How can I keep HEK cells alive while expressing NMDA receptors?

How can I keep HEK cells alive while expressing NMDA receptors?

We are searching data for your request:

Forums and discussions:
Manuals and reference books:
Data from registers:
Wait the end of the search in all databases.
Upon completion, a link will appear to access the found materials.

I am trying to express functional NMDA receptors in HEK293 line cells for single channel recording experiments.

The HEK cells are maintained in the standard way (Thomas & Smart 2005) and transfected with NR1 and NR2x subunit cDNAs and also GFP, using either lipofectamine or calcium phosphate precipitation. GFP expression suggests successful transfection, but cells exhibiting green fluorescence are uniformly swollen and dead and it's more or less impossible to obtain a successful patch. Untransfected cells seem to remain perfectly viable.

On the basis that the toxicity might be a consequence of Ca2+ influx through open NMDARs, I've tried including a cocktail of blockers in the growth medium, including AP5, kynurenic acid and Mg2+, but the transfected cells continue to die.

Can anyone suggest anything else I should be doing to keep the cells alive? Or am I just on a hiding to nothing? Other researchers seem to have managed this (eg Medina et al 1995, Vicini et al 1998) and do not seem to be doing anything substantially different, so I'm a bit at a loss.

Try perhaps lowering the level of expression of your construct. HEK cells are notorious for expressing large quantities of your transfected insert. Do you have it under a CMV promoter? Try a milder one, such as an ubiquitin promoter or similar or, even better, a doxycyclin-inducible system (e.g., clontech's tet off) so you can manage the expression levels through the dox concentration/time of induction.


In acute ischaemic brain injury and chronic neurodegeneration, the primary step leading to excitotoxicity and cell death is the excessive and/or prolonged activation of glutamate (Glu) receptors, followed by intracellular calcium (Ca 2+ ) overload. These steps lead to several effects: a persistent depolarisation of neurons, mitochondrial dysfunction resulting in energy failure, an increased production of reactive oxygen species (ROS), an increase in the concentration of cytosolic Ca 2+ [Ca 2+ ]i, increased mitochondrial Ca 2+ uptake, and the activation of self-destructing enzymatic mechanisms. Antagonists for NMDA receptors (NMDARs) are expected to display neuroprotective effects, but no evidence to support this hypothesis has yet been reported. A number of clinical trials using NMDAR antagonists have failed to demonstrate neuroprotective effects, either by reducing brain injury or by preventing neurodegeneration. Recent advances in NMDAR research have provided an explanation for this phenomenon. Synaptic and extrasynaptic NMDARs are composed of different subunits (GluN2A and GluN2B) that demonstrate opposing effects. Synaptic GluN2A-containing and extrasynaptic GluN2B-containing NMDARs have different co-agonists: d -serine for synaptic NMDARs and glycine for extrasynaptic NMDARs. Both co-agonists are of glial origin.

The mechanisms of cell destruction or cell survival in response to the activation of NMDAR receptors depend in part on [Ca 2+ ]i and the route of entry of this ion and more significantly on the subunit composition and localisation of the NMDARs. While synaptic NMDAR activation is involved in neuroprotection, the stimulation of extrasynaptic NMDARs, which are composed of GluN2B subunits, triggers cell destruction pathways and may play a key role in the neurodegeneration associated with Glu-induced excitotoxicity. In addition, it has been found that synaptic and extrasynaptic NMDA receptors have opposing effects in determining the fate of neurons. This result has led to the targeting of nonsynaptic GluN2B-containing NMDARs as promising candidates for drug research. Under hypoxic conditions, it is likely that the failure of synaptic glutamatergic transmission, the impairment of the GluN2A-activated neuroprotective cascade, and the persistent over-activation of extrasynaptic GluN2B-containing NMDARs lead to excitotoxicity. Fluoxetine, a drug widely used in clinical practice as an antidepressant, has been found to selectively block GluNR2B-containing NMDARs. Therefore, it seems to be a potential candidate for neuroprotection.


► Role of extrasynaptic GluN2B subunit-containing NMDA receptors in excitotoxicity. ► Role of synaptic GluN2A subunit-containing NMDA receptors in neuroprotection. ► Effect of fluoxetine. ► Clinical studies with different glutamate receptor antagonists. ► Selective GluN2B subunit-containing NMDA receptor antagonists.

Access options

Get full journal access for 1 year

All prices are NET prices.
VAT will be added later in the checkout.
Tax calculation will be finalised during checkout.

Get time limited or full article access on ReadCube.

All prices are NET prices.


Mammalian cells such as the human embryonic kidney 293 (HEK-293) and the Chinese hamster ovary (CHO) cells are widely used as hosts to express recombinant proteins to study their structural, biophysical, and pharmacological properties (Baldi et al., 2007 Dalton and Barton, 2014). HEK-293 cells in particular are an attractive heterologous system for expression of membrane proteins not least because they have post-translational modification machineries required for the proper folding and/or optimal biological activity of target proteins. They also exhibit high transfection efficiency, faithful translation, and processing of proteins (Wurm, 2004) that will result in higher protein yields (Backliwal et al., 2008a) as compared to other mammalian cells, e.g., CHO cells (Bollin et al., 2011). These attributes together with the cell size, morphology, rapid division rate, the ease of maintenance, and the low expression of native channels as well as the capacity to express transgenic receptor proteins and ion channels with high fidelity (Thomas and Smart, 2005), have established HEK-293 cells as a host of choice for transient heterologous expression of membrane proteins for structural studies (Nettleship et al., 2008 Chaudhary et al., 2011 Andrell and Tate, 2013), biopharmaceutical (Thomas and Smart, 2005 Jager et al., 2013), and electrophysiology applications (Lemtiri-Chlieh and Ali, 2013).

Despite these advantages, high-level expression of complex membrane proteins such as ion channels and trans-membrane receptors originating from a different species for current-voltage measurements has remained a challenge (Gan et al., 2006 Allen et al., 2009). For electrophysiology in particular, a high expression of proteins is critical for single-cell current recordings in the whole-cell mode. Membrane protein biosynthesis in the host is limited by the different composition of lipid bilayers between human and other species that may prevent proper folding of expressed proteins into their functional native three-dimensional conformations. As such, optimization of plasmids, culture media, growth conditions, or combinations thereof have been undertaken in the past to enhance the expression of membrane proteins (for review, see Jäger et al., 2015). These include the introduction of mild hypothermia (Wulhfard et al., 2008 Lin et al., 2015) and the addition of inhibitors of histone deacetylase to the culture media (Fan et al., 2005 Backliwal et al., 2008b). In addition to the essential elements required for the expression of recombinant protein, vectors can also be synthetically engineered to include optimized introns and codon usage (Gustafsson et al., 2004) and post-transcriptional regulatory elements (Mariati et al., 2010) to increase yields by stabilizing recombinant transcripts.

The delivery of the recombinant vector into the host cell and the detection of expressed proteins are the two critical stages crucial to the expression and study of recombinant proteins in HEK-293 cells. Recombinant proteins expressed in transiently transfected cells can be identified with a method that requires the simultaneous co-transfection of a lymphocyte surface marker antigen (CD8-alpha) or a fluorescent protein plasmid in addition to a second expression vector that contains the gene of interest (Lemtiri-Chlieh and Ali, 2013). This method enables visual identification of individual cells decorated with anti-CD8 antibody coated polystyrene beads (Jurman et al., 1994 Fortin and Hugo, 1999) under the light microscope or fluorescence detection of cells co-expressing the fluorescent proteins (Bestvater et al., 2002 Lin et al., 2015) and has been used for the study of ion channels and receptor proteins including the mammalian N-methyl-D-aspartate (NMDA Ehlers et al., 1996) and α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) receptors (Swanson et al., 1997 Lin et al., 2015), the Arabidopsis thaliana cyclic nucleotide-gated ion channels (AtCNGCs) (Leng et al., 2002 Hua et al., 2003) and Arabidopsis K + transporters (AKTs) (Lacombe et al., 2000 Cherel et al., 2002). While simple, rapid and inexpensive, expression of the CD8-alpha marker antigen or the fluorescent protein in the transfected cells has no direct correlation with the transfection efficiency and expression of the recombinant protein as the success rate of the introduction of both these marker and target gene expression vectors into the cells may be highly variable. Recent methods using fluorescent tags such as a dye-sensitive epitope (Tour et al., 2003 Rudner et al., 2005) or a fluorescent protein fusion (Snapp, 2005) provide a direct and better correlation between fluorescent signal and transfection efficiency and protein expression level although fusion tags such as fluorescent proteins may cause structural constrains that interfere with protein function. These fluorescent detection approaches have enabled successful expression of a number of membrane proteins including the connexin-43 (Gaietta et al., 2002), the AMPA receptors (Ju et al., 2004), the G protein-coupled receptors (Hoffmann et al., 2005) and the human ether-a-go-go-related gene (hERG) channel (Claassen et al., 2008 Huang et al., 2011) that were subsequently used for protein localization and trafficking as well as current-voltage measurement studies.

Although the expression of several membrane proteins and ion channels in HEK-293 cells have been reported previously, a detailed authoritative protocol that describes the key stages in the transient transfection and in-cell detection of recombinant proteins optimized for single-cell applications, is currently lacking. Here, we use the transient expression of an A. thaliana guard cell outward-rectifying K + channel, AtGORK (At5G37500) in HEK-293 cells as an example to assess current commonly used transfection reagents and the fluorescent detection methods, and provide a specific protocol that is easily accessible for the general expression of membrane proteins in HEK-293 cells suitable for biological characterization. As an example, AtGORK represents: (1) a difficult to express multi-pass membrane protein, (2) originates from a different species, and (3) needs to assemble into a heteromeric complex to achieve functionality. These three characteristics can hamper optimal expression of membrane proteins in HEK cells. In addition to this authoritative step-by-step protocol, we also include cautionary measures, and propose optimization strategies and recommendations extendable and amendable for different applications or proteins.


Inhibitory Effect of NPA on N2A-Containing Receptors.

We recorded currents from HEK293 cells transiently expressing GluN1-1a and GluN2A subunits with the whole-cell patch-clamp method. The extracellular solutions contained maximally effective agonist concentrations (1 mM glutamate and 0.1 mM glycine each 𾄀-fold EC50) EDTA at 0.01 mM and had low proton concentration (pH 8, 10 mM HEPBS). Inhibition was measured as decrease in the steady-state current elicited by glutamate following addition of NPA (0� mM). The limited solubility of NPA prevented us from investigating higher concentrations. Fitting the data with the Hill equation predicted �% maximal inhibition, with 1.9 mM NPA producing half-maximal inhibition ( Fig. 1 ). This IC50 value is larger than previously measured in Xenopus oocytes (Costa et al., 2010), where 0.1 mM NPA produced 30% inhibition. We attribute this discrepancy to differences in experimental conditions because these previous measurements were done at pH = 7.4, where 2A receptors are tonically �% inhibited (Traynelis and Cull-Candy, 1990) and in the absence of chelators, when trace zinc and magnesium ions further inhibit 2A currents (Nowak et al., 1984 Paoletti et al., 1997).

NPA inhibition of steady-state glutamate-elicited N1/N2A receptor currents. (Left) Whole-cell responses were recorded from human embryonic kidney 293 cells expressing N1/N2A receptors. Bars indicate glutamate (1 mM) applications (white) and NPA coapplications (gray). (Right) Levels of equilibrium responses measured after each NPA concentration was normalized to the responses when only glutamate was applied. Circles represent means of normalized values across cells. Line represents the fit of the logistic function to means of normalized responses for each concentration: 0.4 mM (n = 4), 1 mM (n = 4), 2 mM (n = 4), 4 mM (n = 5), 10 mM (n = 4). IC50 is expressed as 95% confidence interval.

Single-Channel Kinetics of NPA-Bound N2A-Containing Receptors.

To investigate the mechanism of NPA inhibition we recorded currents from one-channel cell-attached patches of HEK293 cells expressing N1/N2A receptors with extracellular (pipette) solution containing 4 mM NPA (𢏂-fold IC50). Similar to our whole cell recordings, the extracellular solution also contained glutamate (1 mM), glycine (0.1 mM), and EDTA (1 mM) to remove trace divalent cations. We clamped proton concentrations at 10 nM (pH 8) with 10 mM HEPBS ( Fig. 2A ). As controls, we used a set of recordings obtained in identical conditions and the absence of NPA (CTR) ( Fig. 2B ). Both data sets included only records that originated from a single active channel, and we processed and analyzed all records as described in detail previously (Kussius et al., 2009).

Effects of NPA on single-channel activity of N1/N2A receptors. Traces represent steady-state inward sodium fluxes recorded from cell-attached patches that contained in the recording pipette one active N1/N2A channel: (A) with NPA (4 mM), and (B) without NPA (CTR). For each condition, a 20-second segment is illustrated at two time resolutions in top and middle panels, respectively bottom panels expand the underlined segment and are displayed filtered, as for analyses, at 12 kHz. All traces represent Na + currents as downward deflections from a zero-current baseline Po indicates the open probability calculated for the entire parent record.

We found that 4 mM NPA decreased the average equilibrium open probability (Po) of 2A receptors to 38% of the CTR with no change in the single-channel amplitude ( Table 1 ). Thus, the NPA concentration selected was sufficient to produce a substantial effect on channel gating and had no effect on single-channel conductance. Further, we were able to attribute the decrease in Po to an �% increase in the mean closed time, and an �% decrease in mean open time. Next, we examined in closer detail the mechanism by which these kinetic effects arose.


Effects of NPA on average kinetic properties of individual 2A receptors

AmplitudePoMCTMOTnDurationEvents × 10 6
pA msms min
CTR10.9 ± 0.80.53 ± 0.044.7 ± 0.65.4 ± 0.751722.0
NPA10 ± 0.90.2 ± 0.3 * 9.8 ± 0.8 * 2.5 ± 0.4 * 61171.0
%Change +109

NPA-Bound Receptors Had More Frequent Long-Closures and Short-Openings.

The gating mechanism of the 2A-type NMDA receptors has been well characterized kinetically. In the conditions employed in this study, the steady-state gating reaction consists of transitions between five closed and two open states and occasional gating mode shifts. To determine how NPA increased mean closed time and decreased MOT we examined the distribution of closed and open events present in our single-channel records.

Based on a log likelihood criterion, all the one-channel records we obtained with NPA (n = 6) were well described with five closed components (E1 – E5), as we had reported previously for 2A receptors (Kussius et al., 2009). This result indicates that NPA binding did not produce additional closed conformations. Instead, it altered specific closed component areas: E3 and E5 increased 𢏂-fold and E4 increased 𢏅-fold, at the expense of a 2-fold decrease in the short E1 component area. In addition, the duration of E3 increased 1.6-fold ( Fig. 3, A and B ). Generally, the short closed components E1𠄾3 encompass the brief closed events within bursts, and occur along the activation pathway. Therefore, these results strongly imply that NPA can influence receptor activation. Notably, NPA did not change the time constant of the longest closed component (E5), which encompasses the longest desensitized events ( Fig. 3, A and B ), but increased its area, and the increased abundance of desensitized events was accompanied by �% shorter bursts (0.9 seconds versus 0.4 seconds). Note, however, that the observed increase in the E5 area does not necessarily imply that NPA affected desensitized state(s). Desensitization will occur more frequently simply if the occupancies of adjacent states increase, as will be discussed later ( Fig. 4A ).

NPA increased the areas of specific closed components. (A) Closed intervals observed in two records obtained from N1/N2A receptors in the absence (CTR, 416,413 events) and presence of NPA (4 mM, 1,347,473 events). Probability density functions (gray lines) were calculated by fitting kinetic 5C4O state models (see Fig. 4 ) to the displayed data dark lines represent individual exponential components their time constants [τ, milliseconds (ms)] and areas (a, %) are given as insets. (B) Summary of closed time constants and areas in the two conditions (CTR, n = 5 NPA, n = 6). Average values for CTR are given below each component (in ms). *P < 0.05 (Student’s t test).

In contrast to the specific and robust effects described above for closed events, NPA binding had no significant effects on the durations of any of the four open components. It caused a 3-fold increase in the frequency of the shortest open component, leading to an overall decrease in MOT ( Fig. 3, C and D ). During activation, NMDA receptors visit two types of open states: first a short 𢏀.2 millisecond state, and then a longer state whose duration defines the mode of opening: 2 milliseconds for low, 6 milliseconds for medium, and 11 milliseconds for high modes, respectively (Popescu and Auerbach, 2003). Since neither the durations nor the relative areas of the low, medium, or high open states changed with NPA ( Fig. 3D ) we conclude that NPA binding did not affect the receptor’s modal transitions. Rather, overall shorter mean open durations reflect a propensity to dwell more often in the shorter of the two open states, regardless of mode.

