Long-Term Depression and Spike-Timing Plasticity

Long-term depression (LTD) is an activity-dependent, lasting decrease in the efficacy of a chemical synapse, expressed most commonly as a reduction in postsynaptic AMPA-receptor number or conductance, though presynaptic contributions exist in some pathways. Like LTP, LTD is input-specific — only the synapses that experienced the depressing pattern of activity are weakened, not neighboring synapses on the same cell — and it can last from hours in a slice preparation to, by inference, much longer in vivo. LTD is not a passive "decay" of unused connections; it is an actively triggered biochemical cascade that requires calcium influx, specific phosphatases, and receptor trafficking machinery, just as LTP requires calcium influx and specific kinases. The deepest conceptual advance uniting LTP and LTD is the recognition that both are outputs of the same postsynaptic calcium-detection system, read out differently depending on the dynamics of the calcium signal. This is often called the calcium-control hypothesis (developed by Lisman and others): brief, large-amplitude calcium transients — of the sort produced when a strongly depolarized postsynaptic spine unblocks its NMDA receptors during high-frequency presynaptic firing — preferentially activate calcium/calmodulin-dependent protein kinase II (CaMKII) and drive potentiation. Modest, more prolonged, lower-amplitude calcium transients — produced by low-frequency stimulation with only partial NMDA receptor unblocking — preferentially activate the calcium-dependent phosphatase calcineurin (protein phosphatase 2B) and its downstream partner protein phosphatase 1 (PP1), driving depression. One receptor family, one second messenger, two opposite outcomes, sorted by the shape of the calcium signal. Spike-timing-dependent plasticity refines this further by adding a temporal-order dimension that a simple "coincidence detector" account of the NMDA receptor cannot fully explain by amplitude alone. When a presynaptic spike arrives shortly before a postsynaptic action potential (presynaptic input helping to cause the postsynaptic spike), the synapse potentiates — a causally sensible rule, since the input appeared to contribute to the cell's firing. When a presynaptic spike arrives shortly after a postsynaptic spike (the input arriving too late to have caused anything, or arriving amid a refractory, hyperpolarized-then-repolarizing membrane), the synapse depresses. This asymmetric, millisecond-scale temporal window — typically effective within about ±20–40 ms of the postsynaptic spike — is what makes STDP a plausible substrate for encoding causal, sequential, and predictive relationships in neural circuits, a computational role foreshadowed in models of Hebbian Plasticity and the Postulate but only made precise by these experiments.

1. Two calcium regimes, one receptor. In the classic hippocampal CA3–CA1 Schaffer collateral synapse, both LTP and NMDA-receptor-dependent LTD begin with glutamate release and NMDA receptor activation, but the resulting postsynaptic calcium signal differs in amplitude and kinetics depending on stimulation frequency. High-frequency stimulation (~100 Hz bursts, tetanic stimulation) produces strong, summating postsynaptic depolarization that fully relieves the Mg²⁺ block of NMDA receptors, yielding a large, fast calcium transient in the spine. Low-frequency stimulation (~1–5 Hz for many minutes, the classic LTD-inducing protocol) produces weaker depolarization, partial Mg²⁺ unblock, and a smaller, more prolonged calcium rise.

2. Kinase versus phosphatase competition. The large calcium transient activates CaMKII, which autophosphorylates at Thr286, becoming persistently (calcium-independently) active — a molecular switch that outlasts the triggering stimulus and drives AMPA receptor phosphorylation and insertion into the postsynaptic density, the expression mechanism of LTP. The smaller, prolonged calcium transient instead preferentially activates calmodulin-bound calcineurin (PP2B), which dephosphorylates and activates protein phosphatase 1 by removing an inhibitory phosphate from inhibitor-1. Active PP1 dephosphorylates AMPA receptor subunits (notably GluA1 at serine residues phosphorylated by PKA and CaMKII) and dephosphorylates other synaptic proteins that stabilize AMPA receptors at the membrane.

