The t-N-methyl-d-aspartate receptor: Making the case for d-Serine to be considered its inverse co-agonist

The N-methyl-d-aspartate receptor (NMDAR) is an enigmatic macromolecule that has garnered a good deal of attention on account of its involvement in the cellular processes that underlie learning and memory, following its discovery in the mid twentieth century (Baudry and Davis, 1991). Yet, despite advances in knowledge about its function, there remains much more to be uncovered regarding the receptor's biophysical properties, subunit composition, and role in CNS physiology and pathophysiology. The motivation for this review stems from the need for synthesizing new information gathered about these receptors that sheds light on their role in synaptic plasticity and their dichotomous relationship with the amino acid d-serine through which they influence the pathogenesis of neurodegenerative diseases like temporal lobe epilepsy (TLE), the most common type of adult epilepsies (Beesley et al., 2020a). This review will outline pertinent ideas relating structure and function of t-NMDARs (GluN3 subunit-containing triheteromeric NMDARs) for which d-serine might serve as an inverse co-agonist. We will explore how tracing d-serine's origins blends glutamate-receptor biology with glial biology to help provide fresh perspectives on how neurodegeneration might interlink with neuroinflammation to initiate and perpetuate the disease state. Taken together, we envisage the review to deepen our understanding of endogenous d-serine's new role in the brain while also recognizing its therapeutic potential in the treatment of TLE that is oftentimes refractory to medications.


Introduction
Glutamate, a non-essential amino acid, serves as ligand for most of the excitatory neurotransmitter receptors in the vertebrate CNS. It activates both ionotropic (iGluRs) and G-protein coupled metabotropic (mGluRs) receptors. The mGluRs will not be discussed here but see (Niswender and Conn, 2010;Reiner and Levitz, 2018) for a comprehensive review. The iGluRs are ligand-gated channels, which upon activation, allow for the selective/non-selective flow of cations (K + , Na + and Ca 2+ ) across plasma membranes of postsynaptic neurons, their typical locus of expression. They are classified into three functionally distinct subclasses and named according to their first discovered agonists as kainate receptors (KARs), α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid or AMPA receptors (AMPARs), and N-methyl-D-aspartate or NMDA receptors (NMDARs) Traynelis et al., 2010). The NMDARs were the first among the glutamatergic receptors to be discovered (Watkins and Jane, 2006) and deemed unique for requiring both the ligand (glutamate) and a co-agonist (glycine and/or D-serine) for full activation (Johnson and Ascher, 1987;McBain et al., 1989). NMDARs are Ca 2+ permeable as are GluA2-lacking AMPARs expressed by excitatory neurons early on in development (≤P15 in rat) (Kumar et al., 2002;Yin et al., 1999) or inhibitory interneurons (Geiger et al., 1995;Jonas et al., 1994;McBain and Dingledine, 1993;Zhou and Hablitz, 1998). However, unlike AMPA or kainite receptors, NMDARs are almost fully blocked by Mg 2+ at a neuron's resting membrane potential. Relief from Mg 2+ blockade is voltage dependent (Mayer et al., 1984;Nowak et al., 1984) and realized upon membrane depolarizations approaching 0 mV (Beesley et al., 2019;Yi et al., 2018), facilitated in part by AMPARs that are activated concomitantly (Fig. 1). Note that while "active" synapses co-express AMPA and NMDA receptors, "silent" synapses express NMDA but no AMPA receptors (Purves, 2008).

The NMDA receptor subunit composition
In addition to providing binding sites for the ligand glutamate and co-agonist glycine for activation, NMDAR subunits in their di-and triheteromeric configurations regulate various aspects of receptor function including voltage-dependence, Mg 2+ blockade, permeability to monovalent and divalent cations and kinetic properties. For example, the decay kinetics of canonical diheteromeric NMDAR-mediated synaptic currents depend on the type of GluN2 subunit (GluN2A-D) in the receptor subunit composition (Vicini et al., 1998).

The GluN1 subunit
The NMDARs are a family of heterotetrameric proteins that assemble as different combinations of four subunits. These subunits are derived from three gene families comprising seven distinct genes: Grin1, . The protein product of Grin1 (GluN1), which has eight splice variants (Dingledine et al., 1999), is the glycine/D-serine binding obligate subunit of all NMDARs. Each GluN1 variant has its own temporal and spatial pattern of expression throughout postnatal development (Laurie and Seeburg, 1994;Paupard et al., 1997; Zhong et al., Fig. 1. Modes of synaptic transmission and comparison of ion-flow in canonical GluN2-only diheteromeric NMDARs (inset at the bottom right of the postsynaptic neuron) and t-NMDARs (inset at the bottom left of the postsynaptic neuron) during dormancy (A), spontaneous, non-action potentialdependent release (B), and action potential triggered release (C-D) of glutamate. Note how spontaneous release of glutamate (B) and the availability of Dserine (D) affects responses in the two receptor subtypes and their dependence on post-synaptic depolarization for relief from Mg 2+ block, or lack thereof, for activation. 1995). The cell surface expression of NMDARs is thought to rely on this subunit (Kenny et al., 2009;Skrenkova et al., 2019).