Together with the observation that NPA did not alter channel conductance, these results exclude the possibility that NPA reduces NMDA receptor currents by blocking the pore or by lengthening desensitized events, and strongly suggest NPA as an allosteric modulator, which primarily affects the NMDA receptor activation.

Kinetic Models of NPA Actions on 2A Receptors.

To interpret the kinetic changes described above, we used an established 5C2O kinetic scheme to fit the sequence of events detected in each single-channel record and averaged the optimized rate constants within each data set. Although this model is a simplification of the complex rearrangements that occur during NMDA receptor gating, it has been shown previously to reproduce accurately key features of its microscopic and macroscopic behaviors (Popescu et al., 2004 Zhang et al., 2008 Kussius et al., 2009). The fitting results showed that NPA significantly changed only three of the 12 rate constants included in the model: kC1-C2, kO1-C1, and kC1-O1. Of these three rate constants, kC1-C2 and kO1-C1 increased 1.4- and 2-fold, respectively, while kC1-O1 decreased 2-fold ( Fig. 4A ). Thus, this mechanism explains the decrease in current as increased occupancies of states C3 and C2, which reside along the activation pathway ( Fig. 4B ). Next, we used the reaction mechanisms deduced for CTR and NPA conditions to compare the relative free-energy fluctuations experienced by NMDA receptors in the absence and presence of NPA. We aligned the two profiles arbitrarily at the free-energy level of the long open state O2, based on the observation that openings were of similar durations in both conditions. According to this representation, when NPA is present, albeit at subsaturating concentrations, the free energy barrier to receptor activation is higher, due to more stable preopen states ( Fig. 4C ). In other words, when NPA is present receptors require more energy to open, and because they dwell longer in preopen states they also have more chances to desensitize.

The description above presents an average picture of receptor behavior in the presence of subsaturating concentrations of NPA and reflects a dynamic mixture of NPA-free and NPA-bound receptors. To more precisely describe the gating of NPA-bound receptors, we expanded the 5C2O model to contain two tiers, one each for NPA-free (top) and NPA-bound (bottom) states. As a first approximation, we connected the tiers through just one NPA association/dissociation step, which we positioned in turn between each of the like-states along the activation pathway. Essentially, the modeling was conducted as described previously (Amico-Ruvio et al., 2012). Along the NPA-free arm, we fixed all rate constants to the values obtained for the CTR data set whereas along the NPA-bound arm, we fixed only the rate constants that were not significantly changed in the NPA data, to same values as the CTR arm ( Fig. 4A ). We fitted separately each of the five-tiered models, which differed in the location of the connecting step (C3, C2, C1, O1, and O2), to individual single-channel records in the NPA data set. We observed that the NPA dissociation rate constants estimated with these models fell into two categories: they were slow when tiers were connected through C3 to C2, and they were 𢏅-fold faster when tiers were connected C1 through O2. The models with single connection at different states were ranked using a LL criterion (O1>O2Ϭ1Ϭ2Ϭ3). The O1 connected model returned the highest LL ( Fig. 5A ), which is 41 units larger than that of the second-ranked model, where the connection was at O2. These results suggested that the NPA binding may be state-dependent, with binding and dissociation occurring preferentially from open states. Because this model does not allow NPA binding to closed states, it predicts that the receptor should be insensitive to NPA, if the drug is applied just before but not during stimulation with glutamate.

NPA binding kinetics. (A) Tiered model was used to represent the NPA-free (upper) and NPA-bound receptors (lower). Transitions between arms were allowed only between the O1 states. This model was fitted to data obtained at 4 mM NPA (n = 6), and yielded the highest log likelihood. The rate constants (dark color) were fixed to the CTR values obtained as shown in Fig. 4 , whereas the rate constants in red were allowed to vary. The association rate constant (kon) and equilibrium dissociation constant (KD) are indicated. The glutamate binding steps used for simulation are shown (gray). (B) Macroscopic responses to a long (5-second) pulse of 1 mM Glu (black line) were simulated, and were compared with whole-cell currents in the absence (black) or presence (red) of 4 mM NPA preapplication (5 seconds).

To test this prediction, we measured glutamate responses from cells that were perfused with NPA-containing solution (4 mM) prior to glutamate application, but not during glutamate exposure. Results showed that NPA reduced NMDA receptor whole-cell currents even when applied exclusively before glutamate ( Fig. 5B , whole cell), thus strongly suggesting that: 1) NPA can bind to resting, glutamate-free NMDA receptors, and 2) that NPA remains bound and can encumber glutamate-elicited receptor activation even after excess NPA was washed out. Similarly, our results presented in Fig. 1 show that NPA can also bind to actively gating receptors, since NPA application inhibits steady-state currents produced by saturating concentrations of glutamate.

To gather all this information into a comprehensive model we postulated a mechanism where NPA can bind and dissociate from receptors at any point during gating but the kinetics of binding and dissociation will be different for each state. Based on the results presented above we aimed to limit the number of free parameters computed during statistical fittings, by further proposing that: 1) receptor conformations can be assigned to only two NPA affinity classes: low or high 2) the change in NPA affinity occurs simultaneously with the C2𠄼1 transition and 3) it is primarily due to an 𢏅-fold change in dissociation rate. With these assumptions and after fixing NPA-invariant rates to CTR values as described for the previous model, we fitted this extended model individually to single-channel records obtained with NPA. The resulting rate constants are given in Fig. 6A .

Kinetic mechanism of N1/N2A inhibition by NPA. (A) Tiered model with transitions allowed at all states except C4 and C5. Rate constants shown in blue are allowed to vary. The on-rates for all closed states were fixed for the values shown, while the off-rates from C1, O1, and O2 (blue) were scaled as 5-fold of those from C2 and C3 (green). This model was fitted to data obtained at 4 mM NPA (n = 6) and also used for simulations. The glutamate binding steps used for simulation are shown (gray). (B) Macroscopic responses to a long (5-second) pulse of 1 mM Glu (black line) were simulated, and agreed well with whole-cell currents as shown in Fig. 5B . (C) The simulated dose-inhibition curve (red) was compared with the whole cell data (IC50 is expressed as 95% confidence interval). Hill slope was 1.0 for simulations and 1.2 for experimental traces.

Experimental Validation for the Proposed Mechanism of NPA Action.

Next, we asked whether the model in Fig. 6A can predict correctly the observed inhibition of glutamate-elicited currents when applied to active receptors ( Fig. 1 ) or when applied to resting receptors ( Fig. 5B ). The model replicated faithfully the inhibition of steady-state currents illustrated in Fig. 1 , and predicted a dose-response curve that was indistinguishable from that calculated from experimentally recorded whole-cell currents (1.5 mM versus 1.9 ± 0.8 mM) ( Fig. 6C ). Further, currents simulated from NMDA receptor pre-exposed to NPA also reproduced successfully the macroscopic behaviors observed experimentally ( Fig. 5B ). Based on these results, we propose that a tiered model that includes two NPA affinities ( Fig. 6A ), although derived on the basis of several simplifying assumptions, captures the necessary detail to account for experimentally observed single-channel and macroscopic behaviors. Last, we used this model to predict how NPA can affect responses from synaptic and extra-synaptic receptors.

Predicted NPA Effects on Synaptic and Extrasynaptic Receptors.

We evaluated the effect of NPA on NMDA receptor synaptic responses by comparing traces simulated with the CTR and the tiered NPA model in Fig. 6B following brief (1 millisecond) pulses of glutamate (1 mM). The resulting traces showed that NPA (4 mM) would inhibit synaptic-like responses from N2A-containing receptors by decreasing the peak current amplitude by 2-fold and also by decreasing the deactivation time 𢏂-fold (400 milliseconds versus 200 milliseconds). Together these changes would produce �% decrease in the total charge transfer in response to a single synaptic event ( Fig. 7A ).

Predicted NPA effects on synaptic and extrasynaptic N1/N2A macroscopic responses. (A) Simulations of macroscopic responses to a brief (1-millisecond) pulse of 1 mM Glu (black arrow) (left), simulated traces are normalized to peak and are overlaid for comparison of deactivation time course (right). Red arrow indicates the peak of NMDA receptor response in the presence of 4 mM NPA. (B) Macroscopic responses to a long (5-second) pulse of 1 mM Glu (black line) were simulated with the models in Fig. 4B , and were recorded as whole-cell currents in the absence (black) or presence (red) of 4 mM ambient NPA (right) insets show the same traces normalized to peak for comparison of desensitization time course (left) and normalized to steady-state level for comparison of deactivation time course (right).

We mimicked the effects of NPA on extra-synaptic receptors with the tiered model in Fig. 6B by starting all receptors in glutamate-free states (C0) and setting the NPA concentration to 4 mM. Activation was initiated by switching the glutamate concentration from zero to 0.1 mM, after which receptors were allowed to equilibrate for 5 seconds across all 18 states. The results showed that 4 mM NPA would attenuate both the peak and the steady state currents and that the effect would be slightly larger on the steady state current ( Fig. 7B , simulation), but the time course for macroscopic desensitization would be largely unchanged (CTR versus NPA: 2.5 seconds versus 2.0 seconds) ( Fig. 7B , simulation, insert). Also, the results predicted 𢏂-fold faster deactivation in the presence of 4 mM NPA (CTR versus NPA: 430 milliseconds versus 190 milliseconds) ( Fig. 7B , simulation, insert). These results matched closely the whole-cell currents recorded with a similar protocol: for desensitization, the time constants for receptors in the presence and presence of 4 mM NPA are 1.1 ± 0.1 second and 1.2 ± 0.1 second (P = 0.54), respectively for deactivation, the time constants are 720 ± 50 millisecond and 350 ± 20 millisecond, respectively (P < 0.01) ( Fig. 7B ). Based on all the results presented we suggest that NPA would inhibit substantially both synaptic and extrasynaptic N2A receptors, with a larger effect on responses to brief synaptic-like stimuli.

Data Analysis

Recorded currents were analyzed as previously described in detail (Kussius et al. 2009). Microscopic recordings that contained more than one active channel, as indicated by simultaneous openings, were discarded, and the selected one-channel files were processed, idealized, and fitted to kinetic models in QUB. Occasional noise spikes were erased by setting the corresponding region to either a closed or an open current level baseline drift was corrected by resetting the baseline to the zero-current level. Preprocessed data were then idealized with the segmental k-means algorithm, at 150-μs resolution, and modeled with a maximum interval log-likelihood algorithm. The number of closed and open states in the final model was determined incrementally adding closed and open states to an initial 1C1O model (where C is closed and O is open) and setting the increase in log likelihood to an arbitrary threshold of 10 units/added state (two additional rate constants). Values calculated for kinetic parameters in each record were averaged for each data set and are reported as means ± SE.

Macroscopic traces (3–10 traces/cell) were averaged and evaluated by measuring peak current (Ipk) and equilibrium [steady state current (Iss)] amplitudes and by fitting a monoexponential function to the decay phase of the current (Clampfit 10.2 software, Molecular Devices). These analyses provided metrics for macroscopic desensitization as a time constant (τD) and as the Iss-to-Ipk ratio (Iss/Ipk). Values are reported as means ± SE and were obtained for 6–15 cells/condition.

Simulated macroscopic responses were generated in QUB as previously described (Kussius et al. 2009). The models derived from stationary single channel recordings were expanded to include two sequential glutamate-binding steps leading into the C3 state of each model glutamate association and dissociation rate constants were considered equal to those reported for recombinant N1/2A (Popescu et al. 2004) and N1/2B (Amico-Ruvio and Popescu 2010) receptors. Long (5 s) glutamate applications were modeled as instantaneous concentrations jumps between 0 and 1 mM glutamate.

The significance of observed differences between means was evaluated with a two-tail Student's t-test the significance of observed differences in variance was evaluated with a two-tailed F-test. Differences were considered significant for P values of <0.05.

GluN2A and GluN2B subunit-containing NMDA receptors in hippocampal plasticity

N-Methyl- d -aspartate receptor (NMDAR)-dependent synaptic plasticity is a strong candidate to mediate learning and memory processes that require the hippocampus. This plasticity is bidirectional, and how the same receptor can mediate opposite changes in synaptic weights remains a conundrum. It has been suggested that the NMDAR subunit composition could be involved. Specifically, one subunit composition of NMDARs would be responsible for the induction of long-term potentiation (LTP), whereas NMDARs with a different subunit composition would be engaged in the induction of long-term depression (LTD). Unfortunately, the results from studies that have investigated this hypothesis are contradictory, particularly in relation to LTD. Nevertheless, current evidence does suggest that the GluN2B subunit might be particularly important for plasticity and may make a synapse bidirectionally malleable. In particular, we conclude that the presence of GluN2B subunit-containing NMDARs at the postsynaptic density might be a necessary, though not a sufficient, condition for the strengthening of individual synapses. This is owing to the interaction of GluN2B with calcium/calmodulin-dependent protein kinase II (CaMKII) and is distinct from its contribution as an ion channel.

1. Introduction

Deciphering how neuronal networks can robustly store information after even a single learning episode has proved one of the biggest challenges in neuroscience. The best-supported cellular model for learning and memory proposes that pertinent neuronal activity leads to long-lasting changes in synaptic weights distributed throughout the network [1]. Such ‘synaptic plasticity’ has been extensively studied at hippocampal excitatory synapses, where most forms of plasticity require the activation of a particular type of glutamate receptor known as the N-methyl- d -aspartate receptor (NMDAR) for their induction [2]. NMDARs are particularly attractive as molecular mediators of plasticity because of their Ca 2+ permeability and also their coincidence detector properties that result from a voltage-dependent Mg 2+ block [3]. Plasticity encompasses both increases (long-term potentiation, LTP) and decreases (long-term depression, LTD) in synaptic strength and is expressed by changes in α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptors (AMPARs) at the postsynaptic element and/or changes in presynaptic transmitter release. Ultimately, such changes appear to be consolidated by structural alterations. Nevertheless, how relevant neuronal activity triggers meaningful plasticity in vivo remains uncertain. Moreover, if plasticity does support learning and memory, it is not yet clear whether a learning episode triggers both potentiation and depression in parallel across different sets of synapses or, alternatively, primarily one or the other.

This review addresses the controversial field around the hypothesis that NMDARs with distinct subunit compositions are responsible for the induction of the two opposing forms of plasticity, and that altering this subunit balance is an important mechanism to fine-tune synaptic strength. Firstly, it outlines how different subunit compositions affect the functional and dynamic properties of the NMDAR and the implications for its role in signal transmission. Secondly, it assesses whether these NMDAR subunits contribute to distinct forms of plasticity. Finally, as NMDAR-mediated plasticity is thought to be involved in learning and memory [4,5], it summarizes the evidence that different types of learning and memory might be supported by NMDARs with distinct subunit compositions. Several mechanisms might contribute to plasticity in different brain regions, and some of these are NMDAR independent, such as LTP at the mossy fibre synapse in CA3 [6] and metabotropic glutamate receptor-dependent LTD in CA1 [7]. Furthermore, plastic changes at the synapse can have both short- and long-term components [8], and these may be supported by distinct processes and have different NMDAR subunit-dependency for their induction [9]. This review focuses on the NMDAR-dependent forms of long-term plasticity in the hippocampus, in particular at the CA3-CA1 synapse, to facilitate comparison between experiments and because this synapse in particular has been implicated in hippocampus-dependent associative learning.

2. The hypothesis

A number of theories have been proposed to explain how different patterns of neuronal activity can lead to opposite effects on synaptic strength when both forms of plasticity depend on the same type of receptor. A unifying aspect of the postsynaptic-expression theories is that NMDAR-mediated Ca 2+ influx at the postsynaptic element, the ‘spine’, must ultimately be coupled to the different intracellular signalling cascades that mediate changes in synaptic weights.

It was quickly identified that the magnitude of Ca 2+ influx through NMDARs in the traditional high-frequency paradigms used to induce LTP was much greater than that during the low-frequency LTD paradigms. Therefore, it was proposed that different levels of Ca 2+ influx could couple to distinct intracellular signalling pathways to cause the molecular, and ultimately structural, changes at the synapse that underlie the two directions of plasticity [10]. This was also extended to the induction of plasticity by precise spike timing [11]. However, it seemed that a mechanism based on Ca 2+ levels alone may not be sufficiently robust, and it also could not account for some experimental observations thus, Ca 2+ time-course was suggested to be important [12,13]. Postsynaptic Ca 2+ dynamics are tightly regulated by small-conductance Ca 2+ -activated K + channels [14,15], voltage-dependent Ca 2+ channels [16,17] and intracellular Ca 2+ release [18] the Ca 2+ microdomains that these generate could provide a finer level of control in the regulation of bidirectional plasticity.

