptic potential, an IPSP, a vote for "do not fire."

Notice the load-bearing detail: excitatory versus inhibitory is not a property of the transmitter molecule by itself. It is set by the receptor and the ion its channel passes. The very same neurotransmitter can be excitatory at one receptor and inhibitory at another, depending on what that receptor opens. The sign of the signal lives in the lock, not the key.

Integration: one neuron sums thousands of votes

A real neuron is not listening to one synapse. Its dendrites are covered in thousands to tens of thousands of them, some excitatory, some inhibitory, each firing its own little EPSP or IPSP at its own moment. The neuron's job is to add all of that up and make one decision: spike, or stay quiet.

The adding happens in two directions at once. Spatial summation is many synapses across the dendrites contributing at the same time, their voltage bumps overlapping. Temporal summation is one or a few synapses firing in quick succession, before each bump has faded, so the bumps stack. And they do fade, because the dendritic membrane is leaky, that same capacitor-with-leak from S9.1, so a postsynaptic potential shrinks as it spreads toward the cell body and dies away within about ten milliseconds if nothing reinforces it. Distant synapses are quieter votes, recent and clustered ones are louder.

All of these summed, decaying signals converge on one spot: the axon hillock and initial segment, the base of the axon, which carry the highest density of voltage-gated sodium channels in the cell. This is the decision point. If the net depolarization arriving there crosses threshold, those sodium channels ignite and the neuron fires exactly one all-or-none spike, right back to the mechanism of BCI-1.2. If the net falls short, nothing happens. The neuron has quietly weighed thousands of weighted inputs and collapsed them into a single binary output.

That is the behavior worth naming: a neuron acts like a threshold unit that sums weighted inputs and fires when the sum clears a bar. The "weight" of any one synapse is how strongly it moves the postsynaptic voltage, which depends on how much transmitter is released, how many receptors are waiting, and where on the dendrite the synapse sits. How many receptors sit in that membrane is itself set by gene expression (S8), and changing that number is one concrete way a synapse's weight goes up or down over time, which is the whole subject of the next lesson.

The artificial-neuron analogy, and exactly where it breaks

If you have built a neural network, the picture above is uncanny. Inputs, each scaled by a weight, summed together, then pushed through a threshold or activation function to produce an output. That is the McCulloch-Pitts neuron and the perceptron, the founding cartoon of artificial neural nets, and it was drawn on purpose as a caricature of the biology you just derived. The resemblance is real and worth savoring: weighted sum, threshold, fire.

Now the failure edge, because an analogy without its limit is a bug, and this one is load-bearing enough that overselling it produces bad intuitions about what a brain is.

The weights are not trained by backpropagation. There is no global error signal flowing backward through the brain adjusting each synapse to reduce a loss. Biological weights change by local, activity-dependent rules, roughly "neurons that fire together wire together," which is a different learning algorithm with different reach, and it is what BCI-2.2 is about.

The dendrites are not a passive summing wire. They do local nonlinear computation, with regions that can generate their own small spikes and combine their inputs in ways a single linear sum cannot capture. For this reason a single biological neuron is better modeled as a small multilayer network than as one artificial unit. The one-neuron-equals-one-node mapping quietly undercounts by a large factor.

Timing carries information, not just totals. Whether two spikes arrive together or a few milliseconds apart can flip a synapse from strengthening to weakening, and part of the code the brain uses is spikes placed in time, discrete events, not only the continuous real numbers that flow through a standard artificial layer.

And the whole network is chemically reconfigurable on the fly. The metabotropic receptors and neuromodulators from earlier can turn the gain up or down across thousands of synapses at once, so the same wiring computes different functions in different states. A trained artificial net has fixed weights at inference time. A brain does not sit still.

Keep the analogy for what it is good for, the intuition that a neuron weighs inputs and thresholds them, and drop it the moment you start reasoning about how brains learn, how much one neuron computes, or what the signal actually is. The map is useful. It is not the territory.

Key terms

synapse

The junction where one neuron signals another, separated at a chemical synapse by a roughly 20-nanometer gap called the synaptic cleft.

voltage-gated calcium channel

A channel in the axon terminal that opens when the arriving spike depolarizes the membrane, letting Ca2+ flood in down its gradient to trigger transmitter release.

synaptic vesicle

A membrane bubble pre-loaded with neurotransmitter that fuses with the terminal membrane when calcium arrives, dumping its contents into the cleft (exocytosis).

neurotransmitter

The chemical messenger released into the cleft that diffuses across and binds receptors on the postsynaptic cell, carrying the signal that a spike could not.

ionotropic vs metabotropic receptor

Ionotropic receptors are ligand-gated ion channels that open directly and act in about a millisecond. Metabotropic receptors are sensors that act slowly through internal signaling (S9.3).

EPSP and IPSP

Small graded voltage changes in the postsynaptic cell. An EPSP depolarizes toward threshold (excitatory), an IPSP hyperpolarizes or clamps away from it (inhibitory).

spatial and temporal summation

How a neuron adds many small postsynaptic potentials, across different synapses at once (spatial) and from repeated firing before decay (temporal), at the axon hillock.

threshold unit

A model of the neuron as a device that sums weighted inputs and fires one all-or-none spike only if the net depolarization at the axon hillock crosses threshold.

Why release is probabilistic, and why that matters for a decoder

A single presynaptic spike does not reliably release a fixed dose of transmitter. Whether any given docked vesicle fuses is probabilistic, governed by how much calcium got in and how primed the vesicles were, so the same spike can yield a strong response, a weak one, or a failure. This is real biology, not a measurement artifact. It means the "weight" of a synapse is a statistical average, not a crisp constant, and that a burst of identical spikes produces a distribution of postsynaptic effects rather than a repeated identical one. For a brain-computer interface this is one of several reasons the neural signal is noisy at the source, before any electrode or amplifier adds its own noise. A decoder is never fitting a clean deterministic function. It is estimating a signal that is genuinely stochastic in the tissue, which is part of why the honest performance ceiling for reading neural activity sits where it does, a theme the recording and decoding lessons return to.