The GluN3 subunit
The protein products of Grin3a-b (GluN3A-B) have not been as well explored as Grin1 or Grin2, given that they are among the last/latest NMDAR subunits to be discovered and characterized. GluN3 subunits, like GluN1, bind glycine/D-serine (Chatterton et al., 2002;Das et al., 1998) and aid in the surface delivery of NMDARs (Skrenkova et al., 2019). However, they impart a totally different functionality to NMDARs when assembled with GluN1 and/or GluN2 subunits, as discussed further on. Early work suggested that GluN3A-containing NMDARs were unlikely to express in the adult CNS (Ulbrich and Isacoff, 2008), although studies have reported seeing both subtypes, and GluN3A-containing NMDARs have been reported in the brains of both rodents and humans (Al-Hallaq et al., 2002;Beesley et al., 2019;Nilsson et al., 2007;Wong et al., 2002). A more recent study of the spatial and temporal expression of GluN3a mRNA confirmed that GluN3A might indeed be expressed widely throughout development into adulthood with strong expression specifically in CA1 hippocampus, entorhinal cortex, and basolateral amygdala (Murillo et al., 2021). Similarly, GluN3B expression has been shown to increase throughout development in the hippocampus, cerebral cortex, projection neurons and interneurons of the striatum, various cell types of the cerebellum, and motor neurons of the spinal cord (Wee et al., 2008). Given that GluN1 is also expressed in these regions makes it plausible for them to assemble as GluN1-GluN3A/B d-NMDARs with the caveat that they would be activated by glycine, and not glutamate (Wee et al., 2008). Indeed, non-canonical GluN3A-containing d-NMDARs have been reported to express in multiple areas of the adult brain and by chondrocytes in cartilage (Bossi et al., 2022;Grand et al., 2018;Otsu et al., 2019;Takarada et al., 2009), providing evidence contrary to early assumptions of a developmentally-regulated decline in GluN3A expression .
Given that GluN3-containing d-NMDARs are activated exclusively by glycine, it is conceivable that they are rendered constitutively active, rapidly desensitizing or both, making functional characterization and/or assessment of their role in the brain difficult. Indeed, GluN3-containing d-NMDAR function appears to be unmasked only in the presence of antidesensitizing agents like CGP-78608 (Bossi et al., 2022;Grand et al., 2018;Otsu et al., 2019). Furthermore, glycine levels in the synaptic cleft may rise as high as 2 μM during synaptic activation (Harsing and Matyus, 2013) while its potency as an agonist of canonical d-NMDARs in native tissue and/or in expression systems is in the sub micromolar range (EC 50 of <1 μM) (Johnson and Ascher, 1992;Kleckner and Dingledine, 1988), suggesting that glycine may regulate receptor function through activation and/or desensitization (Mayer et al., 1989).
It is unfortunate that GluN3-containing d-NMDARs be considered as bona fide NMDARs because they are unaffected by either glutamate or NMDA (Chatterton et al., 2002) and therefore most likely have been rebranded recently as excitatory glycine GluN1/GluN3A receptors (eGlyRs) (Bossi et al., 2022). Hence, these receptors may be construed as serving a unique hitherto unknown function or being a physiological byproduct of assembling glutamate-activated GluN3-containing triheteromeric NMDARs or t-NMDARs (e.g., GuN1-2A-3A-2A) from GluN1-3A/B, GluN1-2A/B or GluN3A/B-2A/B heterodimers and will be discussed no further in this review (unless stated otherwise, d-NMDARs referred to henceforth relate to glutamate activated GluN2-only d-NMDARs). Note that GluN2A/B-2B/A heterodimers are also plausible but unlikely to be trafficked to the membrane without GluN1 or GluN3A/B.