In parallel to refinements in the theory linking Ca 2+ dynamics and bidirectional plasticity arose the suggestion that the NMDAR itself could intrinsically dissociate synaptic strengthening and weakening in response to activity patterns. The distinct kinetic properties conferred by NMDAR subunits might allow the NMDAR composition to exert a finer level of control over the postsynaptic Ca 2+ dynamics. They also make distinct molecular associations, and thus could independently couple to the downstream kinase and phosphatase pathways that have been established to regulate each direction of plasticity (for review, see [19]). Furthermore, the relative abundance of these subunits at synaptic sites is tightly regulated throughout development, and a transition in subunit dominance correlates with changes in the ease of plasticity induction [20]. All this evidence converged on the attractive suggestion that the type of plasticity induced could be determined by the type of NMDAR subunit activated, perhaps by enabling transmission of distinct signals and tightly coupling them to different downstream signalling molecules. However, the findings from studies that set out to investigate this hypothesis have been inconclusive.

3. NMDA receptors

NMDARs are found both pre- and postsynaptically (figure 1a), and evidence from neocortical areas has set a precedent that these distinct populations of NMDARs may support different plasticity mechanisms [21–23]. This might also apply in the hippocampus, because plasticity can be mediated by changes in presynaptic release probability under some experimental conditions [24–26] for further discussion of presynaptically expressed hippocampal LTP, readers are directed to a recent review by Bliss & Collingridge [27]. Our review focuses on postsynaptic NMDARs, which can be found synaptically, perisynaptically and extrasynaptically. These three populations of receptors are recruited differentially by neuronal activity patterns, suggesting that they may play distinct functional roles however, it remains controversial as to how subunit composition and synaptic location relate.

Figure 1. NMDA receptor location and subunits in synaptic plasticity. (a) NMDARs are found both pre- and postsynaptically, and these two NMDAR populations might play different roles in synaptic plasticity. In the postsynaptic membrane, NMDARs are found synaptically, perisynaptically and extrasynaptically, where they are also likely to perform different functions. (b) During induction of spike timing-dependent LTP, Ca 2+ influx through GluN2B subunit-containing NMDARs (orange arrow) directly activates CaMKII to trigger LTP. Tetanic activation elicits a larger Ca 2+ influx through GluN2A subunit-containing NMDARs (grey arrows), which reaches and activates CaMKII anchored at the postsynaptic density (PSD) by the C-terminal of the GluN2B subunit. In both cases, it is CaMKII activation that triggers downstream signalling cascades mediating LTP expression, suggesting that the presence of the GluN2B subunit at the PSD is important for LTP induction irrespective of whether it supports a majority of the Ca 2+ influx.

NMDARs consist of two obligatory GluN1 subunits and two additional GluN2 or GluN3 subunits that confer the particular properties of the receptor. Each subunit consists of four major domains: the N-terminal domain containing binding sites for allosteric modulators, such as Zn 2+ and ifenprodil the agonist-binding domain, where the binding sites for glycine/ d -serine (on GluN1) and glutamate (on GluN2) are located and where competitive antagonists act the pore domain, accessible to pore blockers, such as phenycyclidine (PCP) and MK801, and, lastly, the C-terminal domain (CTD), which binds to different intracellular mediators. The two predominant GluN2 subunits in the hippocampus are GluN2A and GluN2B, although GluN2C and GluN2D are also present, in particular early in development, but also in low quantities in adulthood [28]. Hippocampal NMDARs can be diheteromeric (GluN1/GluN2A and GluN1/GluN2B) or triheteromeric (GluN1/GluN2A/GluN2B). The expression of GluN2B is high at birth but decreases into adulthood, while GluN2A expression increases with age [20,28–30]. The triheteromeric population is increasingly recognized to form a large proportion of the synaptic NMDARs in the adult brain [9,31–33], but, as these receptors cannot be selectively interrogated pharmacologically, the characteristics of triheteromeric receptors can only be inferred. Thus, their precise role remains enigmatic, though likely important.

(a) Functional and dynamic properties

Measurements of channel currents or postsynaptic responses to single stimuli have revealed notable differences in the kinetics of GluN2A and GluN2B subunit-containing NMDARs, and this is one way in which these channel subtypes could influence the induction of plasticity. Single-channel recordings in human embryonic kidney (HEK) cells showed that GluN1/GluN2A receptors have a higher probability of opening in response to glutamate and also a higher peak open probability than GluN1/GluN2B receptors [34]. These findings have been supported by whole-cell measurements in HEK cells [35] and also acute hippocampal slices [31] where channel open probability was estimated by use-dependent block of NMDARs with MK801. The faster activation and deactivation rates of individual NMDARs comprising GluN1/GluN2A result in whole-cell currents that rise and decay more quickly than those supported by GluN1/GluN2B, and a triheteromeric population apparently has an intermediate decay time constant [36].

The response of NMDARs to multiple stimuli at a range of frequencies over different durations is particularly important in the context of synaptic plasticity. Measurements from single-channel recordings have been used to simulate the responses of GluN2A and GluN2B subunit-containing NMDARs to trains of presynaptic activity and this showed a relationship between stimulation frequency and total charge transfer that differed between the two subunits [34]. It was found that at the very low frequencies (0.1–0.3 Hz) used in some plasticity protocols, GluN1/GluN2B channels supported twice as much charge transfer owing to their slower deactivation rate. At the 1 Hz stimulation favoured for LTD induction, charge transfer through GluN1/GluN2B channels still exceeded that through GluN1/GluN2A, but to a lesser extent, and it was equivalent between channels at 2 Hz. Modelling also revealed that the duration of activity is important at high frequencies of stimulation for the 100 Hz frequency often used in high-frequency induction protocols, the duration of the stimulus train dramatically changed the charge transfer through GluN2A and GluN2B subunit-containing NMDARs. The two channel subtypes permitted equivalent charge transfer for 100 ms of stimulation however, for a 1000 ms stimulus train, the typical duration used to induce LTP, GluN1/GluN2A supported twice the charge transfer of GluN1/GluN2B [34]. Desensitization of GluN2A and GluN2B subunit-containing NMDARs will also contribute to their ability to mediate charge transfer, but the effect will be highly influenced by the subunit composition at the postsynaptic spine modelling has suggested that a larger fraction of GluN2A subunit-containing NMDARs will be desensitized at stimulation frequencies from 5 to 100 Hz, but that the difference between GluN2A and GluN2B will be the lowest at 100 Hz for 1000 ms [37]. Although these modelling studies provide useful information predicting how subunit composition and plasticity protocols may interact, further experimental investigation is essential to test their predictions under physiological conditions.

It is generally assumed that the magnitude of Ca 2+ influx is related to charge transfer, which suggests that differences in charge transfer between channel subtypes would be important in their ability to trigger distinct forms of plasticity that require different postsynaptic Ca 2+ dynamics. However, the relationship may be more complex, because Ca 2+ imaging of spines has revealed that the magnitude of Ca 2+ influx in response to glutamate application does not necessarily correlate with NMDAR-mediated charge transfer [38]. Moreover, a GluN2B subunit-selective antagonist caused a greater reduction in Ca 2+ influx in those spines that supported a higher change in Ca 2+ per uncaging-evoked excitatory postsynaptic current (EPSC) which led the authors to conclude that GluN2B subunit-containing NMDARs support a greater Ca 2+ influx per unit of current [38]. This had previously not been observed in recombinant GluN2A or GluN2B subunit-containing NMDARs in heterologous systems [28], which may be owing to methodological differences, the presence of triheteromeric NMDARs and/or posttranslational modifications of NMDARs in acute brain slices, and so further investigation is required. Nevertheless, if this higher Ca 2+ influx through GluN2B subunit-containing NMDARs holds, then it suggests that this subunit could exert a disproportionate influence on plasticity.

(b) Intracellular associations of NMDA receptors

Other ways in which NMDAR subunits could influence the direction of plasticity relate to their long CTDs, as these allow for many intracellular interactions and modifications by phosphorylation. Many of these direct and indirect molecular associations are unique, as the CTD shows a high level of sequence divergence between GluN2A and GluN2B. This permits separate regulation of their presence or absence at the synapse and could also couple each subunit to different downstream signalling cascades that support either LTP or LTD a few key examples of these interactions with relevance to plasticity are outlined below.

The presence of GluN2A and GluN2B subunit-containing NMDARs at the synapse is stabilized by their interaction with the postsynaptic density (PSD) subfamily of membrane-associated guanylate kinases, which includes PSD-95, PSD-93, SAP102 and SAP97 [39]. Posttranslational modifications can alter how strongly GluN2A and GluN2B subunits bind to these PSD proteins, and this is one mechanism that regulates the availability of these subunits at the synapse: in turn this could influence the induction of plasticity. For example, Cdk2 phosphorylates GluN2B at Ser1480, which disrupts its interaction with PSD-95 and SAP102 and leads to a reduction in synaptic GluN2B [40,41]. Not only is this phosphorylation event unique to GluN2B [41], it is also regulated by synaptic activity [40]. Surface levels of GluN2B can also be regulated by phosphorylation at Tyr1472 by Fyn and Src phosphorylation prevents the clathrin adaptor protein AP-2 binding to the GluN2B internalization motif YEKL, and thus inhibits endocytosis. Cdk5 is likely to be an additional component in this regulatory cascade, because inhibiting it encourages the interaction between Src and PSD-95, and consequently increases Tyr1472 phosphorylation [42]. Another example is the endocytosis of GluN2A, which is controlled by a dileucine motif at a different site (Leu1319 and Leu1320) [43], and thus its surface expression can be regulated independently. It is notable that GluN2B subunits also undergo more frequent endocytosis than GluN2A, and that these two subunits enter into different intracellular pathways, with GluN2B entering into recycling endosomes and GluN2A into late endosomes [43]. This suggests that activity- or neuromodulatory-dependent regulation of GluN2B may be a quicker or more sensitive way to alter the state of a synapse.

In addition to separate regulatory pathways to control the synaptic GluN2A and GluN2B levels, unique associations with different enzymes may also enable the subunits to contribute to opposing forms of plasticity. The most important example is probably the high affinity binding between the GluN2B CTD and the catalytic domain of Ca 2+ /calmodulin-dependent protein kinase II (CaMKII) [44]. There is a basal level of CaMKII present at the PSD bound to NMDARs [45], but this is supplemented by active, phosphorylated CaMKII that translocates to the PSD following the induction of LTP [46,47]. The interaction between CaMKII and GluN2B anchors CaMKII at the synapse in its active conformation [48]. LTP requires active CaMKII (reviewed in [49]), more specifically, the association between GluN2B and active CaMKII [50,51]. Therefore, the ability of GluN2B to mediate tight coupling between Ca 2+ influx and CaMKII, and maintain extra active CaMKII in the vicinity of its substrates, such as AMPARs, to initiate the phosphorylation events that support synaptic strengthening, suggests that this subunit may play a crucial role in LTP. Another notable protein interaction at the postsynaptic spine is that between GluN2B and Ras-GRF1, a Ca 2+ /calmodulin-sensitive, Ras-specific guanosine diphosphate/guanosine triphosphate exchange factor that has been implicated in synaptic plasticity [52,53]. There are also proteins that associate indirectly with NMDARs but still retain a subunit preference, permitting an extra level of complexity in the modulation of NMDAR function by its subunit composition.

4. GluN2A and GluN2B NMDA receptor subunits in plasticity

As illustrated thus far, there are three main ways in which differential subunit composition could affect the type of plasticity induced. Firstly, the different channel properties conferred by these two subunits could encourage sufficiently distinct postsynaptic spine Ca 2+ dynamics to trigger separate downstream plasticity pathways. Secondly, the unique associations made by the CTDs of GluN2A and GluN2B subunits provide a direct mechanism by which the subunits could couple to independent intracellular signalling cascades, and thus play different roles in synaptic plasticity. Finally, changes in the quantity or subcellular location of these subunits could alter the basal state of a spine, and thus affect the future induction of plasticity. Given these possibilities, it is interesting to ask to what extent experimental evidence supports the hypothesis that different NMDAR subunit compositions determine the outcome of a plasticity protocol.

(a) Synaptic plasticity protocols

A wide variety of induction protocols can be used to induce synaptic changes in the hippocampus and the main paradigms are illustrated in table 1. There are considerable differences in the activity patterns used to induce even one direction of plasticity therefore, it is possible that the synaptic changes induced by such diverse paradigms have distinct mechanistic underpinnings. It is important to consider this when delineating the NMDAR subunits involved in plasticity, as results are only truly comparable when the same protocol is applied.

Table 1. Different protocols used to induce LTP and LTD.

Various plasticity protocols have been used in combination with pharmacological and genetic manipulations to investigate the potential role of different NMDAR subunits in LTP and LTD. The first investigation into the roles of distinct NMDAR subpopulations in bidirectional plasticity compared the possible contributions of GluN2A/B and GluN2C/D subunits using broad-spectrum antagonists [54,55]. Subsequent studies have focused on dissecting whether differences exist between GluN2A and GluN2B, as these are the predominant subunits present at the juvenile and adult CA3-CA1 synapse. There are currently highly selective pharmacological tools to block GluN1/GluN2B receptors, primarily ifenprodil and its more potent derivative Ro 25-6981 [56]. Nevertheless, these are activity-dependent blockers, and the agonist concentration affects the extent to which these antagonists inhibit GluN2B subunit-containing NMDARs for example, 3 µM Ro 25-6981 achieves 93% inhibition in the presence of 100 µM NMDA but only 62% in 10 µM NMDA [57]. The study of GluN2A is even more limited by the lack of selective antagonists for this subunit. NVP-AAM077 (hereafter referred to as NVP) is commonly used as a GluN2A-preferring antagonist, but it is now known to be less selective in rodents than had been assumed from recombinant human proteins and it actually shows only a 10-fold preference for GluN2A over GluN2B [56,58]. Therefore, data collected using NVP invariably have some block of GluN2B as well, but the magnitude of this block varies considerably with the concentration used.

Pharmacological and genetic manipulations allow the investigation of different aspects of NMDAR function. Pharmacological approaches are informative about the role of the NMDAR as an ion channel and do not directly address the importance of intracellular associations. They also will indiscriminately affect a particular NMDAR subtype on all cell types, for example interneurons and pyramidal neurons, where they may play different roles. In turn, this could have repercussions for the network that manifest at a circuit as well as at a behavioural level [59]. By contrast, constitutive and inducible genetic knock-outs provide information about the combined role of the NMDAR channel properties and synaptic presence, and can be directed to a specific cell type however, given the changes in subtype composition throughout development, and even on the short timescales of metaplasticity, perturbing the balance of receptors is likely to cause other changes in the network or to recruit compensatory processes in the long term. In addition, pharmacological manipulations have often been performed in rats, whilst almost all the genetic manipulations discussed below have used mice, and species differences have not been excluded.

(b) Long-term potentiation

An early study often cited to argue for the differential involvement of NMDAR subunits in the induction of LTP and LTD found that pharmacological block of GluN2A by NVP prevented tetanus- and pairing-induced LTP in three- to four-week rats but that antagonism of GluN2B by ifenprodil/Ro 25-6981 did not, but did instead block LTD induction [60]. Another study, using the same pharmacological concentrations and age of rats, also found that NVP completely blocked tetanus-induced LTP, though in this study ifenprodil and Ro 25-6981 partially blocked this form of LTP [61]. However, the data in both these studies were collected using a relatively high NVP concentration thus, in addition to the aforementioned lack of NVP specificity, the resultant NMDAR-mediated current block was much greater with NVP (reduction of current: 53 ± 3.0% in [60] and 81 ± 3% in [61]) compared to the small reductions caused by Ro 25-6981 (36 ± 5% in [60] and 32 ± 3% in [61]). Therefore, this apparently selective role of the GluN2A subunit might instead arise from a threshold effect, whereby inhibiting more total NMDAR-mediated current could block LTP, especially given that a high level of Ca 2+ influx is traditionally thought to be required for LTP induction. Indeed, when a study in mice controlled for this by selecting antagonist concentrations to reduce NMDAR-mediated current equally (40% reduction to match that possible by Ro 25-6981), pairing-induced LTP was not impaired irrespective of the antagonist used (NVP, Ro 25-6981 or the broad-spectrum NMDAR antagonist 2-amino-5-phosphonopentanoate (AP5)) [62]. The same antagonist concentrations, as titrated for their NMDAR current reduction in the aforementioned study, also did not impair LTP induced by a theta-burst paradigm in adult mice [63], though this low NVP concentration did partially block tetanus-induced LTP, while ifenprodil did not [64]. This suggests that tetanus-induced LTP is the most sensitive to GluN2A block. In addition to the extent of current reduction, the developmental stage is also likely to account for some of the different results reported. In a study using two-week-old rats, LTP was impaired by both GluN2A and GluN2B antagonism [65], probably because the developmental transition in subunit composition is not yet complete at this age in rats, as GluN2B antagonism reduced LTP in two-week-old rats but not in those over six weeks [66]. Some pharmacological studies have also been conducted in vivo, where, although it is more difficult to determine the exact concentrations that the CA3-CA1 synapse is exposed to, the variability introduced by different slicing angles, techniques and incubation solutions is avoided. Intrahippocampal infusions of Ro 25-6981 or NVP blocked tetanus-induced LTP in vivo in four- to six-week-old rats [67], whereas, when drugs were delivered intraperitoneally, only NVP blocked LTP with Ro 25-6981 having no effect [67,68] this apparent discrepancy is likely to result from different levels of NMDAR block. Therefore, even in vivo studies have not yielded a clear conclusion.