3. Receptor removal: the expression mechanism of NMDAR-dependent LTD. Dephosphorylation destabilizes AMPA receptors in the postsynaptic density and promotes their internalization via clathrin-mediated endocytosis — a process requiring dynamin and clathrin adaptor proteins. The internalized AMPA receptors are shuttled to endosomes; some are degraded, others recycled. Net surface AMPA receptor number falls, synaptic conductance for a given quantum of glutamate release falls, and the EPSP evoked by a standard test stimulus becomes smaller — the electrophysiological signature of LTD. This mechanism was demonstrated directly by manipulations that block clathrin-dependent endocytosis (e.g., dynamin-inhibitory peptides), which prevent LTD without blocking LTP.

4. A second, independent LTD pathway: endocannabinoid-mediated (eCB-LTD). Many synapses, notably in the striatum, cerebellum, and cortex, express a presynaptically-expressed LTD triggered retrogradely. Postsynaptic calcium influx and mGluR (metabotropic glutamate receptor) activation stimulate synthesis of endocannabinoids (2-arachidonoylglycerol, 2-AG, being the principal one at most central synapses). These lipid messengers diffuse retrogradely across the synaptic cleft and bind presynaptic CB1 cannabinoid receptors, which are G-protein-coupled receptors that inhibit presynaptic voltage-gated calcium channels and activate potassium channels, thereby suppressing subsequent neurotransmitter release probability. Sustained eCB signaling can produce a lasting decrease in release probability — LTD expressed presynaptically rather than postsynaptically. This route is central to “heterosynaptic” depression at cortical and striatal synapses and interacts richly with dopamine signaling in basal ganglia learning.

5. Cerebellar LTD: a third mechanism, a distinct trigger. At parallel fiber–Purkinje cell synapses, LTD requires conjunctive activity: parallel fiber input (carrying AMPA-receptor-mediated depolarization and mGluR1 activation) paired within about 100–250 ms with climbing fiber input (a massive, all-or-none complex spike that floods the Purkinje dendrite with calcium via voltage-gated calcium channels). The combination of mGluR1-triggered diacylglycerol/IP3 signaling and the climbing-fiber calcium transient activates protein kinase C (PKC), which phosphorylates AMPA receptor GluA2 subunits and drives their clathrin-mediated internalization, weakening the parallel fiber synapse. This cerebellar LTD is the physiological substrate proposed (Marr, Albus, Ito) for motor learning — for instance, adaptation of the vestibulo-ocular reflex gain when visual and vestibular signals are experimentally mismatched.

6. Spike-timing-dependent plasticity: adding the clock. STDP experiments pair single presynaptic EPSPs with single postsynaptic action potentials (evoked by brief current injection) at a controlled, systematically varied time interval, Δt = t(post) − t(pre). The resulting synaptic weight change, plotted against Δt, produces an asymmetric, biphasic "STDP curve":

• Δt slightly positive (pre before post, within ~0–20 ms): LTP. The presynaptic glutamate release arrives while the postsynaptic membrane is still depolarized from the back-propagating action potential, so NMDA receptors experience strong, well-timed Mg²⁺ unblock exactly during glutamate binding, producing a large calcium transient.
• Δt slightly negative (post before pre, within ~0–20 ms): LTD. Glutamate arrives after the postsynaptic spike has already fired and the membrane is repolarizing/hyperpolarizing (afterhyperpolarization), so NMDA receptors see glutamate with weak, poorly-timed depolarization — a smaller, delayed calcium signal that favors phosphatase-driven depression, or in some circuits recruits distinct calcium-source pathways (e.g., voltage-gated calcium channels sensing the falling back-propagating action potential) that trigger LTD-specific cascades.
• Beyond roughly ±40–100 ms (varies by synapse type), no lasting change occurs — the window closes.