An introduction to t-NMDARs
Triheteromeric NMDARs comprise of three distinct subunits. However, receptors containing one or more subunits from each of the three known NMDAR gene families (GluN1-2A/B/C-3A/B) have been designated as t-NMDARs (e.g., GluN1-2A-3A-2A), to distinguish them from d-NMDARs (e.g., GluN1-2A-1-2A) and GluN2-only triheteromeric NMDARs (e.g., GluN1-2A-1-2B) (Kumar, 2016) (Fig. 2). Note that substituting a GluN1 subunit with GluN3 in GluN2-only triheteromeric NMDAR subunit composition yields t-NMDARs which comply with the requirements of two glutamate and two glycine binding subunits to make functional receptors that can be surface-expressed. Indeed, GluN1-2A-3A-2A-containing t-NMDARs are expressed by neurons within the adult mammalian CNS as evidenced by electrophysiological and pharmacological studies (Al-Hallaq et al., 2002;Beesley et al., 2019Beesley et al., , 2020a. t-NMDAR expression has been confirmed through assays of GluN3A mRNA (Murillo et al., 2021) and protein level expression within the murine brain in areas including the entorhinal cortex, basal lateral amygdala, and CA1 hippocampus using area specific tissue analysis (Beesley et al., 2022). t-NMDAR subunit composition can also be visualized using immunohistochemistry with subunit-specific antibodies targeting extracellular epitopes on NMDAR subunit proteins and high resolution confocal microscopy in the MEA and hippocampus (Beesley et al., 2023). t-NMDARs can also be readily distinguished from d-and GluN2-only triheteromeric NMDARs through mandatory GluN1 subunits of these receptors which require two glutamate and two glycine molecules for activation. It is deemed roughly three to four times more potent than glycine, the canonical co-agonist, in its binding. D-serine concentrations are known to vary significantly across brain regions− from trace amounts in the cerebellum, to high enough (~6.5 μM) to saturate all the glycine binding sites on NMDARs in the frontal cortex (Matsui et al., 1995;Priestley et al., 1995). This suggests that in certain localized environments, D-serine may well be the primary co-agonist rather than glycine and together with glutamate (ligand) drive NMDAR function. Without a brain-wide assessment of D-serine/glycine distribution though, it may be impossible to pinpoint the true co-agonist or understand the full extent of their interplay. D-serine is an inhibitor of glycine-activated, GluN3-containing d-NMDARs as shown using a heterologous expression system (Chatterton et al., 2002), and in vitro, in rodent cartilage-deriving chondrocytes that express both the receptors and serine racemase (SR) to enable the D-serine produced to regulate their differentiation during bone development (Takarada et al., 2009). Initially believed to be an antagonist of glutamatergic GluN3-containing t-NMDARs (Beesley et al., 2020a(Beesley et al., , 2021, we will now consider why D-serine better fits the role of an inverse co-agonist for this class of receptors (Fig. 3). Note that t-NMDARs, like their GluN2-containing di-or triheteromeric counterparts, also require the binding of two glutamate and two glycine molecules for activation. We would argue that although D-serine can compete for the glycine-binding sites on both GluN1 and GluN3 subunits, its antagonistic effects are likely mediated through GluN3 in glycinergic d-and glutamatergic t-NMDARs. This is because its competition with glycine for the binding site on GluN1 activates GluN2-only di-and triheteromeric NMDARs as their alternate co-agonist. D-serine would also compete with glycine for the binding site on GluN3, however, to now suppress activation of glycinergic d-and glutamatergic t-NMDARs like an antagonist would. It should be noted that D-serine's competition is with glycine for the co-agonist binding sites on either GluN1 or GluN3 and not with the ligand glutamate which binds to the GluN2 subunits of these receptors. Hence, it should be considered as an agent that brings about the same effect as a GluN2 subunit-preferring antagonist would except doing so through the glycine-binding site on GluN3 as an inverse co-agonist.
To be considered an inverse co-agonist, D-serine must bind to the same receptor (e.g., t-NMDAR) as a co-agonist (glycine) but induce a pharmacological response opposite to that of the co-agonist, which it does. Indeed, t-NMDAR mediated EPSCs evoked postsynaptically in neurons are diminished significantly by D-serine acting through GluN3 or when glycine is prevented from binding GluN1 to activate the receptor . While a neutral antagonist (e.g., 5, 7-Dichlorokynurenic acid or DCKA) acting at the glycine-binding site on the NMDAR complex has no activity in the absence of the co-agonist (i.e., glycine) or inverse co-agonist (i.e., D-serine) but can block the activity of either, an inverse co-agonist (D-serine) has opposite actions to those of the co-agonist (glycine) but the effects of both can be blocked by the antagonist (i.e., DCKA) as has been verified experimentally . Thus, glycine and D-serine qualify as co-agonist and inverse co-agonists of t-NMDARs respectively by satisfying the functional requirements of increasing the activity of the receptor above its basal level as co-agonist and decreasing the activity of the receptor below the basal level as inverse co-agonist. This leads to the prerequisite that there be a well-defined basal or intrinsic level of receptor activity, or the receptor be constitutively active for the co-agonist and inverse co-agonist to act upon in the absence of any ligand (glutamate). This appears to be a hard criterion to meet/demonstrate for ligand-gated channels let alone for canonical GluN2-only di-or triheteromeric NMDARs which rely on not only the ligand and a co-agonist for activation but are also blocked by Mg 2+ until depolarized postsynaptically. Incorporation of GluN3 into the subunit composition removes the requirement for postsynaptic depolarization by making t-NMDARs independent of Mg 2+ blockade and activatable at potentials closer to the neuron's resting membrane potential. Thus, t-NMDARs have a larger dynamic range of potentials during which they can operate compared with GluN2-only di-or triheteromeric counterparts. We posit that this flexibility imparts t-NMDARs to be constitutively active or responsive to a steady barrage of non-action potential dependent neurotransmitter release at synapses that can be recorded in neurons at rest. Indeed, in the presence of the sodium channel blocker tetrodotoxin (TTX), to inhibit the firing of APs, miniature excitatory postsynaptic currents (mEPSCs), which can range in frequency from 0.1 to 1.5 Hz or greater in various cell types of the medial entorhinal area, in acute brain slices (Jones and Woodhall, 2005;Kumar and Buckmaster, 2006;, constitute a basal level of background synaptic activity that can potentially activate t-NMDARs and render them constitutively active ( Fig. 1). Note that this type of activity, which can potentially be much higher in the intact brain, is distinct from evoked release in which the firing of action potentials by the presynaptic neuron triggers the release of neurotransmitter at the synapse. Thus, spontaneous EPSCs recorded in neurons comprise of action potential-dependent and non-action potential dependent events that can both be transduced by t-NMDARs on account of their rapidly desensitizing AMPAR-like kinetics (Beesley et al., 2019;Pilli and Kumar, 2014). Taken together these observations strongly suggest consideration of D-serine as a bona fide inverse co-agonist of GluN3-containing t-NMDARs. Ligand gated ionotropic receptors for which inverse agonists have been identified include the GABA A receptor whose agonists, like muscimol, bring about relaxant effects while their inverse agonists like Ro15-4513, agitation, convulsive or even anxiogenic effects (Mehta and Ticku, 1988;Sieghart, 1994).
Interestingly, D-serine can antagonize kainate-induced AMPARmediated currents in acutely isolated hippocampal neurons by reducing their affinity for kainate (Gong et al., 2007). Even though original estimates of D-serine concentration in various brain regions Matsui et al., 1995) were deemed an order of magnitude lower than those required for AMPAR inhibition, more sensitive measurement techniques developed recently have revised the estimates to 250 nmol/g in the rat brain, including the hippocampus (Hashimoto, 2002) and over 100 nmol/g (~1.8 μM) in the human cortex (Bendikov et al., 2007;Kumashiro et al., 1995). Our own measurements of D-serine levels in the rat entorhinal cortex were in the 12-25 nmol/g range (Beesley et al., 2020a). Importantly, D-serine is thought to be concentrated in SNARE protein-bearing vesicles whose release from astrocytes is Ca 2+ -dependent and can be realized upon stimulation with AMPA (Mothet et al., 2005) at specific target sites that preclude microdialysis-based measurements of its true concentration. This vesicular form of release is reminiscent of how classical neurotransmitters are made available in the synaptic cleft and hence from this perspective, D-serine lives up to its reputation as a gliotransmitter that can bind to and prime GluN2-only di-and triheteromeric NMDARs for activation or antagonize t-NMDARs at tripartite synapses. More recent studies have indicated that D-serine can be made available through not just astrocytes, but by SR-expressing neurons as well which can convert the L-serine supplied by astrocytes, via the Na + -dependent neutral amino acid (alanine-serine-cysteine) antiporter ASCT, to D-serine following its uptake via the Na + -independent neutral amino acid antiporter Asc-1 Fig. 3. Differential effects of glycine (co-agonist, solid purple circle), D-serine (putative inverse co-agonist, solid red circle) and an unspecified neutral antagonist (small solid black circle) on the conducting ("O" or open central pore, hatched)/non-conducting ("C" or closed central pore depicted by the solid black circle inside the pore) states of glutamate (ligand, solid blue circle)-bound t-NMDARs. (Martineau et al., 2014). It is reported that Asc-1 transporters can also mediate the concurrent non-vesicular release of both D-serine and glycine from neurons to modulate NMDAR synaptic activity and that astrocytic D-serine is not solely responsible for activating/suppressing synaptic NMDARs (Rosenberg et al., 2013). The antiporter properties of ASCT make possible for D-Serine to elicits robust efflux of intracellular L-serine and conversely, for physiological concentrations of L-serine to induce efflux of preloaded D-serine from astrocytes to maintain their synaptic concentration (Ribeiro et al., 2002).
Finally, depletion of endogenous D-serine, hyperactivation of t-NMDARs and the possibility of neuronal hyperexcitability shutting-off presynaptic release of D-serine to render postsynaptic t-NMDARs unchecked, have all prompted us to hypothesize that changes in normal homeostatic functions of astrocytes affected by injury alters their reactive status in a way that switches their neuroprotective role whereby they are no longer a source of D/L-serine in regions vulnerable to TLEmediated neurodegeneration. Indeed, L-serine under these circumstances may be converted by SR to pyruvate (via a β-elimination pathway) instead of D-serine (via a racemization pathway) in astrocytes to meet the high metabolic demand of attending to dead and dying cells during seizure activity and/or in neurons to attend to their hyperexcitable state (Beesley et al., 2020a) (Fig. 4).