As these pharmacological studies have provided little convincing evidence for a highly selective role of either subunit in LTP induction, an alternative hypothesis should be considered, whereby either subunit can support LTP provided they allow sufficient Ca 2+ entry but the subunit that mediates most of the Ca 2+ influx will have a greater importance. This subunit bias will be affected by neuronal activity patterns in behaving animals, whilst the induction protocol will influence the subunit bias in synaptic plasticity studies. The charge transfer modelling of Erreger et al. [34] suggests that the frequency of stimulation affects which subunit predominates specifically, GluN2A should carry a majority of the current in tetanically induced LTP, whilst GluN2B might carry more current in lower frequency protocols [34]. Indeed, Ro 25-6981 blocked the induction of spike-timing-dependent LTP (t-LTP) in acute hippocampal slices from adult mice [69,70], which is induced by a low-frequency paradigm, whereas it did not impair tetanus-induced LTP [70]. However, complicating this interpretation, Zhang et al. [70] found, albeit with a high and thus less selective concentration, that NVP also blocked t-LTP. Furthermore, Gerkin et al. [71] found that an equivalent concentration of Ro 25-6981 did not block t-LTP in cultured hippocampal neurons, but NVP did performing pre–post pairings at 1 Hz, rather than 0.1–0.2 Hz, and the use of cultured neurons, might account for this different result. Other studies have provided further evidence for a subunit bias owing to charge transfer. For example, a pairing protocol should limit the importance of subunit kinetics, as the postsynaptic neuron is held at a depolarized potential to remove the Mg 2+ block. Therefore, neither subunit would be predicted to dominate, and any bias would rather be determined by the numbers of each subunit present at the synapse. Indeed, Berberich et al. [72] found that pairing-induced LTP was not blocked by NVP, Ro 25-6891 or AP5 when used at sufficiently low concentrations to have a minimal impact on charge transfer during induction however, when antagonist concentrations were increased to a level that significantly reduced charge transfer, both GluN2A and GluN2B antagonists caused an equivalent reduction in the magnitude of LTP. It is also important to consider the complication introduced by the large triheteromeric population of NMDARs a recent study found that only high concentrations of Ro 25-6981 could impair theta-burst-induced LTP, possibly because the antagonist starts inhibiting these triheteromers [9].

Overall, therefore, the conclusion from the current pharmacological data, with the caveat that no subunit-selective GluN2A antagonist exists, is that either subunit is capable of supporting LTP induction, provided they can mediate sufficient charge transfer (and associated Ca 2+ influx). Which subunit mediates most charge transfer will be influenced by age, because the relative abundance of these two subunits changes over development [20,28,30]. Within a given age range, the induction protocol used may bias which subunit makes a greater contribution because of the different kinetics of the two subunits. Specifically, at the two extremes of frequency, t-LTP might have a greater contribution from GluN2B subunit-containing NMDARs, while tetanus-induced LTP could have a bigger contribution from GluN2A subunit-containing NMDARs [34] the greater ability of GluN2A subunit-containing NMDARs to drive the induction of tetanic LTP has received additional support from modelling work [73]. The potential importance of stimulation frequency during induction means that, in order to fully explore the role of GluN2A and GluN2B in plasticity, it is vital to further investigate the types of activity patterns that drive synaptic changes in behaving animals.

However, pharmacological antagonists only block the ligand binding and channel properties of the NMDAR, and therefore pharmacological studies do not directly address whether the physical presence of NMDAR subunits at the synapse could play a role additional to that in mediating Ca 2+ influx. The most important other function in plasticity is likely to be related to the direct and indirect intracellular associations made by the GluN2 CTD. Genetic manipulations that remove or alter the GluN2 subunits may shed some light on whether subunit selectivity in LTP induction could arise through these interactions of particular interest is whether the GluN2B subunit plays a crucial role owing to its association with CaMKII.

A conditional genetic knock-out of GluN2B subunits in the CA3 subfield abolished tetanus-induced LTP at the commissural-CA3 and commissural/associational-CA3 synapses [74]. In mice lacking GluN2B in the CA1 and neocortex, LTP induced by two tetani was impaired, but this deficit could be overcome by giving multiple tetani [75]. A smaller effect on LTP was seen in another forebrain-specific GluN2B knock-out mouse, where LTP was deficient only when induced by a pairing protocol and not by a stronger protocol using two tetani [76]. Other genetic manipulations that indirectly affect the levels of GluN2B have also shown considerable LTP impairments. For example, kinesin-like protein KIF17 transports GluN2B subunit-containing NMDARs to the synapse, and KIF17 knock-out mice [77], or mice carrying mutations in KIF17 that disrupt its loading or unloading of GluN2B [78], show reduced synaptic GluN2B and also abolished (KIF17 knock-out) or impaired (KIF17 mutant) LTP induced by a single tetanus. However, when interpreting results from genetic manipulations as evidence for a role of GluN2B in LTP, it must be remembered that such long-term manipulations could trigger compensatory changes at the cellular level, such as increased recruitment of CaMKII by other PSD components, or even at the network level, through changes in inhibition, for example. Of particular note, GluN2B knock-out lines also show small reductions in GluN2A [77,78], GluN1 [74] or spine density [74,75]. Given that GluN2B subunits must therefore contribute to the regulation of NMDAR levels and postsynaptic structural integrity, it is difficult to attribute fully the LTP impairments found following chronic reduction of GluN2B to a unique structural/interaction role of the GluN2B CTD in LTP.

Therefore, it is interesting that shorter term manipulations, despite only reducing the GluN2B content, have produced stronger effects than complete genetic knockouts. For example, GluN2B knockdown with RNA interference (RNAi) caused a profound reduction in the magnitude of LTP induced in two-month-old rats even though a very strong LTP protocol (four tetani) was used [79]. LTP was also abolished in slices incubated for 2 h with a membrane-permeable GluN2B C-terminal peptide that disrupted the GluN2B–PSD interaction and caused a consequent decrease in synaptic GluN2B content [80]. Nevertheless, as mice with reduced synaptic GluN2A [81] or GluN2A knock-out [82] also exhibit reduced LTP, which can be recovered by a stronger induction protocol with multiple tetani [83], these experiments alone do not provide convincing evidence of a unique role of the intracellular interactions made by GluN2B.

The converse experiments, using subunit-selective overexpression, show that LTP is enhanced in adult [84] and aged [85] mice and adult rats [86] by increasing GluN2B. Cdk5 knock-out mice with elevated levels of synaptic GluN2B also show enhanced LTP [87]. By contrast, GluN2A-overexpressing mice do not show increased LTP [88], suggesting that GluN2B may play a unique role. However, elevated LTP was observed in dysbindin knockout mice that had enhanced GluN2A-mediated currents [89] thus, it is not completely clear whether the enhanced LTP in GluN2B-overexpressing rodents arises from higher levels of Ca 2+ influx owing to the increased NMDAR-mediated current or because of an important contribution of the GluN2B subunit itself. Therefore, data from knock-out and overexpression studies suggest that the GluN2B subunit is important for the induction of LTP, but do not convincingly distinguish between a unique role of GluN2B at the synapse or a general threshold effect. In the latter case, the reduction or increase in synaptic NMDAR content produced by these genetic manipulations would alter the magnitude of LTP simply because of non-specific changes in the amount of Ca 2+ influx during induction.

Rather than attempt to demonstrate a unique role for GluN2B by excluding this threshold possibility, studies have instead focused on manipulating the distinguishing features of the GluN2B subunit, in particular its intracellular interactions. The most convincing evidence for a uniquely important role for the interaction between GluN2B and CaMKII in LTP comes from experiments directly perturbing this association. The first data came from work by Barria & Malinow [50] they showed that LTP was blocked in organotypic hippocampal slices transfected with a GluN2B construct impaired in binding to CaMKII. Moreover, the LTP impairment caused by transfecting recombinant GluN2A, which drives a switch in synaptic NMDAR content from GluN2B to GluN2A, could be ameliorated by transfecting a mutated form of GluN2A with an enhanced ability to bind CaMKII [50]. This further suggests that an important function of wild-type GluN2B is to maintain sufficient CaMKII at the synapse. These findings have since been extended to acute slices an inducible mutation that weakens the GluN2B–CaMKII interaction impaired LTP induced by both a tetanic and 10 Hz protocol [51], while a knock-in mouse with two point mutations that decrease the GluN2B–CaMKII interaction also showed a large reduction in two-tetani-induced LTP [90]. This interaction is regulated by both components as inhibiting CaMKII, which initially disrupts the GluN2B–CaMKII association, leads to downregulation of synaptic GluN2B content within 2 h and a concomitant impairment in LTP [80]. Inhibition of CaMKII was removed, and hence CaMKII kinase activity restored, before LTP was induced, and therefore the LTP deficit observed by Gardoni et al. [80] could be attributed to an altered synaptic subunit composition.

One reason why the GluN2B–CaMKII interaction is so important for LTP might be its effect on AMPARs. Phosphorylation of Ser831 on GluA1 may be necessary for AMPAR insertion, and this phosphorylation does not occur when the GluN2B–CaMKII interaction is blocked [90], which may hinder synaptic strengthening. Thus, through its strong affinity for CaMKII, it does seem that GluN2B may play a special role at the synapse additional to its function as an ion channel. This was particularly clear when, at an age when pharmacological inhibition of GluN2B with Ro 25-6981 no longer blocked pairing-induced LTP, LTP was impaired by RNAi knock-down of GluN2B [91]. This difference seems directly related to a unique role of the GluN2B CTD as, following RNAi knock-down of GluN2B and the resultant LTP impairment, LTP could be restored by an RNAi-resistant GluN2B or a chimaera of GluN2A with the GluN2B tail, but not a chimaera of GluN2B with the GluN2A tail [91]. Thus, the physical presence of GluN2B may be important irrespective of its contribution to Ca 2+ influx. This explains how GluN2B could have a unique and fundamental role in LTP induction, despite pharmacological block of GluN2B not impairing the LTP induced by certain paradigms.

However, this does not mean that the GluN2A CTD makes no contribution to LTP. It has been shown that GluN2A subunit-containing NMDARs alone can trigger a form of LTP dependent on the Ras-GRF2/Erk Map Kinase pathway [52], suggesting that unique interactions may occur. Indeed, a mouse line with a GluN2A C-terminal truncation showed impaired tetanus-induced LTP, with the remaining LTP supported by GluN2B-subunit-containing NMDARs [66,92]. However, as the truncation mutation also caused an overall reduction in GluN2A levels in this mouse line, as well as a possible switch to a pure triheteromeric NMDAR population, it is possible that the LTP impairment arose from a threshold effect from reduced Ca 2+ influx rather than a unique role of the GluN2A CTD. Supporting the threshold interpretation, stronger induction protocols expected to promote higher levels of Ca 2+ influx could overcome the LTP deficits in mice with a GluN2A C-terminal truncation both in this [66] and another study [93]. Thus, the evidence for a privileged role of GluN2A in LTP induction is weaker.

One vital consideration when investigating a selective role of NMDAR subunits in LTP is that CA3-CA1 synapses are not a uniform population. Changes in the extracellular field or whole-cell response to a plasticity protocol are, therefore, the summed changes from a heterogeneous population of synapses that may respond differently to the same activity pattern. Structural imaging and molecular labelling studies have revealed that, even in the adult brain, there are different categories of postsynaptic spine shape that also have distinct receptor signatures [94]. Imaging of individual spines has shown that GluN2B antagonists selectively reduce glutamate uncaging-evoked EPSCs and Ca 2+ transients in small spines [38] indicating that GluN2B is concentrated in such spines. This is of particular interest because chronic imaging experiments have revealed that, whilst the volume of all spines increases following an LTP induction protocol, this expansion is only preserved in spines initially classified as small, but is transient in others [95]. This suggests that these distinct types of spines, with their different NMDAR subunit compositions, may engage differently in plasticity processes, and thus could play independent roles even in the mature brain. Therefore, using the variability in synapse structure and composition may be a fruitful additional approach to study the role of NMDAR subunits in LTP and information processing in general. Electron microscopy and immunolabelling have revealed that these different spine populations are asymmetrically distributed in the adult mouse brain specifically, postsynaptic CA1 spines receiving input from the left CA3 are mainly small and rich in GluN2B subunit-containing NMDARs, whereas those receiving input from the right CA3 tend to be larger and richer in AMPARs [94,96]. This enabled these different types of spine, with their different GluN2B content, to be targeted optogenetically to test the hypothesis that GluN2B subunits are particularly important for LTP induction. By injecting channelrhodopsin-2 (ChR2) under the control of a Cre-dependent promoter into either the left or right CA3 of CaMKII-Cre mice, the left or right excitatory CA3 input onto CA1 could be selectively recruited. It was found that only left CA3-CA1 synapses showed LTP induced by spike timing-dependent or theta-based protocols [69]. Moreover, this was likely a result of the higher GluN2B content of spines receiving left CA3 projections, as Ro 25-6981 caused a greater reduction in NMDAR-mediated current in left-injected mice and also blocked the observed plasticity [69]. Although some GluN2B was present at the right CA3-CA1 synapses, they appeared incapable of expressing LTP an explanation is that they are already at the maximum possible synaptic strength that can be maintained. It will be interesting to test whether a similar asymmetry is found following stronger induction paradigms such as tetanus-induced LTP, though current limitations of ChR2 kinetics in driving pyramidal neurons at a sufficiently high frequency preclude direct testing of this.

In summary (figure 1b), the most parsimonious conclusion from the available evidence is that both GluN2A and GluN2B subunits can support the requisite Ca 2+ influx to induce LTP. However, which subunit predominates in mediating this Ca 2+ influx, and thus contributes more to LTP induction, depends on the protocol and also developmental stage, as the latter affects subunit abundance at the postsynaptic spine. Provided the NMDARs present can support sufficient Ca 2+ influx, the GluN2B subunit seems to play a unique additional role because of its association with CaMKII. This interaction targets CaMKII to the vicinity of Ca 2+ influx at the PSD and also provides a binding site for the additional activated CaMKII recruited following an LTP induction protocol. Finally, GluN2B helps maintain CaMKII activation and localize it at the PSD in the vicinity of its cellular substrates enabling it to trigger the downstream pathways that mediate the expression of LTP.

(c) Long-term depression

Given the possible bias towards an important contribution of the GluN2B subunit to LTP induction, an opposite bias might have been expected for NMDAR-dependent LTD. However, studies into the NMDAR subunit dependence of LTD have produced even more contradictory results than those for LTP, making it hard to draw unifying conclusions. The different types of findings are illustrated below and possible explanations for the controversies are offered. One major complication in this field is the strong dependence of LTD induction on developmental age, because small changes in age affect the magnitude or even presence of LTD [97]. Nevertheless, this may provide information about the mechanisms involved, especially given the changing pattern of subunit expression across development [20,30]. Very few studies have addressed subunit involvement in spike timing-dependent LTD in the hippocampus, and therefore the following discussion will focus on low-frequency-induced LTD. Nevertheless, groups investigating this form of LTD have often drawn opposing conclusions about subunit-specificity, despite using very similar experimental preparations. For example, some pharmacology-based studies have found that a concentration of NVP that blocked LTP did not impair LTD in culture [71], in acute slices [60] or in vivo [68], suggesting that GluN2A subunits are not required for LTD induction. By contrast, other studies have found that NVP concentrations that impaired LTP also blocked LTD [61,65]. It is possible that NVP causes a substantial reduction in NMDAR-mediated current that prevents LTP and, depending on the precise concentrations used, sufficient Ca 2+ influx remains to trigger the downstream molecular cascades that mediate LTD. The importance of antagonist concentration in determining experimental conclusions is illustrated by the findings of Fox et al. [67], where, of two NVP concentrations administered intraperitoneally that blocked LTP in vivo, only the higher concentration blocked LTD. Overall, therefore, the extent to which GluN2A subunits are engaged in LTD induction under physiological circumstances is not yet clear, but there certainly is no pharmacological evidence that they play a mandatory or unique role.