The back-propagating action potential (bAP) — an action potential initiated at the axon initial segment that actively (though with some attenuation) invades the dendritic tree — is the biophysical signal that lets the synapse "know" the postsynaptic cell just fired, and its precise timing relative to glutamate release and to the resulting NMDA-receptor-gated calcium transient is what the whole STDP window is built from.

Dudek & Bear (1992) — establishing frequency-dependent bidirectional plasticity. Working in hippocampal slices, Serena Dudek and Mark Bear delivered prolonged low-frequency stimulation (900 pulses at 1–3 Hz) to the Schaffer collateral pathway and recorded field EPSPs in area CA1. They found that this protocol produced a robust, lasting depression of synaptic strength (roughly 20–50% below baseline, stable for at least an hour, the operational criterion for "long-term" in a slice), and — critically — that this LTD was blocked by the NMDA receptor antagonist APV, exactly as LTP is blocked by APV. By systematically varying stimulation frequency, they showed a continuous relationship: very low frequencies produced maximal depression, intermediate frequencies produced no net change, and high frequencies produced potentiation — a single frequency-response curve later formalized as the Bienenstock-Cooper-Munro (BCM) sliding-threshold model, which posits that the very same synaptic history sets a moving threshold separating LTD-inducing from LTP-inducing activity. This was the first clean demonstration that LTD is not simply "LTP running backward" mechanically but is a distinct, NMDA-receptor-gated process occupying the opposite end of the same input continuum.

Mulkey & Malenka (1992) and Mulkey, Herron & Malenka (1993) — the phosphatase cascade. Rebecca Mulkey and Robert Malenka used pharmacological occlusion and blockade experiments to dissect the biochemistry downstream of the NMDA-receptor calcium signal in hippocampal LTD. They showed that intracellular loading of postsynaptic neurons with inhibitors of calcineurin (protein phosphatase 2B) or of protein phosphatase 1 blocked the induction of LTD without affecting baseline transmission or LTP, directly implicating a phosphatase cascade — as opposed to the kinase cascade (CaMKII, PKC) required for LTP — as the molecular decision point. This was among the first demonstrations that a single receptor (NMDAR) can trigger opposite plasticity outcomes depending on which downstream enzyme class the calcium signal recruits, giving direct experimental support to the calcium-amplitude/calcium-control hypothesis.

Ito, Sakurai & Fujita (1982) and subsequent work — cerebellar LTD. Masao Ito and colleagues showed that conjunctive stimulation of parallel fibers and climbing fibers onto Purkinje cells produced a long-lasting depression of the parallel-fiber-evoked Purkinje cell response, whereas either input alone did not. This was one of the earliest discoveries of any form of LTD (predating much of the hippocampal LTD literature) and directly supported the Marr–Albus–Ito theoretical framework proposing cerebellar LTD as the substrate of motor learning; later behavioral work confirmed that mice with mutations disrupting cerebellar LTD (e.g., in mGluR1 or specific PKC isoforms) show impaired vestibulo-ocular reflex adaptation and eyeblink conditioning, linking the molecular mechanism causally to a measurable behavior.

Markram, Lübke, Frotscher & Sakmann (1997) — the first clean STDP demonstration in cortical slices. Using dual patch-clamp recordings from synaptically connected pairs of pyramidal neurons in rat neocortical slices, Henry Markram and colleagues paired single presynaptic EPSPs with postsynaptic action potentials at controlled millisecond delays. They found that repeated pairing in which the presynaptic EPSP preceded the postsynaptic spike by less than about 10 ms produced robust potentiation, while pairing in which the EPSP followed the spike produced depression — establishing that the order, not merely the coincidence within a stimulation window, of pre- and postsynaptic activity sets the sign of plasticity, and that the window for this "causal" detection was on the order of tens of milliseconds, far finer than earlier tetanus-based LTP protocols could resolve.

