NMDA, glutamate & memory

Glutamate is the most abundant amino acid in the diet and one of the most abundant molecules in the brain, and almost none of that pool is transmitter. The brain has to keep a metabolic building block and a signalling molecule strictly separated, and the way it does so shapes everything else on this page.

The supply chain runs through astrocytes. Released glutamate is taken up largely by surrounding glia, converted to glutamine, which is inert as a signal, shipped back to the neuron, and converted back to glutamate there before being loaded into vesicles. That glutamate-glutamine cycle exists because glutamate cannot simply be left lying around. Dietary glutamate is a separate matter entirely; there is no evidence for brain damage in humans from glutamate in food, and the blood brain barrier is the reason ^[3].

The critical fact is the clearance mechanism. Acetylcholine is destroyed in the cleft by an enzyme. Glutamate is not: transporter proteins in the membranes of glia and neurons are the only significant mechanism for removing glutamate from the extracellular fluid, and there is nothing else ^[4]. Those transporters do more than mop up. They shape the time course of synaptic events, determine how much transmitter escapes to receptors outside the cleft, control cross-talk between neighbouring synapses, and feed glutamate back into the synthesis of GABA, glutathione and protein ^[4].

Two consequences follow, and they run through the rest of this article. First, ambient extracellular glutamate has to be held very low, because the receptors are sensitive and the resting level is what determines whether they are quietly active all the time. Second, anything that impairs transport, most obviously the energy failure of a stroke, raises glutamate everywhere at once with no backup system to catch it. That is the mechanism of excitotoxicity, and it is a transport failure before it is a receptor problem.

The NMDA receptor's defining trick is that glutamate alone is not enough to open it. At rest the channel is physically plugged by a magnesium ion, sitting in the pore and held there by the negative resting voltage inside the cell. Two groups reported that independently in 1984, and the voltage dependence of that block is the entire basis of coincidence detection ^[5][6]. Depolarising the cell pushes the magnesium out. Glutamate arriving at a depolarised cell therefore opens the channel; glutamate arriving at a resting cell does not.

There is a second requirement that took longer to appreciate. The receptor needs a co-agonist bound at a separate site before it will respond to glutamate at all. Glycine was shown to potentiate the NMDA response in 1987 ^[7], and in the forebrain the endogenous ligand at that site is largely D-serine, a D-amino acid that mammals were not thought to make. It is synthesised by serine racemase, an enzyme enriched in glial cells ^[8], and it acts as an endogenous ligand for the receptor's glycine site ^[9]. The implication is worth pausing on: astrocytes hold a vote in whether a given synapse is allowed to change, because they supply the permission molecule.

Put together, the receptor asks for glutamate, a co-agonist, and prior depolarisation, and only then does it pass current. What comes through is the actual message. The NMDA channel is unusually permeable to calcium, and calcium entering a dendritic spine is the trigger for the enzymatic cascade that follows. That is the physical meaning of "fire together, wire together": the receptor is a molecular AND gate whose output is a calcium pulse ^[1][14].

The co-agonist site is also the most accessible pharmacological handle on the whole receptor, which is why compounds like D-cycloserine, D-serine and sarcosine have all been studied as ways to make learning easier rather than harder.

Long-term potentiation was discovered in 1973 and has been worked on continuously ever since; thousands of papers later it remains the most compelling cellular model of learning available ^[11]. The sequence is now reasonably well agreed. A burst of high-frequency activity opens NMDA receptors, calcium floods a dendritic spine, and CaMKII, a calcium and calmodulin-dependent kinase, is activated. CaMKII activation is both necessary and sufficient to induce LTP, and it has held that status for more than twenty years ^[12][13].

What CaMKII then does is the part that makes the synapse stronger rather than just briefly excited. It drives more AMPA receptors into the postsynaptic membrane. AMPA receptors, not NMDA receptors, carry ordinary fast excitatory transmission, so adding them turns up the volume on that synapse for every future signal ^[12][14]. The NMDA receptor is the detector and the AMPA receptor is the effector, which is why compounds that potentiate AMPA are a separate class of interest; see ampakines.

The evidence tying this to behaviour is unusually direct. Chronic infusion of the NMDA antagonist AP5 into rat ventricles blocked LTP in the living animal and produced a selective impairment of place learning, the kind of learning that depends on the hippocampus, while leaving visual discrimination learning untouched; the inactive isomer of the drug did nothing ^[10]. That is close to the cleanest experiment available: the same molecule blocks the cellular mechanism and the corresponding behaviour, and spares behaviour that should not depend on it.

One honest limit. LTP is a model of memory rather than memory itself. It is measured in slices and in anaesthetised animals with stimulation patterns no natural experience produces, and the step from a strengthened synapse to a recalled event is still largely inferred ^[11][1]. The distinction matters commercially, because a compound shown to enhance LTP in a slice has demonstrated something real about a mechanism and nothing at all about whether a person will remember more. See BDNF and neuroplasticity for the parallel version of this problem.

Turn the NMDA receptor down and you get dissociation, a sense of detachment from body and surroundings, along with impaired memory formation for the duration. That is how ketamine, phencyclidine and the research-chemical dissociatives work; see dissociatives. The interesting part is that drugs which all block the same pore behave completely differently, and the reason is kinetics rather than target.