D-serine
D-serine and D-aspartate are the most abundant D-amino acids found naturally in mammals. They are believed to be the only D-enantiomers derived from intrinsic enzymatic reactions (racemization) and are often referred to as the canonical D-amino acids (Fuchs et al., 2005;Sasabe and Suzuki, 2019). However, while there is much debate about the intrinsic origins of D-aspartate (Horio et al., 2013;Ito et al., 2016;Kim et al., 2010;Tanaka-Hayashi et al., 2015), D-serine, originally believed to be confined to prokaryotes, has been detected in various organs of a number of different species, including humans (Hashimoto and Oka, 1997;Wolosker et al., 1999), and in multiple structures of the CNS (Beesley et al., 2020a;Hashimoto and Oka, 1997;Klatte et al., 2013;Panatier et al., 2006). Within the brain, it has been found in proximity to regions of NMDAR expression, particularly in the prefrontal and temporal cortices in humans (Chouinard et al., 1993;Kumashiro et al., 1995) and within the rodent hippocampus and cerebral cortex (Hashimoto et al., 1993;Schell et al., 1995). D-serine is regarded as a co-agonist of GluN2-only di-and triheteromeric NMDARs, perhaps even more potent than glycine (Kleckner and Dingledine, 1988;Le Bail et al., 2015;Mothet et al., 2000;Traynelis et al., 2010). However, it inhibits glycine-activated GluN3-containing d-NMDARs (Chatterton et al., 2002) and work from our laboratory has shown that it antagonizes glutamatergic t-NMDARs, preventing their overactivation which can lead to excitotoxic cell death (Beesley et al., 2019(Beesley et al., , 2020a. As alluded to above, our current thinking is that D-serine better fits the role of an inverse co-agonist for these receptors, given the important distinction between a neutral antagonist, that can reduce the effects of an agonist or co-agonist (like glycine) by preventing its binding but having no effect by itself, and an inverse agonist or co-agonist which can diminish receptor function independently of the agonist or co-agonist. Increasingly recognized as contributors to neurological disorders (Paoletti et al., 2013), NMDARs are a veritable therapeutic target. Accepting and harnessing the subtle differences in their mechanisms of action is crucial to our understanding of their role and versatility in cellular processes as well as in designing appropriate molecules like D-serine with which to intervene in a number of neurological diseases including TLE (Beesley et al., 2020a;Klatte et al., 2013) and schizophrenia (Hashimoto et al., 2003) where its endogenous expression is differentially regulated with age, brain region and type of disease. Together, these observations provide an intriguing narrative that D-serine is locally regulated/dysregulated in normal/diseased states. D-serine is synthesized from L-serine by the enzyme serine racemase (SR) and co-enzyme pyridoxly-5-phosphate (PLP) Wolosker et al., 1999;Ye et al., 2021). Racemase enzymes have been classified as PLP-dependent or PLP-independent depending on the requirements of a co-enzyme, and SR is a PLP-dependent racemase (Conti et al., 2011) (Fig. 4). There are many eukaryotic SR genes that have been cloned to date, in mice (Strisovsky et al., 2005), plants (Fujitani et al., 2006(Fujitani et al., , 2007, yeast (Goto et al., 2009), and even soil-dwelling slime molds (Ito et al., 2012). Their enzymatic properties have been well-studied and documented given their high biological importance. Interestingly, little is known about prokaryotic SRs even though they are known to be in existence (Kubota et al., 2016). D-serine was initially thought to be synthesized by SR in astrocytes after its initial discovery in type II astrocytes of the gray matter (Schell et al., 1995); however, it is now evident that SR is expressed in both neurons and glia (astrocytes and microglia) (Beesley et al., 2020a;Panatier et al., 2006;Wu and Barger, 2004;Wu et al., 2007). It has been detected in various peripheral tissues/organs including the pancreas, adrenal glands, and testes of both mice and rats and SR mRNA has been found in multiple peripheral tissues/organs in humans (Xia et al., 2004), including pancreatic β-cells, where it is thought to regulate insulin production (Ndiaye et al., 2017;Suwandhi et al., 2018). However, there is evidence to support that peripheral D-serine may be acquired independently of SR, through dietary means, given that no differences in D-serine levels were found between control and SR knock-out mice (Horio et al., 2011). This suggests that D-serine in mammals may both be made endogenously or acquired exogenously.

L-serine
L-serine, like D-serine, diffuses poorly across the blood-brain barrier (Pernot et al., 2012), yet unlike D-serine (Hasegawa et al., 2019), it is not nephrotoxic and considered safe by the FDA. L-serine is widely available as a health supplement and well tolerated at relatively high doses (Ye et al., 2021). Indeed, multiple reports suggest that L-serine can be an effective neuroprotectant Metcalf et al., 2018) and has even found use in treating metabolic conditions such as alcoholic fatty liver disease (Sim et al., 2015). It has been reported to ameliorate Fig. 4. Schematic of key enzymatic reactions involved in the synthesis of Dserine and pyruvate from L-serine and their roles in the brain. Note that D-serine is synthesized from L-serine by the enzyme serine racemase (SR) and co-enzyme pyridoxly-5-phosphate (PLP). Racemase enzymes have been classified as PLPdependent or PLP-independent depending on the requirements of a coenzyme, and SR is a PLP-dependent racemase.
the inflammatory response to neurological trauma and aid in the management of traumatic brain injury (TBI) by reducing the neurological deficit score, lesion volume, neuron loss, cytokine levels (TNFα, IL1-β and IL-6), GFAP-positive astrocytes, Iba-1-positive microglia, and hyperpolarizing neurons through strychnine-sensitive glycine receptors that enhance Cl − inflow (Zhai et al., 2015). It nonetheless is conceivable that L-serine mediated reductions in astrocyte/microglia numbers and the mitigation of inflammation could simply be due to the curtailment of neuron loss as has been shown from work in our laboratory with D-serine which when introduced to specific brain regions of rats subjected to an epileptogenic insult prevents neuron loss and astrogliosis in these and interconnected regions thereby curtailing neuroinflammation and epileptogenesis. Given the similarities, it is conceivable that exogenously administered L-serine could be converted to D-serine within the brain by serine racemase expressing neurons/astroglia and exert its neuroprotective effects through t-NMDARs (Beesley et al., 2020a).