There has been a particular focus on a possible role of GluN2B subunit-containing NMDARs in LTD. Some groups find that the GluN2B-specific antagonists Ro 25-6981 and ifenprodil block LTD in acute slices [60,64], in culture [71] and in vivo [67,68], even when a larger NMDAR-mediated current reduction by NVP has no effect [60,68,71]. However, as with the contradictory findings reported with GluN2A antagonism, many other studies have found that the same manipulation has no effect on LTD induction in acute slices, despite using equivalent or higher concentrations of GluN2B antagonists [61,65,98,99]. The slice orientation used may partially explain the discordance between these findings, since Bartlett et al. [100] found that coronal slices (as used in [60]) had a form of LTD that was sensitive to Ro 25-6981, whereas sagittal slices (as used in [65]) had a GluN2B-independent form of LTD, likely owing to the stronger preservation of cholinergic projections with a sagittal cutting angle. Indeed, when they blocked muscarinic acetylcholine receptors, it unmasked a GluN2B-dependent LTD [100]. The contradictory results found in slices cut in the transverse angle, where studies have found either GluN2B-dependence [53,64] or GluN2B-independence [98,99] of LTD, might result from subtle variations in the integrity of cholinergic fibres.

One further factor contributing to the variability in the sensitivity of LTD to GluN2B antagonism may be the extent to which extrasynaptic GluN2B subunit-containing NMDARs are recruited. The traditional low-frequency induction protocol is sometimes combined with application of the glutamate-uptake inhibitor threo-β-hydroxyaspartate (TBOA), which would likely increase activation of the extrasynaptic receptor population. Contrasting effects have been reported following TBOA application: it has not affected GluN2B involvement in LTD [65], it has produced a GluN2B-dependent LTD [101], or it has even decreased the magnitude of LTD [98]. The last finding suggests that extrasynaptic, possibly GluN2B subunit-containing, NMDARs could exert a negative modulatory role on LTD under some experimental conditions. This suggestion is supported by a study finding that ifenprodil block of GluN2B actually enhanced the magnitude of LTD [102]. However, it cannot be excluded that the discrepancy between these findings results from the complex mode of action of the non-competitive GluN2B antagonists used, as the extent of their block varies with agonist concentration [57], and such concentrations are likely more variable when glutamate uptake is blocked. Overall, therefore, the contradictory findings mean the precise nature of the role played by GluN2B subunit-containing NMDARs in LTD is anything but clear, and the induction of LTD may be particularly sensitive to experimental conditions.

Transgenic approaches have also not provided conclusive support for a unique role of either subunit in LTD and, in fact, most of the findings have been negative. A reduction in synaptic GluN2B levels, achieved by disrupting the association of GluN2B with PSD-95 in an acute manner, did not alter LTD, despite being sufficient to impair LTP [80]. The GluN2A knock-out mouse also showed no LTD impairment, though the same study also found no reduction in LTP [103]. Furthermore, overexpression of GluN2B did not affect the magnitude of traditional 1 Hz stimulation-induced LTD [86], and neither did direct [88] or indirect [89] GluN2A overexpression, although GluN2A overexpression did reduce LTD induced by 3 or 5 Hz stimulation without affecting LTP [88]. The only clear positive findings have been that GluN2B knock-out mice [75] and KIF17 knock-out mice, which have a consequent reduction in synaptic GluN2B [77], both have impaired LTD. However, this is not evidence for a subunit-selective role of GluN2B in LTD because these manipulations also caused reduced LTP. Overall, there is no clear evidence from transgenic approaches that either subunit plays an irreplaceable role in LTD induction, though impairments have been more frequently associated with GluN2B-specific manipulations.

A final possibility is that one subunit makes a more important contribution to LTD through its unique intracellular interactions. This may be because it facilitates coupling of Ca 2+ influx to the downstream pathways that mediate a reduction in synaptic strength. Alternatively, NMDARs could play a different role in LTD, where their activation is still required to trigger downstream signalling cascades, but not to mediate Ca 2+ influx. Evidence for this latter proposal is the demonstration that NMDAR-dependent LTD could still be induced despite NMDAR-mediated ion flux being blocked [104]. This suggests that the basal level of Ca 2+ , rather than Ca 2+ influx through NMDARs, is necessary for LTD induction and may explain why no clear findings have emerged from studies using subunit-specific pharmacological antagonists. It also implies that any subunit-selective function of NMDARs would emerge solely through their unique intracellular associations. As predicted by the requirement for CaMKII in LTP induction alone, genetic disruption of the GluN2B–CaMKII association did not impair LTD [51]. However, other interactions have been identified that could be important. For example, LTD alone was impaired in mice with a genetic knock-out of the p75 neurotrophin receptor [105]. The specific mechanism responsible for this deficit was reported to be that pro-brain-derived neurotrophic factor could not activate the p75 neurotrophin receptor to enable LTD and, because the mouse line also had an overall reduction in GluN2B levels, it suggests a possible signalling role of GluN2B. Another important association might be that between Ras-GRF1 and the GluN2B CTD, as LTD is impaired in Ras-GRF1 knock-out mice [53]. The level of Ras-GRF1 may determine the type of depression induced under physiological conditions, as LTD is GluN2B-dependent when this association predominates [100]. Furthermore, acetylcholine concentration is inversely correlated with the level of the GluN2B–Ras-GRF1 interaction [100], suggesting that activity of the cholinergic system could modulate the subunit dependence of LTD induction.

In summary, the data are inconclusive, but neither the GluN2A nor the GluN2B subunit seems to play an obligatory role in the induction of LTD. Nevertheless, the GluN2B subunit has been implicated more frequently, perhaps because of its unique associations with components of pathways that modulate or mediate LTD.

5. Metaplasticity

The basal state of the synapse is likely to play a crucial role in determining the nature of its changes in response to a particular input. It has even been suggested that it affects whether the LTP expression mechanism is pre- or postsynaptic [24]. Given that there is no convincing evidence for complete subunit selectivity in the induction of either direction of plasticity, it is conceivable that GluN2A and GluN2B subunits play a subtler role, biasing the synapse towards the induction of either LTP or LTD. Thus, it has been investigated whether altering the relative or absolute levels of either subunit changes the basal state of a synapse sufficiently to influence the future induction of plasticity.

Activity-dependent changes in the patterns of subunit expression take place in the forebrain during development [20,30], and even a short history of neuronal activity can change the subunit balance at the CA3-CA1 synapse in young rodents [106–108]. To investigate whether changes in subunit composition alter the response of the synapse to further activity, Xu et al. [109] used 600 low-intensity pulses at varying frequencies to ‘prime’ the synapse this priming was insufficient to change synaptic weights but did alter the GluN2A/GluN2B ratio. Following this priming, different plasticity protocols were applied. The magnitude of LTP was enhanced, and LTD reduced, by low-frequency priming that decreased the GluN2A/GluN2B ratio. By contrast, an elevated LTD and suppressed LTP was seen if the synapse had been primed with high-frequency stimulation that increased the GluN2A/GluN2B ratio. They also used a plasticity protocol that did not change synaptic weights in a ‘naive’ slice and found that this threshold protocol could induce LTP if preceded by low-frequency priming, but induced LTD if preceded by high-frequency priming. Moreover, the effect of priming on the outcome of this threshold protocol could be replicated by direct manipulation of the GluN2A/GluN2B ratio using partial block by pharmacological antagonists (having titrated concentrations with AP5 so that the NMDAR-mediated current reduction itself could not explain the effect) [109]. This result suggests that synaptic activity that increases the relative GluN2B content of a synapse biases it towards LTP, whereas a relative increase in GluN2A encourages the induction of LTD. The effect of prior activity has also been measured at the single synapse level [110]. Sparse transfection of neuronal cultures with a construct that blocks presynaptic vesicle release was used to silence some inputs size-matched postsynaptic spines, one with a silenced input and the other with an active input, were then exposed to a glutamate uncaging-based LTP induction protocol. The silenced synapses showed LTP and ensuing spine growth with a low-intensity protocol, whereas a higher number and longer duration of glutamate uncaging events were required to induce LTP and structural changes in spines that had been receiving an active input. Moreover, GluN2B appeared to mediate the increased propensity for potentiation, as the GluN2B-mediated current was enhanced at silenced synapses [110]. Therefore, both these studies suggest that metaplastic changes that cause a relative or absolute increase in the synaptic GluN2B content will bias a synapse towards LTP.

Neuromodulatory factors are known to influence how a synapse responds to a given activity pattern, and it is possible that they alter the basal state of the synapse in a subunit-selective manner. Dopamine levels modulate plasticity for example, dopamine application augmented tetanus-induced LTP via a cascade involving the D1/5 receptor, PKA and Src family kinases [111]. This LTP enhancement was blocked by Ro 25-6981 [111], suggesting that an increase in GluN2B subunit-containing NMDARs may be the expression mechanism for dopamine-dependent metaplasticity, which is in line with the findings from activity-based metaplasticity studies. However, not all metaplasticity studies have supported the conclusion that GluN2B encourages LTP. Priming by activation of G protein-coupled receptors and their downstream actuators in the Src family of kinases showed that metaplastic changes increasing GluN2B had the opposite effect on plasticity to the aforementioned studies [112]. Instead, Src kinase (activated by the pituitary adenylate cyclase-activating peptide 1 receptor PAC1R) was shown to phosphorylate GluN2A, and thus enhance GluN2A-mediated currents, whereas Fyn kinase (activated by the D1/5 receptor) phosphorylated GluN2B to enhance GluN2B-mediated currents. In turn, this influenced the response to a series of different frequency induction protocols: the switch from LTD to LTP induction occurred at 10–20 Hz without prior drug treatment, but exposure to PAC1R-activating drugs and a consequent GluN2A enhancement meant that LTP was induced at lower frequencies by contrast, the LTP induction threshold was shifted to higher frequencies if the D1/5 receptor had been activated, and hence GluN2B was increased [112]. This, therefore, suggested that GluN2A promotes LTP induction. Overall, further investigation into metaplasticity may help unravel how neuronal activity can alter synaptic weights and whether there is a contribution from changes in NMDAR subunit composition to synaptic priming effects. This is especially important given that learning and memory does not take place in an isolated fashion but against a background of prior neuronal activity and neuromodulatory input.

6. NMDA receptor subunits in behaviour

Synaptic plasticity is widely considered a cellular model for learning and memory [3,113], and there were early indications of an important role for NMDARs in certain forms of learning and memory [4,5]. In parallel to investigations into whether there is a subunit-selective contribution to LTP and LTD, subunit-specific manipulations have also been used to interrogate whether GluN2A and GluN2B subunits are important for behavioural tasks with different cognitive demands. In some cases, impairments in one direction of plasticity could be correlated with performance deficits.

GluN2A knock-out mice were the first to be studied behaviourally. The initial characterization of these mice suggested relatively pervasive memory deficits compared with C57BL/6 controls [82], but this was later attributed to the high degree of CBA strain character remaining in their genetic background, a line known to perform worse on the Morris water maze (MWM) task [114]. Once on an almost pure C57 background, the behavioural deficits found were relatively limited: GluN2A knock-out mice and mice lacking the GluN2A CTD showed no evidence of long-term memory deficits, as mice were not impaired on spatial tasks acquired over a number of days (the MWM and radial arm maze) [115]. Mice with a knock-out of Neuropilin and tolloid-like protein 1 (Neto1), a component of the NMDAR protein complex, have an approximately one-third reduction in GluN2A at the PSD and an LTP impairment, but also acquired the MWM at a normal rate [81]. Furthermore, pharmacological manipulations have produced a similar result Ge et al. [68] used the GluN2A subunit-preferring antagonist NVP injected intraperitoneally at a concentration that impaired LTP in vivo, and saw no deficits of NVP-treated rats either in MWM acquisition or consolidation. Thus, GluN2A does not seem necessary for incrementally acquired long-term memory tasks.

However, GluN2A manipulations have been associated with short-term spatial memory deficits, because GluN2A knock-out mice and mice lacking the GluN2A CTD were impaired on a non-matching-to-place T-maze task and also in their ability to distinguish between arms based on how recently they had been visited within a trial on the radial arm maze [115]. In further support of a role for GluN2A in short-term spatial memory, infusion of NVP into the CA1 impaired performance on a delayed alternation T-maze task [116]. It also seems that GluN2A subunit-containing NMDARs may contribute to the rapid formation of context or object representations. GluN2A knock-out mice, which had an increased threshold for LTP induction, were only impaired in a hippocampus-dependent contextual fear-conditioning paradigm when the task demands were increased by reducing context exposure before shock delivery [83], suggesting a deficit in the quick formation of a context representation. Neto1 knock-out mice were impaired in a displaced object recognition task, but not in a novel object recognition task [81] both these tasks require relatively rapid learning about an object, but only the displacement task has a clear spatial component and clear hippocampal involvement (but also see [117]). Overall, mice with no or reduced GluN2A do show some learning impairments, which are limited to short-term memory and the rapid acquisition of spatial information. In order to test whether this is a particular role of GluN2A or a non-specific effect of reduced NMDAR-mediated currents (especially given the lack of NVP selectivity), these results should be compared to those from mice with equivalent genetic and pharmacological manipulations of GluN2B levels.

There is considerable evidence that pharmacological and genetic manipulations of the GluN2B subunit produce similar impairments to those described above, suggesting that the short-term memory deficits observed in GluN2A-deficient mice are primarily owing to a general reduction in NMDAR-mediated current. For example, mice with postnatal GluN2B knock-out in pyramidal neurons of CA1 and the dentate gyrus (DG) of the hippocampus showed very similar impairments to global GluN2A knock-out mice, namely a short-term spatial memory deficit in spontaneous alternation in the T-maze but normal performance in the MWM [76]. This finding is supported by pharmacological data from rats, where infusion of Ro 25-6981 or ifenprodil into the CA1 region of the hippocampus impaired performance in tasks requiring short-term spatial memory, though long-term memory was not tested [116]. Rats were impaired on a delayed alternation T-maze task with 5- and 30-s delays, but they only made significantly more win-shift errors than controls after a 30-s delay. Rats also had longer escape latencies on trial 2 of a delayed matching-to-place water maze task, but only when the retention interval was 10 min, not 30 s [116]. These two findings suggest that GluN2B subunits may become more important when delay times increase. It should be noted, though, that pharmacological antagonism of NMDARs will also affect interneurons [59]. In general, the similar short-term memory impairments seen following manipulations that remove or reduce either the GluN2A or GluN2B subunit suggest that this phenotype is primarily owing to a decrease in NMDAR-mediated currents rather than a selective contribution of either subunit.

Nevertheless, there is evidence that GluN2B subunits could play a role additional to that shared with GluN2A in short-term memory, because other studies have found more pervasive impairments following GluN2B disruption. For example, KIF17 knock-out mice, which have reduced synaptic GluN2B levels owing to a transport impairment, performed worse on the MWM task [77,78], a novel object test with long delays [77] and contextual fear-conditioning [77]. A CA1 and neocortical GluN2B knock-out mouse line was deficient in MWM acquisition, fear-conditioning and T-maze spontaneous alternation [75]. However, these impairments might be attributable to extra-hippocampal deficits [76]. Therefore, it is reassuring that a more acute manipulation involving RNAi-mediated knock-down of GluN2B in the hippocampi of rats also showed slower acquisition of the MWM task [79]. The authors also found a correlation between GluN2B level and MWM performance within an aged population of rats [79]. However, the nature of the GluN2B involvement cannot be correlated with one direction of plasticity as RNAi caused an LTP deficit but LTD was not investigated, while both the aforementioned genetically altered mouse lines had deficits in both LTP and LTD. Nevertheless, these studies suggest that GluN2B may play a role in long-term memory that was not seen in mice with GluN2A knock-out or reduction, though additional studies using acute manipulations restricted to the hippocampus are required to verify this conclusion.

To dissect further a possibly unique role of GluN2B in learning and memory, it is useful to investigate the behavioural deficits associated with a selective impairment of one direction of plasticity. For LTP, a selective impairment arises if the association between CaMKII and GluN2B is weakened. This manipulation has been found to cause deficits in acquisition of both the MWM task, which could be overcome by extended training, and also in a delayed spatial win-shift eight-arm radial maze task, where mice had to remember the baited arms from the first session, and then avoid them in the next trial to find the food reward [51]. The authors ascribed this deficit to an impairment in forming a spatial map. Another mouse line with genetic disruption of the CaMKII–GluN2B interaction engineered in a different way showed fewer behavioural impairments, being normal in MWM acquisition and probe trial performance at 1–2 h post-training, but, interestingly, did exhibit long-term consolidation deficits, as they performed worse in a probe test 24 h after training [90].

Furthermore, in contrast to the impairments seen following GluN2A overexpression [88], GluN2B overexpression has been shown to improve performance of adult [84] and aged mice [85] and rats [86] in a battery of tests including context and cued fear-conditioning, object recognition memory at longer retention intervals, MWM acquisition and spatial short-term memory. Cdk5 knock-out mice, which have increased synaptic GluN2B owing to a reduction in its degradation were also found to have improved contextual learning in a context-dependent fear-conditioning paradigm [87]. All these behavioural improvements were associated with increased LTP, while LTD, when investigated, was found not to change [86]. Stronger evidence for a link between subunit-specificity in LTP and learning comes from tasks where learning and postsynaptic responses can be measured simultaneously. Using the trace-conditioning paradigm in rats, Valenzuela-Harrington et al. [118] showed that acquisition of this task was associated with an increased strength of the medial perforant pathway-DG synapse of the hippocampus. Moreover, systemic administration of Ro 25-6981 blocked both task acquisition and the changes in synaptic strength. Therefore, possibly through its contribution to LTP induction via CaMKII, GluN2B seems to have an important role in learning and memory.