Memantine is the clearest illustration. It is an open-channel blocker with low affinity, a fast off-rate and strong voltage dependence, which means it leaves the channel quickly whenever the cell depolarises normally. Under ordinary synaptic activity it gets out of the way; under the sustained, low-level overactivation of disease it stays in. That combination is why it is tolerated at all, when higher-affinity blockers of the same site failed clinically because of intolerable side effects ^[15]. It is approved for moderate to severe Alzheimer's disease on the strength of a placebo-controlled trial showing a modest benefit ^[16].

Ketamine took the opposite path and became interesting for a reason nobody predicted. The first placebo-controlled trial gave a single intravenous dose of 0.5 mg/kg to seven patients with major depression and saw the 25-item Hamilton score fall by 14 points within 72 hours, against no change on saline ^[17]. A properly randomised trial in treatment-resistant depression confirmed a rapid antidepressant effect six years later ^[18]. The mechanism appears not to be the block itself but what follows it: ketamine rapidly activates the mTOR pathway, increases synaptic signalling proteins, and produces new functional spine synapses in prefrontal cortex, and blocking mTOR abolishes both the synaptogenesis and the behavioural effect ^[19]. In other words the antidepressant action looks like a plasticity effect downstream of a brief blockade, which is the reverse of how the drug is usually described.

The catch across the whole class is that impairing the learning receptor impairs learning. Every drug in the table below degrades memory formation while it is active, and the ones with high affinity and slow off-rates degrade it for longest.

Glutamate's power is also its danger. Too much of it over-activates NMDA receptors, floods neurons with calcium and kills them, a process called excitotoxicity that contributes to stroke damage and to several neurodegenerative diseases ^[2].

The classic experiment separated it into two stages. Exposing cultured cortical neurons to toxic glutamate produces an early phase of swelling that depends on extracellular sodium and chloride, and a later, slower phase of degeneration that depends on extracellular calcium; remove the calcium and the delayed killing largely goes away ^[20]. The lasting damage is the calcium arm, which is why the calcium-permeable NMDA receptor rather than the sodium-carrying AMPA receptor became the target.

Then came a genuine surprise. NMDA receptor activity is not simply bad in large amounts; where the receptor sits changes the sign of the outcome. Stimulating synaptic NMDA receptors drives nuclear calcium signalling and builds a neuroprotective programme, while stimulating extrasynaptic ones activates a distinct genomic programme that promotes cell death, with the two acting in opposition on the same intracellular pathways ^[21]. Neuroprotection therefore means enhancing one population while disrupting the other, and no drug that simply blocks the receptor can do that.

This is the best explanation available for one of the larger failures in neuropharmacology. A long list of NMDA antagonists including selfotel, aptiganel, eliprodil, licostinel and gavestinel all failed in clinical trials for stroke and traumatic brain injury. The proposed reason is that glutamate is destructive only during the acute phase immediately after injury, and afterwards resumes its normal physiological duties, which include promoting neuronal survival; blocking synaptic NMDA transmission during that window hinders recovery ^[22]. The drugs were not too weak. They were doing the right thing at the wrong time.

The practical version for a reader of this site: more glutamate is never the goal, and NMDA agonists such as NMDA itself are strictly laboratory tools rather than anything to consume. Magnesium's role is worth a line too, since magnesium L-threonate is sold on this logic; magnesium is the physiological pore blocker, so raising it slightly is a plausible dampener on excess NMDA activity, but the human evidence for a cognitive effect from any oral magnesium form remains thin.

If blocking the learning receptor impairs learning, the obvious question is whether enhancing it helps. The honest answer is that this has been tried carefully, in the best-designed setting anyone could pick, and the result is small.

The setting is fear extinction. Extinction is learning, it depends on NMDA receptors, and psychotherapy for phobias is a controlled, repeatable dose of exactly that learning, so a drug that made extinction stick better ought to show up cleanly. D-cycloserine, a partial agonist at the receptor's co-agonist site, was given to 28 people with acrophobia before two sessions of virtual-reality exposure therapy in a randomised, double-blind, placebo-controlled design; the D-cycloserine group improved more, at one week and at three months ^[23]. It was a genuinely elegant result and it launched a large literature.

The individual participant data meta-analysis is the one to read. Raw data were obtained for 21 of 22 eligible trials, covering 1047 participants, across specific phobia, social anxiety, panic disorder, obsessive-compulsive disorder and post-traumatic stress disorder. D-cycloserine was associated with greater improvement from before to after treatment, with an effect size of about 0.25, a small effect; the advantage at follow-up did not reach significance on the primary comparison; concurrent antidepressants did not moderate it; and none of the prespecified patient-level or study-level moderators predicted who responded ^[24]. A small, real, so far unpredictable augmentation is the accurate summary.

The other routes are earlier and thinner. Raising the co-agonist directly with D-serine or indirectly with sarcosine, which blocks the glycine transporter, has been studied mostly as an add-on in schizophrenia rather than for cognitive enhancement. Potentiating the effector side rather than the detector is the ampakine approach, and no ampakine has completed a positive controlled trial in a healthy population either.

What is genuinely not known here is worth stating plainly. Nobody has shown that enhancing NMDA function improves memory in a healthy adult. Everything positive comes from settings where the system was impaired or where a specific piece of learning was being trained, and the excitotoxicity literature is a standing argument that a system tuned this finely is unlikely to have a spare margin lying unused ^[21][22]. The receptor is a coincidence detector, and detectors are not usually improved by being made less selective.