Serine racemase
SR has a dual role in mammalian biology as an enzyme catalyzing racemization of L-serine to D-serine and in the β-elimination of L-serine to pyruvate (Fig. 4), with the latter enzymatic reaction favored ~4-fold over the former (Graham et al., 2019;Nelson et al., 2017). This bifurcation may have evolved to maintain endogenous D-serine levels in the brain and serve as a checkpoint for its synthesis, although the negative impacts of excessive D-serine are not yet fully understood. In its role as an inverse co-agonist, D-serine could keep overactivation of GluN3A-containing t-NMDARs and the Ca 2+ influx through these receptors in check. A number of studies have shown that alterations of certain key amino acids in the 'triple serine loop' (identified by the three serine residues 150 through 152 in aspartate racemase) affect racemization activity that can be manipulated to influence SR kinetics (Foltyn et al., 2005;Uda et al., 2016Uda et al., , 2017. Amino acids 152 through 155 are key contributors to overall racemization efficiency within human SR and changes in amino acid residues within this region can significantly alter SR properties. For instance, mutations H152S; H152S and P153S; H152S and N154S; H152S, P153S and N154S can all render mouse SR enzyme racemase-only to various degrees of efficiency relative to wild type SR and get rid of the eliminase activity altogether (Graham et al., 2019;Uda et al., 2017). There is also evidence that glycine can serve as a SR inhibitor (Dunlop and Neidle, 2005) and therefore changes in local glycine concentrations, particularly in disease states like TLE, could potentially affect levels of endogenous D-serine. Stargazin is a well described accessory protein that binds to and facilitates integration of AMPARs into the plasma membrane, their recycling, and regulation of desensitization/deactivation (Priel et al., 2005;Tomita et al., 2005). Often considered an auxiliary subunit of the AMPA receptor, Stargazin can bind PSD-95 (post synaptic density protein-95) via PDZ domains 1 and 2 on its C-terminal, helping to anchor AMPARs to the plasma membrane (Schnell et al., 2002). PSD-95 is also known to interact with GluN2 subunits of the NMDAR (Kornau et al., 1995) thereby forming an AMPA/NMDA receptor complex together with Stargazin and PSD-95 (Ma et al., 2014). Interestingly, both PSD-95 and Stargazin interact with SR (Giaccari et al., 2022) and when bound to Stargazin, the enzymatic activity of SR is reduced, thereby diminishing D-serine/pyruvate availability locally at the synapse. Activation of AMPARs by glutamate causes dissociation of this complex and the resumption of SR enzymatic activity (Ma et al., 2014). It is intriguing to note that SR can be tethered to the membrane in proximity to the AMPA/NMDA receptor complex and be regulated by AMPA-dependent, GluN2-containing di-and triheteromeric NMDARs and/or AMPA-independent, GluN3-containing t-NMDARs.