Some behavioural deficits have also been related to a GluN2B-dependent LTD impairment. Ge et al. [68] administered Ro 25-6981 intraperitoneally to rats, which impaired LTD but not LTP in vivo, and found that, while acquisition of a MWM task was not impaired by this manipulation, rats that received Ro 25-6981 on the training day could not recall the platform location 24 h later owing to a consolidation-related deficit. This compromised performance could be mimicked by injection of a peptide that prevented AMPAR endocytosis (Tat-GluA23Y), suggesting that the impairment was related to an LTD-like process. However, an alternative interpretation is that this manipulation disrupted homoeostatic resetting of the network and that this process is required for consolidation.

A finding reported by a number of groups is a possible link between LTD and reversal learning. Duffy et al. [101] found that subcutaneous administration Ro 25-6981 blocked LTD and impaired reversal learning in mice, and deficient reversal learning in the MWM was also found in mice with GluN2B knock-out in the hippocampus, though LTD was not tested [76]. In rats, GluN2B-dependent LTD could not be induced in naive animals in vivo, but could be induced once animals experienced MWM training when rats that had previously acquired the MWM were tested on a reversal phase, those rats treated with either Ro 25-6981 or the Tat-GluA23Y peptide were impaired in reversal learning [119]. In further support of a role of GluN2B-dependent LTD in reversal learning, enhancing LTD by subcutaneous administration of d -serine increased the rate at which the mice learnt the reversal phase in the MWM [101], while a mouse line with enhanced GluN2B-mediated currents showed no improvement in the acquisition of an MWM task, but did show faster reversal learning [87]. Thus, many of the studies investigating a role of NMDAR subunits in behaviour have found a ‘perseverance phenotype’ that correlates with an impairment in GluN2B-dependent LTD.

There is a large body of evidence to suggest that synaptic plasticity supports learning and memory [4,5,118,120], but the link is not yet resolved [121,122]. Therefore, studies that correlate plasticity and behaviour have been informative, irrespective of whether the underlying, possibly subunit-selective, mechanisms are known. Nevertheless, correlations between plasticity deficits and learning and memory impairments do not provide a causative link. Moreover, we should be cautious in concluding that such studies provide definitive evidence given the possible compensatory mechanisms and network changes that may be triggered by the experimental manipulations described here. Of course, even if plasticity does support learning and memory, we still do not know what types of plasticity would be engaged in tasks with different cognitive demands, and thus, with all this considered, it is no surprise that the findings from these studies are complex and often contradictory. In some cases, correlations have been found between an impairment in one direction of plasticity and task performance, and occasionally particular subunits implicated. However, equally, some manipulations that caused large impairments in the magnitude of plasticity had little or no effect on the behavioural tests chosen, only affected certain aspects of the tasks, or even only caused an impairment when task difficulty was increased or demands were changed, such as by introducing a delay. The aforementioned hemispheric asymmetry in the distribution of the GluN2B subunit and LTP in the mouse CA3-CA1 synapse may help further elucidate the role of synaptic plasticity in learning and memory, and also determine if there is a particular subunit involvement. Although it is not known how this asymmetry arises in development and is maintained during adulthood, evidence that it may be important in memory processing comes from inversus vicerum mice that show disrupted asymmetry and display memory impairments [123].

7. Summary

The hypothesis that different GluN2 subunits selectively mediate different directions of plasticity is not supported by the available studies. Although the lack of specific GluN2A antagonists prohibits the conclusive testing of this hypothesis, when manipulations control for the magnitude of NMDAR-mediated current reduction, there is no clear evidence, as yet, that either subunit plays an irreplaceable role in LTP induction. Nevertheless, the different kinetics conferred on the NMDAR by the GluN2A and GluN2B subunits means the contribution of each subunit to charge transfer may vary according to the pattern of presynaptic activity, and therefore they may not play an equal role in potentiation under physiological conditions.

Irrespective of the induction paradigm used, however, most evidence suggests that the GluN2B subunit has a greater importance for LTP induction. An explanation for why GluN2B can exert a strong influence on LTP in the adult hippocampus, despite comprising a smaller proportion of GluN2 subunits than early in development, would be if GluN2B subunit-containing NMDARs carry more Ca 2+ influx per unit of current [38]. An alternative explanation, which is more in line with the prevailing evidence, is that GluN2B is particularly important for LTP induction because its presence at the spine anchors CaMKII at the PSD. Thus, GluN2B subunit-containing NMDARs enable activation of the downstream signalling cascades that mediate synaptic strengthening, irrespective of whether they support a majority of the Ca 2+ influx. This suggests that GluN2B is a necessary factor for LTP induction under naturalistic conditions, but, importantly, whether or not it is sufficient to support the requisite Ca 2+ influx may depend on the pattern of presynaptic activity. Moreover, although spine-based imaging studies and asymmetry studies have suggested that small spines, which tend to be GluN2B-rich, have a greater propensity for the induction and maintenance of LTP, the presence of GluN2B may not be sufficient if downstream signalling pathways and structural changes are already saturated.

The LTD literature is ambiguous, although a greater number of studies and the developmental profile of LTD would support a more central role of GluN2B in its induction. Exactly why this may be is not clear, but could relate to intracellular associations distinct from that with CaMKII. Most plasticity studies investigate the response to induction protocols at either the cellular or extracellular field potential level, and thus record the summed changes at many synapses. However, synapses are not equal, and instead have different structural and molecular signatures, and this may help explain at least some of the contradictory findings in the plasticity literature. Overall, given that GluN2B seems important for both LTP and LTD, a parsimonious hypothesis is that this subunit confers the synapse with malleable properties.

Given the contradictory conclusions of plasticity studies, it is unsurprising that neither subunit has yet been found to have a clear role in learning and memory. What the current literature does suggest is that both GluN2A and GluN2B support short-term memory processes. There is also some evidence that GluN2B subunit-containing NMDARs may make a greater contribution to learning when information must be retained for a longer delay or there is incremental task acquisition across a number of days.

It may seem surprising that short-term memory tasks often appear the most susceptible to genetic lesions or pharmacological block of NMDARs, especially in light of the proposal that NMDAR-dependent LTP supports task acquisition through incremental learning [124]. However, a number of factors make it premature to conclude that this necessitates a revision of the theory linking plasticity and learning. Firstly, NMDARs have roles additional to their involvement in plasticity, as further described later. Secondly, a global knock-out would also remove NMDARs located on interneurons, and hence may alter the network dynamics important for short-term memory [125]. Thirdly, in the few studies where the subunit knock-out is restricted to pyramidal cells in the hippocampus, navigation systems may still be affected, and how that could interact with the cognitive demands of short- and long-term memory tasks is not known. Fourthly, even localized manipulations have been chronic, and so compensatory mechanisms are likely to be recruited it is conceivable that short-term memory requires greater online processing power or rapid, large-scale network changes, either of which might need a higher level of Ca 2+ influx to drive them, and thus would be more sensitive to manipulations that reduce NMDARs and hence total Ca 2+ influx. Finally, the fact that these manipulations also often impair LTP does not necessarily imply a link between LTP and short-term memory, it only shows that both processes rely on NMDARs. Indeed, short-term potentiation also requires NMDARs and may have subunit preference [9].

In some studies, it has been possible to relate a behavioural impairment to a selective deficit in one direction of plasticity. In particular, there is evidence that the gradual acquisition and consolidation of information may be supported by the GluN2B–CaMKII interaction, which is also important for LTP. Another common finding is that the mechanisms that support reversal learning may involve a GluN2B-dependent LTD-like process. One caveat of studies relating learning impairments with LTD deficits is that many laboratories find that they cannot induce LTD in adult rodents, despite rodents of that age still being able to perform the associated memory tasks. This suggests that either the low-frequency paradigms used to induce LTD do not successfully mimic the type of activity that triggers reduction of synaptic weights in vivo, or, even, that synapse-specific reductions in strength do not underlie learning instead, they may be more important for circuit refinement during development. One way to reconcile these observations would be that synaptic strengthening always supports the acquisition of new information, but a concomitant depression is also necessary in behavioural tasks that require re-learning (reversal learning, extinction) or refinement of learning (recognition memory) to increase the gain of the newly potentiated synapses above noise or ‘background’. In this scenario, a limited number of synapses would potentiate and carry new information, but an overall depression would be observed in extracellular field or whole-cell measurements that record populations of synapses. Such a form of synaptic depression would be functionally distinct from purely homoeostatic mechanisms, as it would contribute to information coding.

8. Future directions

To fully investigate the subunit selectivity hypothesis, priorities for the future should include the development of tools to selectively block GluN2A subunit-containing NMDARs. In addition, establishing techniques to target triheteromeric NMDARs is fundamental to investigate their role it appears likely that some of the confusion in the current literature may result from the differing extents to which these receptors have been affected by the manipulations used. Combining subunit labelling with fluorescent probes and high-resolution live imaging could help to study this population in single spine-based plasticity studies. Alternatively, if a posttranslational modification, such as a disulfide bridge, could be introduced to link GluN2A and GluN2B subunits without disrupting their kinetic properties, this would help investigation at the level of whole-cell or extracellular field recordings. Given that a number of studies have suggested that the triheteromeric population may be substantial, or even dominant, in the adult brain, it is likely that this NMDAR composition has an important function. This may be because triheteromers have intermediate channel kinetics, and so combine the faster response and integrative properties provided by the GluN2A subunit with the important intracellular associations conveyed by the GluN2B subunit—in particular allowing Ca 2+ influx to be closely coupled to downstream signalling molecules, such as CaMKII.

In order to understand whether the links between subunit involvement in plasticity paradigms and behaviour are meaningful, it is important to demonstrate conclusively that changes in synaptic weights can support learning and memory. This could then be dissected further to establish the types and synaptic location of plasticity that support behavioural tasks with different cognitive demands. In the past, technical limitations have prevented interrogation of a causal link between plasticity and learning and memory. However, new advances may help us test this hypothesis more directly. As the neural code is not yet resolved, driving learning to guide a complex behaviour by replacing experience with artificially introduced, specific changes in synaptic weights may not be on the immediate horizon, although significant steps in this direction have recently been taken [126]. Other approaches may still be fruitful: the newly developed optogenetic and pharmacogenetic technologies can be used to target synapses with distinct molecular and structural compositions and, when used in combination with molecular trickery, are likely to be a particularly powerful way to investigate whether different types of synapse play distinct computational roles. Such acute manipulations have already revealed new information about hippocampal function that differs from what had been suggested by pharmacological lesion studies [127]. Furthermore, given the technological advances in wireless recording devices, it will be important to record postsynaptic responses in vivo during behaviour to provide a read-out of the manipulations made. Such recordings would also show the extent of synaptic weakening in reversal tasks. In terms of the subunit-selectivity hypothesis, the current evidence suggests that the extent to which each subunit is involved in plasticity is affected by the induction protocol used owing to their distinct kinetics. Therefore, to understand the importance of GluN2A and GluN2B in physiological conditions and in relation to learning and memory, emphasis must be placed on investigating the natural activity patterns that occur during behaviour and which drive long-lasting changes in synaptic strength.

However, it should be remembered that NMDARs do not only play a role in plasticity, but also in slow or tonic response components of neural activity for example, they have been shown to support slow responses in the thalamus and sensory tuning in the neocortex. Moreover, NMDARs may not even be required for certain forms of learning it has recently been shown that mice with NMDAR knock-out in the DG and dorsal CA1 can still perform tasks typically considered to be NMDAR-dependent, for example the MWM [121]. This does not mean that, under normal circumstances, NMDARs are not involved in this type of learning, as other mechanisms or hippocampal regions could take over following prolonged deletion. Nevertheless, it will be important to investigate whether NMDARs at all hippocampal synapses are involved in learning and memory or, alternatively, whether only a subset is, while the rest of the hippocampal NMDARs perform computations similar to those in sensory neocortex. The hippocampal asymmetry reported in humans [128] and, more recently, in the mouse [69,94,96], makes it possible that the hippocampus consists of parallel networks, one being primarily a response network involved in sensory coding and navigation, whereas the other supports learning of associations. Despite the huge advances in our understanding of hippocampal function, it seems that there is still much more to uncover.


Design Strategy for Discovery of Aptamer Targeting the Open-Channel Conformation of AMPA Receptors

In this work, we chose to use the GluA2Qflip AMPA receptor as the target of in vitro selection (Experimental Procedures and Figure S1, Supporting Information). GluA2 is one of the four AMPA receptor subunits, and can form homomeric, functional channel by itself, like any other AMPA receptor subunits (9). GluA2 is considered a key subunit that mediates excitotoxicity (25). The 𠇏lip” isoform of GluA2Q or GluA2Qflip, generated by alternative splicing, is known to desensitize less rapidly than the 𠇏lop” isoform (5, 26). The edited or the Q isoform (i.e., glutamine at the glutamine/arginine or Q/R editing site) is calcium-permeable, whereas the R isoform is not (27). An abnormal expression of the Q isoform of GluA2 is linked to neurological disorders such as ALS (28).

To make it practically possible to apply an in vitro evolution approach to identifying aptamers against the open-channel conformation of AMPA receptors, we specifically designed the following experiments. First, we used a saturating agonist concentration to “titrate” the receptor population to maximize the fraction of the open-channel conformation of GluA2Qflip. In other words, we wanted to present specifically the open-channel conformation of GluA2Qflip as the target of the selection in anticipation of finding aptamers that would specifically recognize the open-channel conformation. Second, the open-channel conformation lasts no more than a millisecond or so (depending on the glutamate concentration) after glutamate binding (5), whereas the binding reaction between the receptor and RNA library requires at least 30 min to complete (19) (Experimental Procedures and Figure S1, Supporting Information). Thus, we had to “trap” the open-channel conformation long enough for the binding reaction. To do so, we decided to choose kainate as the agonist. Kainate is capable of producing a non-desensitizing current response with GluA2 after kainate binds to it, indicative of a persistent existence of the open-channel conformation (29). Experimentally, we preincubated the cell membrane containing the GluA2Qflip receptor with 1 mM kainate (i.e., this was a saturating concentration). Third, we used a noncompetitive inhibitor, i.e., GYKI 47409, to elute putative RNAs that might bind to the same site or mutually exclusive sites(s) (Figure S1, Supporting Information). GYKI 47409 is a 2,3-benzodiazepine derivative and has an inhibition constant (KI) of

3 μM for the open-channel conformation of GluA2Qflip or

2-fold higher affinity than towards the closed-channel conformation (Pei and Niu, unpublished data). The overall design of our experiments was to specifically identify an aptamer that targeted the open-channel conformation by its binding to a noncompetitive site.

Identification of an RNA Aptamer that Inhibits AMAP Receptors

The GluA2Qflip channels were transiently expressed in HEK-293S cells, and the membrane fragments harboring the entire functional receptors were used for in vitro selection (19). To suppress potentially hazardous enrichment of nonspecific RNAs bound to any other “targets”, such as lipids, we also carried out the negative selection as in rounds 5, 10 and 13, in a total of 14 selection cycles, in which plain HEK-293 cell membrane lacking only the GluA2Qflip receptors was used to absorb these nonspecific RNAs. In contrast, the positive selection rounds involved the use of GYKI 47409, as mentioned before, to elute potentially useful RNAs. The eluted RNAs were amplified by RT-PCR, and an enriched RNA library was then transcribed for a new round of selection (Figure S1, Supporting Information). After 14 cycles, we identified some enriched sequences ( Figure 1A ). An enriched sequence was one with at least two copies in the entire sequence pool of 83 clones (i.e., 43 clones from round 12 and 40 clones from round 14). The putative inhibitory property of these enriched sequences was then functionally tested by the use of whole-cell current recording with GluA2Qflip expressed in HEK-293 cells. Based on the whole-cell recording results (see representative traces in Figure 1B ) or the ratio of the current amplitudes in the absence and presence of an aptamer, A/A(I) (shown on the right column in Figure 1A ), we concluded that AG1407, the most enriched sequence, was one of the most potent inhibitors. A further test of AG1407 at the same aptamer concentration but with increasing glutamate concentrations showed that AG1407 inhibited the open-channel conformation of GluA2Qflip (Figure S2, Supporting Information).