Bi & Poo (1998) — quantifying the STDP curve. Guo-qiang Bi and Mu-ming Poo, recording from cultured hippocampal neurons with paired whole-cell recordings, systematically varied Δt across a wide range (roughly −100 ms to +100 ms) and plotted synaptic weight change as a continuous function of timing, producing the now-canonical asymmetric STDP curve: a steep potentiation peak for small positive Δt, a broader depression trough for small negative Δt, and a rapid falloff to no change beyond about ±40–50 ms. They further showed that the LTP side of the curve required NMDA receptor activation and postsynaptic calcium rise, and that disrupting the precise timing (e.g., by broadening the postsynaptic spike or altering dendritic backpropagation) reshaped the curve — directly implicating the back-propagating action potential as the timing signal read out by NMDA receptors. This paper is the definitive quantitative reference for the STDP curve reproduced in essentially every neuroscience course and textbook since.

Feldman (2000) and subsequent in vivo work — STDP outside the culture dish. Daniel Feldman demonstrated STDP-like rules operating in vivo/in intact cortical circuits (rat barrel cortex layer 2/3), showing that natural patterns of whisker-evoked activity, timed relative to postsynaptic spiking, reshape receptive fields in a manner consistent with the STDP curve — extending the phenomenon from reduced preparations (cultures, slice pairs) to the intact, behaving sensory system, and supporting the idea that STDP is not merely an artifact of paired stimulation protocols but a plausible in vivo learning rule.

Example 1 — Predicting the sign of plasticity from a stimulation protocol (BCM-style reasoning). A hippocampal slice physiologist delivers three different stimulation protocols to the Schaffer collateral pathway and records the resulting change in field EPSP slope, normalized to baseline (100%), one hour later:

Using the calcium-amplitude framework: Protocol A produces small, prolonged postsynaptic depolarization and calcium influx — sufficient to activate calcineurin/PP1 but below the threshold for strong, persistent CaMKII autophosphorylation — yielding LTD. Protocol C produces large, fast, summating depolarization via near-simultaneous EPSPs, strong NMDA receptor unblocking, and a large calcium transient that saturates CaMKII autophosphorylation at Thr286 — yielding LTP. Protocol B sits near the "modification threshold" (θm in BCM terms) where phosphatase and kinase activation are roughly balanced, producing no net change — this is the crossover point predicted by the sliding-threshold model, and its location shifts (moves rightward, requiring stronger stimulation for LTP) if the cell's average recent activity has been high, a homeostatic feature that keeps the system from runaway potentiation. Quantitatively, if we model the net synaptic weight change as depending on a threshold calcium concentration [Ca²⁺]_θ1 (below which nothing happens), [Ca²⁺]_θ2 (above which LTP dominates), and depression occurring between them:

• Protocol A: peak [Ca²⁺] ≈ 300 nM, sustained ~15 min → between θ1 (~200 nM) and θ2 (~500 nM) → LTD.
• Protocol C: peak [Ca²⁺] ≈ 1–5 μM transient, lasting milliseconds per burst but repeated → exceeds θ2 → LTP.

This illustrates why the same NMDA receptor, the same postsynaptic cell, and the same presynaptic terminals can be pushed in opposite directions purely by reshaping the temporal profile of calcium entry.

Example 2 — Reading an STDP experiment: predicting outcome from spike timing and diagnosing a "silent" synapse. A graduate student records a monosynaptic connection between a layer 5 pyramidal cell (presynaptic) and a layer 2/3 pyramidal cell (postsynaptic) in a cortical slice, and runs three pairing protocols, 60 repetitions each at 1 Hz, while recording EPSP amplitude before and 30 minutes after pairing:

The student now adds APV (an NMDA receptor antagonist) to the bath and repeats Trial 1. The expected result, based on the mechanism (bAP-timed NMDA receptor calcium influx as the trigger for the LTP side of STDP), is that potentiation should be abolished, since without functional NMDA receptors there is no calcium-permeable coincidence detector to translate the millisecond timing into a lasting weight change — this is exactly the control Bi & Poo and Markram et al. performed, and it is what allowed them to attribute the timing-dependence specifically to NMDA-receptor-gated calcium dynamics rather than to some independent, receptor-nonspecific timing mechanism (e.g., presynaptic facilitation). If instead the student find that Trial 1 (Δt = +8 ms) produces no initial EPSP at resting membrane potential but the connection becomes detectable only after pairing, this is the classic signature of "unsilencing" a silent synapse — one that expresses NMDA receptors but has few or no functional AMPA receptors at baseline. LTP-inducing pairing, by driving CaMKII-mediated insertion of AMPA receptors into a previously AMPA-receptor-poor postsynaptic density, converts a silent synapse into a functional one — a developmentally important complement to STDP that shows the same molecular machinery used for graded strengthening of mature synapses can also switch a synapse on entirely.

Mark Bear (with Serena Dudek) provided the pivotal demonstration that LTD is an NMDA-receptor-dependent, frequency-graded counterpart to LTP rather than an independent, unrelated phenomenon, work that grew directly out of his earlier theoretical collaboration with Leon Cooper and Elie Bienenstock on the BCM model of a sliding modification threshold. Robert Malenka and Rebecca Mulkey then dissected the downstream biochemistry, showing that calcineurin and protein phosphatase 1 — not kinases — execute hippocampal LTD, giving the calcium-control hypothesis its first direct pharmacological support. On the timing side, Henry Markram (with Bert Sakmann's group) and, independently and more systematically, Guo-qiang Bi and Mu-ming Poo established that the order of pre- and postsynaptic spiking, resolved to single milliseconds using paired intracellular recordings, determines the sign of plasticity — turning "cells that fire together wire together" into a precise, causally interpretable rule that has since become foundational to computational neuroscience and machine-learning-inspired models of synaptic learning.

Understanding LTD and STDP has reshaped how neuroscientists model learning, memory, and disease. In computational neuroscience and artificial intelligence, STDP rules are used directly as biologically-plausible local learning algorithms in spiking neural networks and neuromorphic hardware (e.g., Intel's Loihi chip, IBM's TrueNorth), because STDP requires only locally available information (pre- and postsynaptic spike times) rather than a global error signal, making it attractive for low-power, brain-inspired computing. In clinical neuroscience, non-invasive brain stimulation protocols such as paired-associative stimulation and certain transcranial magnetic stimulation paradigms are explicitly designed around STDP timing rules to induce LTP-like or LTD-like changes in human corticospinal excitability, with translational interest in stroke rehabilitation, chronic pain, and dystonia. In psychiatric and neurodevelopmental research, disrupted balance between LTP and LTD — for instance excessive mGluR-dependent LTD in Fragile X syndrome, linked to loss of function of the FMR1 gene product FMRP, which normally represses local dendritic translation of proteins required for AMPA receptor internalization — has become a major mechanistic and therapeutic target, with mGluR5 antagonists tested in clinical trials to normalize excessive LTD. In addiction neuroscience, endocannabinoid-mediated LTD at glutamatergic synapses onto striatal medium spiny neurons is disrupted by chronic drug exposure (cocaine, in particular), and restoring normal eCB-LTD has been proposed as a strategy for reversing drug-induced synaptic pathology. Finally, cerebellar LTD remains the leading cellular model for a specific, well-characterized form of motor learning (vestibulo-ocular reflex adaptation, eyeblink conditioning), providing one of the clearest end-to-end links from molecule to synapse to circuit to measurable behavior in the mammalian brain, and setting the stage for how Molecular Mechanisms of Memory Consolidation converts these transient synaptic changes into stable, structural, long-term memory, and how Neural Circuits and Network Dynamics scales single-synapse rules like STDP up to population-level computation.

Ask the tutor: How does dopamine modulate or reverse the sign of spike-timing-dependent plasticity in different brain circuits?