t-NMDARs and ion selectivity
Subunit composition also affects voltage-dependent properties of NMDARs, especially those that incorporate the GluN3 subunit, and we have taken advantage of the differences in current-voltage (I-V) relationship profiles to distinguish between GluN3-containing and GluN3lacking NMDARs native to acute slices cut from intact brains. Pharmacologically isolated EPSCs for GluN3-containing t-NMDARs, measured using peak amplitudes at various holding potentials in the range − 85 to +20 mV, were found to have outwardly rectifying (OR) I-V profiles with EPSCs reversing polarity at hyperpolarized holding potentials instead of ~0 mV as observed for GluN2-only di-and triheteromeric NMDARs (Beesley et al., 2019;. The upshot of examining voltage dependence of the EPSCs was the realization that t-NMDARs have reduced affinity for Mg 2+ and increased selectivity for Ca 2+ over Na + Kumar 2012, 2014). Indeed, t-NMDAR-mediated EPSCs reverse polarity closer to the equilibrium potential for K + (E eq , K + ) and away from E eq , Na + , instead of ~0 mV, the reversal potential (E rev ) in d-NMDARs established jointly by Na + and K + . The permeability of GluN3-containing t-NMDARs to Cl − has not actually been assessed even though NMDARs have been suggested to be impermeable to anions (Mayer and Westbrook, 1987;Sharma and Stevens, 1996). Although Ca 2+ contributes minimally to E rev , owing to the relatively smaller concentrations inside and outside neurons under normal conditions, K + efflux through these receptors (OR current at hyperpolarized potentials) is countered by Ca 2+ influx (E eq , Ca 2+ ≫ 0 mV), contributing to an overall increase in intracellular positivity. Therefore, GluN3-containing t-NMDARs are excitogenic and deemed highly Ca 2+ permeable. This has been demonstrated using ion-substitution experiments in brain slices (Beesley et al., 2020b) and confirmed using modeling studies demonstrating how GluN3 alters the selectivity filters of canonical d-NMDARs rendering them selective for Ca 2+ over Na + , and in so doing, making them ~2 to 4-fold more Ca 2+ permeable   (Fig. 5). These observations, made in native t-NMDARs, are consistent with incorporating GluN3A into the subunit composition of oocytes-expressed GluN1/GluN2B-containing d-NMDARs which show ~5 to 10-fold reductions in NMDA-evoked Na + currents (Sucher et al., 1995). Given that co-injection of GluN3A with GluN1/GluN2 subunits in cell expression systems has also been reported to yield NMDARs with decreased unitary conductance and Ca 2+ permeability that is recapitulated in acutely isolated cortical neurons from very young (postnatal day 8) wild-type mice (Sasaki et al., 2002), there is a need for caution in interpreting receptor function across various experimental platforms.
Based on the permeability of NMDARs with different subunit stoichiometries to monovalent (i.e., Na + ) and divalent (i.e., Ca 2+ ) cations and the sequence of amino acids lining the selectivity filters within the pore forming membrane helices (M2) of these subunits, we have proposed GluN3 to be a regulatory subunit for ion selectivity in t-NMDARs (Beesley et al., 2020b). These studies have helped dispel the commonly held belief that NMDARs are non-selective for monovalent and divalent cations (Beesley et al., 2020b;Kumar and Kumar, 2021) (Fig. 5). Thus, while both Na + and Ca 2+ can displace themselves or each other from within the pore of the d-NMDAR channel, Na + can displace Ca 2+ only in d-NMDARs but not t-NMDARs due to the change in total ring charge. Ca 2+ , on the other hand, can displace itself or Na + in both d-and t-NMDARs. This suggests that once the pore of a t-NMDAR becomes occupied by Ca 2+ , it becomes impermeable to Na + and selective for Ca 2+ . Hence, without GluN3, the control over selectivity of both monovalent and divalent cations appears to be lost and this situation is rescued by incorporating GluN3 in the making of t-NMDARs, which then acquire selectivity for Ca 2+ over Na + . Taken together, this suggests that ion selectivity of the NMDARs under consideration is governed by the subunit-specific configuration of their selectivity filters and the charge permeability equation . These insights into ion selectivity were whetted in two independent brain regions where t-NMDARs are expressed, the somatosensory barrel cortex, where both d-and t-NMDARs can be found at separate inputs onto single neurons, and the MEA, where they are co-expressed at the same synapse.
Charge screening for monovalent and divalent cations may in fact begin in the extracellular vestibule itself of the NMDAR channel where charge neutralization or conversion of hydrophilic residues identified by a cluster of charged residues and a proline, the DRPEER motif, positioned C terminal to M3, and unique to the NR1 subunit alters fractional Ca 2+ currents in a manner consistent with its forming a binding site for Ca 2+ (Watanabe et al., 2002). Sequence alignment of the DRPEER motif in GluN1 (aa 658-663) with that of GluN3A (EKIYEE; aa 771-776) or GluN3B (DKTYEE; aa 671-676) reveal an overall net negativity that is greater than or equal to that of GluN1 (~1-2 depending on whether or not the last R is considered) suggesting that GluN3A/B subunits, like GluN1, might indeed screen for Ca 2+ , shepherding it from the extracellular vestibule of the channel (M3; aa 658-663 on GluN1) to the Fig. 5. Ion-selectivity in NMDARs depends on their subunit composition (insets at the bottom left of the postsynaptic neuron). Glutamatergic GluN2-only d-NMDARs are selective for and permeable to Na + , K + and Ca 2+ (A); Glycinergic GluN3-only d-NMDARs are selective for and permeable to Na + , K + but not Ca 2+ (B); Glutamatergic t-NMDARs are permeable to K + and prefer Ca 2+ over Na + (C) and consequently bring in ~2 to 4-fold more Ca 2+ than their GluN2-only triheteromeric counterparts (D). narrow constriction within the pore (M2; aa 608-624 on GluN1) where the selectivity filters reside. Additionally, differences in amino acids at the 'N site' of the selectivity filters are known to influence the binding efficacy of Mg 2+ (Wollmuth et al., 1998a(Wollmuth et al., , 1998b. Thus, asparagine which is present at the "N" site in both GluN1 and GluN2 subunits is replaced by glycine in GluN3 subunits and the N+1 site, which is presumed to facilitate Mg 2+ block, becomes positively charged with arginine replacing serine in GluN1-2A/B/C/D containing di-or triheteromeric NMDARs. Hence, in addition to increased Ca 2+ permeability, t-NMDARs exhibit significantly reduced or no Mg 2+ -sensitivity enabling them to activate independently of AMPARs and conduct outwardly-rectifying current over a broader range of membrane potentials compared with d-NMDARs, starting closer to a neuron's resting membrane potential (Beesley et al., 2019;. This increase in dynamic range has tremendous consequences for the vulnerability of neurons to Ca 2+ -induced excitotoxicity under conditions of hyperexcitability and hypersynchrony which define TLE (Beesley et al., 2020a(Beesley et al., , 2022.