Identification of the Minimal, Functional Aptamer Sequence

Next we systematically truncated aptamer AG1407 in order to identify the minimal, yet functional sequence. Guided by the secondary structure prediction using the Mfold program (30), we constructed and functionally tested shorter versions of AG1407 ( Figure 2A ). Based on A/A(I) value (shown in red color at the bottom of each predicted structure in Figure 2A ), we found the 56-nucleotide (nt) version of AG1407, termed as AG56, was the functional aptamer with the minimal length of sequence. In contrast, either shortening the three-way junction by deleting UUGUGA sequence (i.e., the 46 nt RNA) or removing the bulge at the U50 position (i.e., 45 nt RNA) or truncating the base-paired stem in the first stem-loop region (i.e., 48 nt RNA) ( Figure 2A ) resulted in nonfunctional RNAs. Therefore these structural elements are essential in the folding of AG56 as a functional aptamer. Consequently we used AG56 as the minimal aptamer for all of the functional characterizations described below.

The minimal, functional sequence of the aptamer and the AMPA-receptor subtype selectivity. (A) In truncating the sequence of AG1407 to identify the minimal, functional 56-nt sequence or AG56, only the most stable secondary structure of each sequence, with the free energy listed below that structure, was considered, based on the prediction by Mfold. The black and brown colored letters represent the constant and variable sequence regions, respectively. A “scissor” indicates where a particular sequence was cut, and the green letters represent the cut-off nucleotides. The inhibitory function, shown as A/A(I) value, is listed at the bottom of each truncated structure. (B) By whole-cell current recording assay, AG56 selectively inhibited the open-channel conformation of all AMPA receptor subunits (left panel) (see further explanation and statistical analysis in Figure S3, Supporting Information). Yet, AG56 did not affect GluK1Q and GluK2Q, two representative kainate receptor channels, nor GluN1A/2A and GluN1A/2B NMDA receptor channels. For each of the receptor types tested, the glutamate concentration was chosen to be equivalent to

95% fraction of the open channels. Specifically, the glutamate concentration was 0.04 mM (for the closed-channel conformation)/3 mM (for the open-channel conformation) of GluA1flip, 0.1 mM/3 mM for GluA2Qflip, GluA3flip, and GluA4flip, as well as 0.04 mM/3 mM for GluK1 and GluK2Q.

Functional Characterization of Aptamer AG56 by Whole-Cell Recording

AG56 was functionally characterized in a series of experiments. First, like its predecessor sequence AG1407 (Figure S2, Supporting Information), AG56 selectively inhibited the open-channel, but not the closed-channel, conformation of GluA2Qflip ( Figure 2B , left panel). Furthermore AG56 similarly inhibited the open-channel conformation of all other AMPA receptor subunits, i.e., GluA1, 3 and 4, although the inhibitory effect of AG56 on GluA4 was weak ( Figure 2B left panel, and Figure S3, Supporting Information). Yet, AG56 had no inhibitory effect on any of the closed-channel conformations ( Figure 2B left panel). AG56 did not affect either the kainate receptor channels (i.e., GluK1 and GluK2) or the NMDA receptor channels (i.e., GluN1A/2A and GluN1A/2B) ( Figure 2B right panel). It should be noted that GluN1A/2A and GluN1A/2B are two dominant NMDA receptor complexes in vivo (31) and neither GluN1A nor GluN2A or GluN2B can form a functional channel by itself (32). These results thus suggest that AG56 is an AMPA receptor-subtype selective inhibitor by targeting the open-channel conformations, and AG56 is without any unwanted, cross activity against other glutamate receptor subtypes.

Mechanism of Inhibition of AG56 on GluA2Qflip: Homologous Binding Studies

We further elucidated the mechanism of action of AG56 on the GluA2Qflip receptor channel expressed in HEK-293 cells. In this study, we first determined the inhibition constant of AG56 to be 0.95 ± 0.20 μM (the solid line in Figure 3A ) for the open-channel conformation of GluA2Qflip at 3 mM glutamate concentration where almost all of the channels were in the open-channel conformation (this was because the EC50 value of GluA2Qflip with glutamate was 1.3 mM and the channel-opening probability of GluA2Qflip was near unity (21)). In contrast, AG56 did not inhibit the closed-channel conformation of GluA2Qflip, as verified by a series of aptamer concentrations ( Figure 3A ) but at 100 μM glutamate concentration where most of the receptors was in the closed-channel conformation (21). This result could be explained by a noncompetitive mechanism by which AG56 bound to a regulatory site or noncompetitive site, and such a site was accessible from both the closed-channel and the open-channel states or conformations yet only the interaction of the aptamer with the open-channel conformation resulted in inhibition. Alternatively, this result could be explained by an uncompetitive mechanism, such as an open-channel blockade model, by which AG56 would only inhibit the open-channel conformation, because the uncompetitive site would only be accessible through the open-channel conformation (33) (see the two mechanisms in Appendix). To differentiate these two mechanisms, we first carried out a homologous competition binding assay (20) and found that AG56 not only bound to the closed-channel conformation (i.e., the unliganded, closed-channel receptor form) but did so with an affinity, i.e., Kd = 68 ± 40 nM ( Figure 3B left panel) similar to that for the open-channel conformation, i.e., Kd = 80 ± 23 nM ( Figure 3B right panel). This result was consistent with a noncompetitive mechanism, based on the fact that AG56 was found to bind to the closed-channel conformation in addition to its binding to the open-channel conformation. This result, however, was inconsistent with the uncompetitive mechanism.

Characterization of inhibition constant, binding affinity and mechanism of action of AG56 with GluA2Qflip expressed in HEK-293 cells. (A) By whole-cell current recording assay, AG56 inhibited the open-channel (measured at 3 mM of glutamate concentration), but not the closed-channel (measured at 0.1 mM of glutamate concentration) conformation of GluA2Qflip (the dotted line indicates no inhibition or A/A(I) = 1). The inhibition constant, KI, for the open-channel conformation of GluA2Qflip by AG56 was determined to be 0.95 ± 0.20 μM (i.e., the upper solid line). (B) The homologous competition binding of AG56 to GluA2Qflip receptor was plotted for both unliganded, closed-channel form (solid circle) and open-channel form (hollow circle). The binding constant, Kd, of AG56 to the closed-channel (left panel) and the open-channel (right panel) forms of GluA2Qflip was determined to be 68 ± 40 nM and 80 ± 23 nM, respectively, based on triplicate data sets. The binding constant was calculated using eq. 1 in Experimental Procedures. (C) The laser-pulse photolysis measurement of the effect of AG56 on the channel-closing rate constant or kcl (left panel) and channel-opening rate constant or kop (middle panel) with GluA2Qflip expressed in HEK-293 cells. Specifically, at 100 μM photolytically released glutamate concentration, the kobs value, which reflected kcl, was decreased from 2,200 s 𢄡 (control or 𢄠.5 μM AG56, black trace, left panel) to 1,600 s 𢄡 (+0.5 μM AG56, red trace, left panel). At 340 μM photolytically released glutamate concentration, the kobs value, which reflected kop (middle panel), was 5,128 s 𢄡 and 4,405 s 𢄡 in the absence and presence of 0.5 μM AG56. The difference, however, or Δkobs = kobskobs’ = Δkcl was invariant even when glutamate concentration increased (the right panel). Here Δkcl = kclkcl’ where kcl’ is the inhibited kcl value and the kcl is the channel-closing rate constant without AG56 (see equ. 9 in Appendix). Each data point represents at least one measurement from a single cell where kobs is the control rate constant and kobs’ is the rate constant in the presence of 0.5 μM AG56.

Mechanism of Inhibition of AG56 on GluA2Qflip: A Laser-Pulse Photolysis Measurement of the Effect of AG56 on the Channel-Opening Rate Process

We further characterized the mechanism of inhibition of AG56 on the channel-opening kinetic process of GluA2Qflip. Using a laser-pulse photolysis technique, together with a photolabile precursor of glutamate or caged glutamate, which provided a time resolution of

30 microsecond (22), we specifically measured the effect of AG56 on both the channel-closing (kcl) and the channel-opening (kop) rate constants (23) ( Figure 3C left and middle panels, respectively). This experiment enabled us to simultaneously follow not only the rate of channel opening but also the current amplitude, prior to channel desensitization (23) ( Figure 3C , left and middle panels). The magnitude of kcl reflects the lifetime (τ) of the open channel (i.e., τ = 1/kcl) and the effect of an inhibitor on kcl thus reveals whether or not it inhibits the open-channel conformation (23). In contrast, kop reflects the closed-channel conformation and the effect on kop reveals whether the inhibitor inhibits the closed-channel conformation (23) (see the rate equations and quantitative treatment of the rate data in Appendix). Experimentally, at a low glutamate concentration (i.e., 100 μM photolytically released glutamate) where kcl was measured (23), AG56 inhibited the rate of channel closing, consistent with a noncompetitive mechanism by which it inhibited the open-channel conformation. Yet, AG56 did not affect the current amplitude ( Figure 3C left panel). This was not surprising because the amplitude observed at this low glutamate concentration (i.e., 100 μM photolytically released glutamate) was dominated by the closed-channel receptor population (notice this was consistent with the amplitude measurement shown as the red dashed line in Figure 3A ).

However, when the concentration of glutamate increased and kop became measurable (23) (see Appendix), AG56 did not inhibit kop ( Figure 3C middle and right panels). This result suggested that the inhibition of the rate by AG56 could be completely ascribed to the inhibition of kcl by AG56 such that the difference between the observed rate constant of channel opening or Δkobs in the absence and presence of AG56 at the same AG56 concentration was invariant in spite of increasing glutamate concentration ( Figure 3C right panel, and see the mechanistic treatment of the rate data, specifically equ. 9, in Appendix). In other words, the fact that Δkobs remained the same (or Δkobs = constant), verified in a series of increasing glutamate concentrations, was entirely consistent with the prediction by equ. 9 (see additional explanation about equ. 9 in Appendix). The lack of an inhibitory effect of AG56 on kop ( Figure 3C middle and right panels) further demonstrated that AG56 did not inhibit the closed-channel conformation.

On the other hand, AG56 reduced the current amplitude at a higher glutamate concentration (see the difference in peak current amplitudes between the red and blue traces in Figure 3C middle panel and see an additional trace in Figure S4, Supporting Information, for a higher current amplitude inhibition at a higher glutamate and inhibitor concentration). Again, this was expected because the current amplitude at a higher glutamate concentration began to reflect more on the open-channel receptor population. Furthermore, the effect of AG56 on the current amplitude from the rate measurement ( Figure 3C middle panel, and Figure S4, Supporting Information) was entirely consistent with the amplitude measurement using a rapid solution flow method ( Figure 3A ). Taken together, the results from the binding site/affinity assessment ( Figure 3B ) and the chemical kinetic characterization of the effect of AG56 on both kcl and kop ( Figure 3C ) as well as the amplitude measurement ( Figure 3A ) are consistent only with AG56 being a noncompetitive inhibitor selective to the open-channel receptor conformation. Conversely, an uncompetitive mode of action is inconsistent with our data, i.e., AG56 would only bind to the open-channel conformation, but not the closed-channel conformation. A competitive mode of action is also inconsistent with our data, i.e., AG56 would only be effective as an inhibitor at a low ligand concentration and would inhibit kop, but not kcl.

7. Physiological and Pathological Roles of the CB1R

Given the widespread distribution of CB1Rs in the human body, it is reasonable for one to speculate a broad spectrum of physiological roles of the CB1R [3,9,63,126]. Indeed, the CB1R and the endocannabinoid system are largely involved in various aspects of central neural activities and disorders, including appetite, learning and memory, anxiety, depression, schizophrenia, stroke, multiple sclerosis, neurodegeneration, epilepsy, and addiction [3,9,126,127]. The CB1R is also involved in physiological and pathological conditions in the PNS and peripheral tissues, including pain, energy metabolism, cardiovascular and reproductive functions, inflammation, glaucoma, cancer, and liver and musculoskeletal disorders [63]. The expression of CB1R remarkably fluctuates in many pathological conditions, underscoring its critical role in a wide spectrum of biological activities [69]. Interestingly, in some cases, both positive and negative alterations in CB1R expression and functionality have been reported [69]. Moreover, the administration of CB1R agonists exert biphasic effects in several conditions [128]. On the other hand, the widespread presence of the CB1R limits the therapeutic application of CB1R ligands due to various side effects. These facts underscore the significance of understanding and manipulating the endocannabinoid system in a condition-specific manner.

CB1R has been found to inhibit GABA and glutamate release from presynaptic terminals, which confers the CB1R with the ability to modulate neurotransmission [60,129]. This has been proposed as a plausible underlying mechanism of CB1R-mediated neuroprotection against excitotoxicity, a prominent pathological process of many neurological disorders, including epilepsy and neurodegenerative diseases [34,130,131]. To date, numerous studies have shown that the CB1R plays a neuroprotective role against excitotoxicity induced by various stimuli [131,132,133,134]. It has been demonstrated recently that in mouse brain, the neuroprotective effect exerted by CB1R against excitotoxicity is restricted to the CB1R population located on glutamatergic terminals [130]. In addition to the prominent inhibitory effects on Ca 2+ influx and glutamate release, CB1R-mediated neuroprotection also involves inhibition of nitric oxide (NO) production, reduction of zinc mobilization, and increase of BDNF expression [134,135,136]. Recent studies have implicated a direct physical interaction between CB1Rs and NMDARs in the presence of histidine triad nucleotide-binding protein 1, which allows CB1Rs to negatively regulate NMDAR activity and protects neural cells from excitotoxicity [136,137].

Specifically, altered expression of the CB1R and other elements of the endocannabinoid system have been observed in various neurodegenerative diseases, such as Alzheimer’s disease (AD), Parkinson’s disease (PD) and Huntington’s disease (HD) [3]. The upregulation of the CB1R and endocannabinoid system activity has been observed in the basal ganglia of experimental models of PD, which could be a mechanism to compensate the degenerated dopaminergic neurons of the substantia nigra, or a pathological process that contributes to the worsening of the disease [138]. Interestingly, decreased endocannabinoid system activity has also been reported in PD models [128]. Moreover, both the FAAH inhibitors and CB1R antagonists have been shown to alleviate the motor symptoms in PD models [128]. Similarly, although changes of CB1R expression in AD patients or animal models are still controversial, the activation of the CB1R has been shown to prevent amyloid β-induced neurotoxicity in several cell models [139,140,141,142,143,144]. In addition, the activation of the CB1R has been reported to be beneficial in AD animal models with memory deficits and cognitive disorders [145,146,147]. On the other hand, studies have emphasized the beneficial potentials of the CB1R in HD pathogenesis. In 1993, decreased expression of the CB1R was first reported in the substantia nigra of HD patients via autoradiography [148]. Further studies revealed a progressive loss of CB1Rs as an early sign of HD, which occurred before the onset of actual neurodegeneration, and hastened the worsening of HD [149]. This observation was confirmed at the mRNA level as well as with CB1R immunoreactivity in several transgenic HD mouse models (reviewed in [3]). A recent study described downregulation of the CB1R not only in medium spiny projection neurons (MSNs) but also in a subpopulation of interneurons that are selectively preserved in both transgenic HD mice and HD patients [150]. Delayed loss of CB1Rs in HD transgenic mice R6/1 was seen in enriched environment, accompanied by delayed onset of motor disorders and disease progression [151]. Moreover, in HD transgenic mice R6/2, CB1R knockout leads to the worsening of motor performances, increased susceptibility to 3-nitropropionic acid, and exacerbated striatal atrophy and Huntingtin (Htt) aggregates [133,152]. Selective increase in CB1R expression in MSNs improves the survival of excitatory projection neurons, but does not promote the motor performances of HD transgenic R6/2 mice [153]. Administration of THC has been reported to ameliorate motor disorders, striatal atrophy, and Htt aggregates in transgenic mice, although controversy exists [133,154]. Activation of the CB1R inhibits glutamate release while increases BDNF release from presynaptic terminals in mice [131]. Further investigation in HD cell models revealed that CB1R activation can protect striatal cells against excitotoxicity through increased BDNF expression via PI3K/Akt pathway [133]. These observations support a critical and possibly beneficial role of the CB1R in neurodegenerative diseases.

The historical record of the anti-epileptic effects of the CB1R dates back centuries [1]. Case reports on the beneficial effects of cannabinoids on epileptic patients became available only after the identification of THC [155,156]. However, studies also suggested increased seizure frequency after marijuana smoking [157]. This paradoxical effect of cannabinoids on epilepsy is not only seen in human studies but has also been reported in animal models [158,159]. Activation of the CB1R by AEA has been shown to inhibit electroshock-induced seizures in rats [159]. Conversely, CB1R activation in FAAH knockout mice displays increased susceptibility to kainic acid-induced seizures [158]. The alteration of the endocannabinoid system following epilepsy is cell type-specific. This concept is supported by previous animal studies showing that CB1R retrograde signaling is selectively enhanced at inhibitory but not excitatory synapses, resulting a persistent potentiation of DSI but not DSE in febrile seizures, which leads to hyper-excitability of neurons, thus contributing to the exacerbation of seizures [160,161]. Moreover, this CB1R-mediated enhanced suppression of inhibitory neurons is phase-dependent as well. Hippocampal tissues from epileptic patients in the acute phase of epilepsy display decreased CB1R density, especially in the dentate gyrus, whereas in patients in the chronic phase of epilepsy, an upregulation of CB1R has been observed [162,163,164,165].