The role of t-NMDARs in synaptic plasticity and information processing
NMDARs are mediators of synaptic plasticity in the brain, and this has been an area of active research for over three decades now. Many excellent books and reviews have been written and one only needs to consult PubMed for a worthy reference on this topic e.g., (Malenka and Nicoll, 1993;Tsien et al., 1996). Briefly, the basis of their role in synaptic plasticity stems from involvement in the phenomenon of long-term potentiation (LTP), a cellular mechanism for strengthening synapses through enhancements in the efficacy of neurotransmission, that is widely regarded as the basis for learning and memory. NMDARs make this possible through a) their voltage dependent properties, including Mg 2+ blockade, which dictate the conditions under which they activate and the nature of the postsynaptic response they elicit and b) their ability to influx Ca 2+ , a vital signaling molecule, that drives a cascade of intracellular events, activating kinases like the auto-phosphorylating Ca 2+ /calmodulin-dependent protein kinase II (CaMKII) (Lisman et al., 2012) and/or the cAMP-dependent protein kinase (PKA) (Park et al., 2021) whose catalytic subunits together with CaMKII phosphorylate serine/threonine residues on AMPA and NMDA receptors enabling them to potentiate responses by changing their conductance, mean open time and/or the probability of opening. The regulatory subunits of the kinases translocate instead to the nucleus where they serve as transcription factors to turn on immediate early genes that bring about synaptogenesis (Kandel, 2013). Synapses with just NMDA but no AMPA receptors are deemed "silent" despite fast ongoing synaptic neurotransmission through AMPARs that provide the initial depolarization to rid existent d-NMDARs of their voltage-dependent Mg 2+ blockade for activation. The fact that NMDARs require both a presynaptic signal, in the form of the ligand glutamate, and a postsynaptic signal, in the form of depolarization induced Mg 2+ unblock, to activate, has led to the notion that they serve as "co-incident detectors" of synaptic activity to presumably mark synapses for strengthening in accordance with Hebb's postulate (Hebb, 2002). Through changes in subunit composition, NMDARs can vary the amount of Ca 2+ they bring in such that synapses may also be weakened through a counter-LTP phenomenon called long-term depression or LTD (Massey and Bashir, 2007).
In addition to their role in synaptic plasticity, NMDARs influence the nature of information processing at the synapse by determining whether APs shunt EPSPs (Hausser et al., 2001;Koch, 1999). Thus, EPSPs with slower kinetics (i.e., NMDA) are shunted less than those with faster kinetics (i.e., AMPA) and are hence more effective at temporal summation. This feature restricts the temporal window for integration in neurons that lack mechanisms to prevent shunting of EPSPs during rapid firing of APs (Hausser et al., 2001). Accordingly, slower NMDAR-dependent EPSPs will be less susceptible to shunting by APs and therefore more efficient at synaptic integration compared with kinetically faster NMDAR-dependent EPSPs that are usually more efficient at coincidence detection and amenable to spike-timing-dependent plasticity (Kumar and Huguenard, 2003). This suggests that the temporal windows for summation of synaptic inputs onto neurons could vary, and AP-mediated shunting of NMDAR mediated EPSPs generated in these pathways could influence the way in which they integrate these synaptic inputs. Given that synaptic inputs evoking EPSPs through GluN2A containing d-NMDARs are likely to have faster kinetics than those evoking EPSPs through GluN2B containing d-NMDARs, the possibility that the two inputs are processed differently by neurons cannot be ruled out. Neurons can therefore be considered as integrate-and-fire devices from the perspective of the slower NMDAR-dependent local intracortical information processing, whereas they could operate as detectors of temporal coincidence of synaptic inputs and synchrony from the viewpoint of the faster, long-range, information processing (e.g., through the corpus callosum) (Kumar and Huguenard, 2003). Note that postsynaptic potentials are processed differently in the coincidence detection and temporal integration schemes, and these are two fundamentally distinct modes of information processing within complex neural networks (Konig et al., 1996;Shadlen and Newsome, 1994). While much is known about the role of GluN2-only di-and triheteromeric NMDARs in synaptic plasticity and information processing at the synapse, very little information is available on whether/how GluN3-containing t-NMDARs participate in these processes. Work from our laboratory has indicated that t-NMDARs mediate pathway-specific synaptic plasticity in the somatosensory barrel cortex Kumar, 2012, 2014) and likely in the entorhinal cortex where colocalization of t-and d-NMDARs at the same/single synaptic input onto excitatory neurons has been demonstrated (Beesley et al., 2019).
Synaptic colocalization of distinct NMDAR subtypes is thought to endow entorhinal cortical neurons with the ability of encoding distinct patterns of neuronal activity through single synapses. Primary whisker motor cortex (L1) synaptic inputs onto layer 5 pyramidal neurons in the rat somatosensory cortex are composed of GluN3A-containing t-NMDARs whereas thalamic/striatal (Str) synaptic inputs onto the same neurons are composed of GluN2A-containing d-NMDARs. In querying synaptic plasticity of these pathways, we found that L1, but not Str, synapses were potentiated following delta burst stimulation (dBS; 0.1-4 Hz, inter-burst interval of ~5 s). This potentiation was blocked by Dserine and/or intracellular BAPTA (1,2-bis(o-aminophenoxy)ethane-N, N, N′, N′-tetraacetic acid) suggesting that it was GluN3-specific and dependent on elevations in intracellular Ca 2+ . Interestingly, GluN2Bpreferring antagonists suppressed baseline L1 responses but did not prevent induction of dBS-evoked potentiation. Unlike L1, Str synapses were maximally potentiated following theta burst stimulation (tBS; 4-7 Hz, inter-burst interval of ~250 ms) and this potentiation was blocked with BAPTA and/or GluN2A-preferring antagonists. Furthermore, while dBS was both necessary and sufficient to potentiate L1 synapses, tBS was most effective in potentiating Str synapses suggesting co-dependence of synaptic potentiation on behaviorally inspired, brain wave-tuned, frequencies of burst stimulation and subunit composition of the underlying NMDARs (Pilli and Kumar, 2014). A model for predicting the likelihood of enhancing synaptic efficacy based on Ca 2+ influx through these receptors and integration of EPSPs at these inputs suggested that when inter-burst intervals are large, such as those for dBS, there is not much diminution of either burst amplitude or area under burst, as a function of burst number in the train for either L1 or Str synaptic inputs. However, much to our surprise, when inter-burst intervals are shortened from dBS to tBS, there is drastic attenuation of both burst amplitude and area under burst for L1 inputs, but not the Str inputs. Compared with Str responses, L1 responses are degraded rapidly, beginning as early as the second burst, and these differences are sustained throughout all bursts in the train. When inter-burst intervals are shortened further from tBS to alpha-burst stimulation (aBS; 8-12 Hz, inter-burst interval of ~115 ms), both L1 and Str responses now show attenuations of their burst parameters as a function of burst number in the train and the reductions are comparable in both pathways supporting the conclusion that integration of EPSPs at L1 and Str inputs during bursts is differentially affected by the inter-burst interval. This implies that Ca 2+ -dependent potentiation of L1 and Str synapses is indirectly regulated by the frequency of burst stimulation and subunit stoichiometry of the underlying NMDARs. Given that potentiation of synapses generally relies on the amount of Ca 2+ that enters through NMDARs during a burst ([Ca 2+ ] burst ) and as alluded to before, "burst efficiency" or the efficiency of Ca 2+ influx through NMDARs during a burst (η burst ), suggests that even though these parameters may be dependent on subunit composition, they are likely modulated by distinct aspects of NMDAR function. Thus, while [Ca 2+ ] burst may depend on Ca 2+ permeability, ion selectivity and single channel conductance, η burst may rely on factors such as integration of EPSPs, voltage-dependence of activation, Mg 2+ sensitivity and receptor desensitization/inactivation (Pilli and Kumar, 2014). These studies of t-NMDAR function have offered valuable insights into the dependence of synaptic efficacy on the interval of burst stimulation or the temporal nature of the synaptic input.

The role of t-NMDARs in temporal lobe epilepsy
Intriguing evidence obtained almost 3 decades ago suggested that Dserine could be used as an anticonvulsant (Loscher et al., 1994) but little to no information was forthcoming regarding its mechanisms of action until work from our lab re-examined its role as an inhibitor of t-NMDARs in the context of TLE (Beesley et al., 2020a(Beesley et al., , 2021. TLE is a serious neurological condition affecting around 1% of the world's population (Ren and Curia, 2021). Within the United States, it is estimated that ~50 new epilepsy cases per 100,000 individuals are diagnosed each year of which 9 are confirmed as TLE (i.e., ~30,000 new patients affected by TLE each year) (Asadi-Pooya et al., 2017). TLE is a focal epilepsy originating from specific areas of neuronal hyperexcitability within the brain and spreading to more distant secondary regions causing widespread neurodegeneration (McNamara et al., 2006). Despite availability of approved anti-epileptic medications, there is a high prevalence of drug resistance in as many as a third of the epilepsy cases diagnosed with TLE being the most common form of drug-resistant epilepsy (Laxer et al., 2014;Picot et al., 2008). TLE is a chronic insidious disease characterized by spontaneous recurrent seizures whose hallmark pathological features include marked loss of neurons in layer 3 of the medial entorhinal area (MEA) and within the hippocampus, particularly in the CA1 subfield. Cells loss has generally been attributed to hyperexcitability and hypersynchrony of neurons within these regions (Kumar and Buckmaster, 2006;Santos et al., 2021) and is invariably accompanied by neuroinflammation and astrogliosis in which both reactive astrocytes and microglia are upregulated within regions of neurodegeneration beginning as early as 1-5 days following the initial precipitating injury in animal models of the disease (Drexel et al., 2012;Shapiro et al., 2008;Wyatt-Johnson et al., 2017). Neuron loss and astrogliosis are also present throughout the hippocampus (Beesley et al., 2020a(Beesley et al., , 2021Covolan and Mello, 2000;Fujikawa, 1996;Morin-Brureau et al., 2018). TLE can therefore be viewed as a neurodegenerative disorder given that seizures associated with it bring about stereotypic neuron loss and circuit reorganization in specific regions of the brain. We attribute the refractory nature of anti-TLE medications in part to lack of clear understanding of basic mechanisms underlying its pathogenesis.
Recent work from our laboratory examined how focal administration of D-serine affects neuronal and glial function in the epileptic brain using the rat pilocarpine model of TLE (Beesley et al., 2020a). D-serine curtails neuron loss, prevents neuroinflammation and astrogliosis in the MEA and hippocampus, and effectively halts epileptogenesis. This is important given that D-serine, unlike many other anti-epileptic drugs, is well-tolerated as it is made in the brain. However, D-serine does not breach the blood brain barrier easily and there is a need for creative ways to make it available locally at the zones of neuronal vulnerability.
We hypothesized that the anti-convulsive, anti-seizure effects of D-serine observed are due to its inverse co-agonist actions on GluN3A subunit-containing t-NMDARs within the MEA (Beesley et al., 2020a). This hypothesis was tested using area specific tissue analysis (ASTA) in which brain chads (150-350 μm) sampled from specific areas within acute brain slices were used to assay protein expression. Indeed, GluN3A was found to express on a gradient in layer 3 of the MEA with lower expression in the medial relative to mid or lateral portions of the MEA where neuron loss was most conspicuous (Beesley et al., 2022). A similar gradient pattern of expression was observed between the CA1a (juxtaposed to the subiculum) and CA1c (juxtaposed to CA2) subdivisions of the hippocampus where greater cell loss, circuit disruption and gliosis was correlated with heightened expression GluN3A (Beesley et al., 2021(Beesley et al., , 2022Masurkar et al., 2017). This correlation between higher GluN3A expression and increased neuronal retention observed following focal application of D-serine in the rat pilocarpine model of TLE suggests that it acts through t-NMDARs as their inverse co-agonist. Consistent with this theory is the fact that endogenous D-serine levels in these brain regions are depleted under epileptic conditions rendering neurons vulnerable to Ca 2+ induced excitotoxicity (Beesley et al., 2020a). D-serine-mediated rescue of TLE pathology remains an actively pursued area of research in our laboratory and others to seek out therapeutic and interventional strategies to prevent/cure TLE.

Future directions
Work on t-NMDARs is far from being complete. Many open questions remain unanswered and yet there is plenty to be optimistic or even excited about dwelling into what their role might be in normal physiological processes of the brain, including synaptic plasticity and information processing, as discussed in this article. We would be interested in knowing where else in the brain, other than MEA and hippocampus, are t-NMDARs expressed, and to what purpose. We would want to visualize the subunit composition of these receptors to better delineate their structure and function from GluN3-containing d-NMDARs or GluN2only di-and triheteromeric NMDARs, particularly in the hippocampus, the testbed for studies relating learning and memory. We would want to better understand how these receptors affect behavior and the development of a conditional knock out animal would go a long way in facilitating efforts on this front. From the pathophysiological perspective, we would like to learn what expression of these receptors in regions like the amygdala might have on phenomena like fear conditioning and PTSD. Within the MEA, we would like to learn more about the involvement of these receptors in Alzheimer's disease which, like TLE, also entails loss of neurons and neuroinflammation. In this regard, we would be particularly interested in knowing whether D-serine would be helpful in any way in mitigating some of these deficits as seen with TLE. Finally, we would like to explore ways to boost endogenous D-serine levels in the brain to harness its therapeutic potential by keeping the deleterious effects of t-NMDAR hyperactivation in check, especially during conditions of seizure-related hyperexcitability.

Declaration of competing interest
The authors declare the following financial interests/personal relationships which may be considered as potential competing interests. Sanjay S. Kumar reports financial support was provided by National Institute of Neurological Disorders and Stroke.

Data availability
No data was used for the research described in the article.

Appendix A. Supplementary data
Supplementary data to this article can be found online at https://doi. org/10.1016/j.neuropharm.2023.109654.