Despite the low expression of CB1R in hypothalamus, cannabinoids are long known for their effects to stimulate appetite, prominently in a CB1R-dependent manner [166]. Endocannabinoids levels are increased in the rat hypothalamus during fasting and return to normal levels after food consumption [167]. The stimulation of appetite and feeding behavior is observed after direct injection of endocannabinoids and is abolished by the administration of CB1R antagonists [167]. Furthermore, activation of ventral striatal CB1Rs inhibit GABAergic neurons, resulting in a hypophagic but not an orexinergic effect [168]. A recent study has demonstrated that CB1R-induced feeding behavior is promoted by the activation of hypothalamic POMC neurons [81]. In addition to the hypothalamus, olfactory process have been proposed to be involved in the positive regulation of CB1R-mediated food intake [169]. Moreover, crosstalk between CB1Rs and the important hormones involved in appetite regulation, including ghrelin, leptin, and orexin, has been extensively reported [68,166]. CB1Rs expressed in the GI tract also are involved in metabolic process and energy balance, as discussed in the previous section. These studies suggest that CB1R-mediated regulation of appetite involves at least two aspects: through the regulation of CNS region related to appetite, and through the modulation of metabolic hormones and digestive functions on site. Rimonabant, a CB1R antagonist, displayed remarkable anti-obesity effects, yet the accompanying psychiatric side effects lead to its withdrawal from the market [170]. An up-to-date review by Koch have summarized the recent progress on elucidating the role of CB1R in appetite control [171].

The regulation of pain is one of the earliest medical applications of cannabinoids [1,2]. Numerous studies have documented the analgesic effects of cannabinoids in different types of pain, including chemical, mechanical, and heat pain, as well as neuropathic, inflammatory, and cancer pain [172,173]. The endocannabinoid system also is involved in the regulation of nociception [3]. A newly published review paper has discussed the preclinical and clinical studies on the role of endocannabinoids in the control of inflammatory and neuropathic pain in details [173]. In addition to the CB1R, there also is evidence supporting the involvement of the CB2R and TRPV1 in cannabinoid-mediated regulation of pain [174,175]. Furthermore, the phytocannabinoids have drawn much attention nowadays in the field of antinociception and other neurological disorders. CBD, for instance, has been shown to modulate chronic pain in several studies [173]. The drug with brand name Sativex, containing equal amount of THC and CBD, is used to treat several kinds of multiple sclerosis associated symptoms including chronic pain [176]. Despite the fact that CBD has negligible affinity to the CB1R and CB2R, recent studies have suggested that it is an allosteric modulator and an indirect antagonist of CBRs, with the ability to potentiate the effect of THC [177].

Cannabinoids used in cancer are best-known for their palliative effects, including reducing nausea and vomiting, alleviating cancer pain, and stimulating appetite [178,179]. It has been argued that cannabinoids can exert anti-tumor effects directly through the inhibition of cell proliferation and induction of apoptosis, or indirectly through the inhibition of angiogenesis, invasion and metastasis [180]. Numerous studies using synthetic/endo-/phyto-cannabinoids and endocannabinoid system regulators in various cancer cell lines support this notion [181]. The antitumor effects of cannabinoids have also been observed in various animal tumor models [180]. In general, an enhanced endocannabinoid system is seen in tumor tissues [179,182,183]. However, the role of upregulated endocannabinoid system activity is still controversial as contrasting results have been reported supporting a proliferative as well as an anti-proliferative role of cannabinoids on cancer cells [180,181]. Interestingly, a bimodal effect of cannabinoids on cancer cell growth has also been observed, with low concentrations being proliferative and high concentrations being pro-apoptotic [184].

Organization, control and function of extrasynaptic NMDA receptors

N-methyl d -aspartate receptors (NMDARs) exist in different forms owing to multiple combinations of subunits that can assemble into a functional receptor. In addition, they are located not only at synapses but also at extrasynaptic sites. There has been intense speculation over the past decade about whether specific NMDAR subtypes and/or locations are responsible for inducing synaptic plasticity and excitotoxicity. Here, we review the latest findings on the organization, subunit composition and endogenous control of NMDARs at extrasynaptic sites and consider their putative functions. Because astrocytes are capable of controlling NMDARs through the release of gliotransmitters, we also discuss the role of the glial environment in regulating the activity of these receptors.

1. Introduction

N-methyl d -aspartate receptors (NMDARs) have long been of major interest to neuroscientists owing to their multiple implications in neuronal physiology during development and adulthood. Following the seminal discovery of long-lasting synaptic plasticity by Bliss & Lomo [1], NMDARs have been known for their key role in long-term potentiation (LTP) and long-term depression (LTD) of synaptic transmission [2]. However, because these glutamate-gated ion channels are highly permeable to calcium, they not only relay a physiological signal into neurons, but can also trigger intracellular signalling cascades that can ultimately lead to cell death. In fact, the phenomenon of excitotoxicity [3] and its causal link with excessive glutamate release [4] were described years before the existence of synaptic plasticity was established, but NMDARs were only demonstrated to be the main source of calcium responsible for glutamate-induced excitotoxicity in the late 1980s [5–7]. Ever since, these receptors have been renowned for their duality. They are both essential and potentially harmful receptors whose activity needs to be finely and tightly controlled. Such regulation is achieved by many intracellular and extracellular factors, including intracellular partners, magnesium block, endogenous allosteric modulators, as well as agonist and co-agonist binding. Interestingly, many of these regulatory actions are highly sensitive to the subunit composition of the receptor [8,9]. Along with the development of subunit-specific pharmacological antagonists, evidence for specific roles of NMDAR subtypes has thus accumulated over the past decades, with particular attention on the GluN2B-containing heterodimers [8,9]. In 1995, like other transmitter-gated ion channels, NMDARs were also found to be present in significant proportions at extrasynaptic sites [10–13]. Relationships between receptor location, their GluN2B content and their roles in the direction of synaptic plasticity and in cell death were soon assumed [14,15]. This generated a new craze for NMDARs and laid the groundwork for a meaningful understanding of how these receptors impact cellular physiology and pathology. More recently, NMDARs have gained a new interest, as d -serine, a potential gliotransmitter released by astrocytes ([16], but see [17]) was found to act as an endogenous co-agonist on their glycine-binding site [16,18–21]. This discovery has made glia a possible regulator of NMDAR function and marked the emergence of an exciting cross-talk between the field of NMDARs and the expending domain of glial cells. In this review, we discuss the latest findings on the organization, the endogenous control and the functions of NMDARs at extrasynaptic sites, and what role glia may play in the activity of these receptors.

2. Defining the extrasynaptic space

Making sharp delineation of synaptic and extrasynaptic spaces is challenging, especially when one tries to reconcile data from microscopy and electrophysiology. This difficulty arises from the fact that what one considers as extrasynaptic based on morphological clues can still be part of the synaptic space as far as signal transmission is concerned. From the morphological point of view, receptors are often considered extrasynaptic if they lie more than 100 nm away from the post-synaptic density (PSD) [22,23]. However, this morphological definition does not take into account glutamate spillover during synaptic transmission, which probably activates receptors located on the neck of dendritic spines (i.e. perisynaptic receptors) or on the dendritic shaft. From the electrophysiological point of view, a convention of synaptic physiologists defines synaptic NMDARs as those recruited during afferent stimulation at low frequency (less than 0.05 Hz), or in response to spontaneous glutamate release producing miniature excitatory responses [24–29]. Extrasynaptic NMDARs correspond to receptors that are not activated during such conditions [24–29]. This electrophysiological definition of synaptic NMDARs may encapsulate receptors located not only at synapses but also at peri- or extrasynaptic locations that possibly contribute, even mildly, to low-frequency synaptic transmission. Because it is difficult to determine to what extent glutamate spills over during basal synaptic transmission, it is unclear to what degree the morphological and electrophysiological definition of the synaptic and extrasynaptic spaces overlap. In fact, the precise delineation of the synaptic–extrasynaptic frontier seems specific to the parameter that one considers, and it is likely that an all-encompassing definition of this boundary cannot be found.

3. NMDAR organization at extrasynaptic sites

Despite some occasional observations of NMDARs at extrasynaptic locations [10–13,30–36], the first study to directly and extensively investigate the organization of post-synaptic NMDARs at extrasynaptic sites was published in 2010 [22]. Interestingly, it seems that extrasynaptic NMDARs are organized as clusters, with a discrete distribution along the perisynaptic zone (45%) and dendritic shaft (55%) [22]. In particular, these clusters often correspond to contact area between a dendrite and adjacent cell processes, including glial (approx. 25%), axonal-like (approx. 50%) or dendritic (approx. 25%) processes. These extrasynaptic sites spread to an average width of less than 100 nm, which means they are substantially narrower than synaptic sites that have an estimated width of 195 nm in the CA1 region [37]. Interestingly, extrasynaptic NMDAR clusters seem to colocalize with intercellular adhesion molecules such as β-catenins and some scaffolding proteins such as SAP102 and PSD95 in culture [22], although it is not clear in this work whether any specific MAGUKs are preferentially interacting with extrasynaptic NMDARs. These observations strongly suggest that extrasynaptic NMDARs are clustered and organized at the plasma membrane in a similar manner to their synaptic homologues. NMDARs are neither evenly nor randomly distributed along the dendritic shaft, and probably are not as free to diffuse as demonstrated in cultures [38–40]. This challenges the common idea that extrasynaptic NMDARs are essentially a reserve pool of receptors, waiting to be recruited to synapses. In fact, the finding that these receptors are clustered to very specific cell contact areas suggests that they may serve a distinct signalling function, independent from their synaptic homologues. The work from Harris & Pettit [26] already supported this idea by demonstrating that synaptic and extrasynaptic NMDARs represent distinct and stable pools of receptors at the surface of CA1 neurons in hippocampal slices.

4. GluN2B-NMDARs at extrasynaptic sites

It is generally believed that extrasynaptic NMDARs are enriched in GluN2B-containing heterodimers (GluN2B-NMDARs). This would fit with the idea that extrasynaptic NMDARs constitute a distinct population, serving a specific function. This idea is also quite satisfying when one considers that GluN2B-NMDARs have a higher affinity for glutamate, whose concentration is lower in the extrasynaptic space [8,9,41,42]. Finally, this localization of GluN2B-NMDARs would also make sense regarding the interaction of NMDARs with their intracellular partners. Indeed, the C-terminal domain (CTD) of GluN2 subunits dictates the scaffolding protein the receptor can interact with and, thus, strongly influences the surface localization of NMDAR subtypes (as reviewed in great detail in [8,9]). It was proposed that GluN2A-NMDARs associate preferentially with PSD95, which is believed to limit their field of action to the PSD and increase their half-life time before recycling [38,43–47]. Because they tend to interact with SAP102 rather than with PSD95, GluN2B-NMDARs are thought to be more mobile, widespread and not confined to synapses in adults [45,47]. In fact, the expression of GluN2A and GluN2B subunits closely follows the expression of PSD95 and SAP102, respectively, during development [47], and evidence suggests that (i) neonatal GluN2B-NMDARs are held at synapses through their interaction with SAP102 during early stages of development (ii) the expression of PSD95 increases throughout postnatal development and competes with SAP102 for insertion into the PSD and (iii) this displaces SAP102 and GluN2B subunits outside of synapses, whereas PSD95 and GluN2A become predominant at synaptic sites [22,45,48,49]. Along with the fact that neurons of the visual cortex and superior colliculus still show significant expression of SAP102 and GluN2B subunits in adulthood [45,47], these observations together strongly fuel the idea that GluN2B-NMDARs exist at extrasynaptic sites more abundantly than GluN2A-NMDARs, which are preferentially found at synapses.

However, only a few studies have functionally investigated this aspect in brain slices, and very often indirectly. Harris & Pettit [26], who specifically tackled that question, did not find a particular enrichment of GluN2B-NMDARs at an extrasynaptic location compared with synapses, in the CA1 region of rat hippocampus, as they identified large amounts of GluN2B-NMDARs at both locations. By contrast, Fellin et al. [50], who performed pioneering work on extrasynaptic NMDARs activated by glia-derived glutamate, observed a relative enrichment of GluN2B-NMDARs outside of synapses, in line with similar observations [51–53]. Unfortunately, these studies were all performed in slices obtained from young rodents (P10–P22), at a time during development when the switch from GluN2B- to GluN2A-subunit-containing NMDARs is still occurring at synapses [9]. For instance, Harris and Pettit found, as expected, that GluN2B-NMDARs disappear from synapses throughout the first month of age, but regrettably carried out the comparison of synaptic versus extrasynaptic composition in juvenile animals (P15) [26] and concluded there was no difference in the GluN2B content. Additionally, it is not known whether extrasynaptic NMDARs also undergo a developmental change of their subunit composition. Hence, because it has been mostly conducted in slices from immature animals or in cultures, the study of NMDAR subunit composition at synaptic and extrasynaptic sites has, so far, essentially led to contradictory results and confusion.

Even when tackled in adult rodents, the situation does not seem to be any clearer. In hippocampal CA1 pyramidal neurons of adult rats, GluN2B-NMDARs were shown to be absent at synapses, but could still be found at extrasynaptic locations [19]. However, the replacement of GluN2B- with GluN2A-NMDARs at synapses is not ubiquitous and complete, and GluN2B-NMDARs still represent a major portion of synaptic NMDARs in many other regions of the adult central nervous system (CNS) [16,54] as reviewed in [9]. Therefore, even though it is broadly accepted, the idea that extrasynaptic receptors are essentially GluN2B-NMDARs, or are at least enriched in GluN2B-NMDARs compared with their synaptic counterpart, is still open for debate.

5. Agonist control of NMDAR at extrasynaptic sites

In 1989, the existence of a tonic current mediated by NMDARs was reported in principal neurons of the hippocampus under mild depolarization [55]. Similar to what is known for GABAR-mediated tonic currents [56], NMDAR-mediated tonic current seems to be allowed by sufficient concentrations of ambient agonist [42,55] and generated by receptors located outside of synapses [55,57,58]. Indeed, NMDAR-mediated tonic current persists when synaptic activity is suppressed with tetrodotoxin, and conversely synaptic NMDAR activity remains unaffected after NMDARs mediating tonic current are blocked with MK801 [58], which establishes those receptors as distinct, and somehow distant, from synaptic receptors. However, the origin of the so-called ambient glutamate that allows such tonic activation is still unclear, and little evidence for the source exists, aside from the observation that it is of non-synaptic origin [58,59]. Interestingly, the amplitude of NMDAR-mediated tonic current is strongly enhanced either by blocking glutamate reuptake through the glutamate transporter 1 (GLT1) [58,60], which is predominantly present on glial cells and mediates 95% of glutamate transport, or when slices are challenged with glial toxins [60]. Additionally, a recent study showed that the extent of NMDAR-mediated tonic current depends on glial coverage in the supraoptic nucleus of the hypothalamus [60]. Hence, the activity of NMDARs that mediate tonic current probably relies on the regulation of extrasynaptic glutamate tone by glia. The corollary of these observations is that this subset of extrasynaptic NMDARs is constitutively activated. However, extrasynaptic NMDARs can also be recruited experimentally by exogenous applications of NMDA or glutamate. This indicates that another subpopulation of extrasynaptic NMDARs exists (but see below) that can be activated by phasic, ‘ectopic’ release of glutamate from non-active zone sites. The main manifestation of such phasic activity of extrasynaptic NMDARs is slow (decay time on the order of seconds), infrequent (approx. 0.05 Hz) inward currents of large amplitude (hundreds of pA), generated in principal neurons of the CA1 region of the hippocampus [50,51,53] and in the superficial layers of the dorsal horn [52,61]. Such currents, termed slow inward currents (SICs), persist in the absence of neuronal and/or synaptic activity [51] and are completely suppressed by glial inhibitors [52]. Evidence points to the fact that such currents are caused by glutamate release from glia onto extrasynaptic NMDARs [33,41,50,62,63]. However, the exact pathway involved in this process is still under investigation. It has been shown that SIC frequency and amplitude increase upon activation of astrocytic G-protein coupled receptors (GPCRs) in the hippocampus [50,51], such as PAR1 receptors [53]. Concomitantly, stimulating the PAR1 receptor on astrocytes was recently shown to trigger calcium-dependent release of glutamate through Best1 channels in the hippocampus [64–66]. Thus, SICs could be caused by astrocyte-derived glutamate release through Best1 channels onto extrasynaptic NMDARs, upon activation of PAR1. Although this appealing theory would fit with earlier observations that calcium activity in the astrocytes is needed for SICs to occur [50,53,61], it would be in striking contradiction to the idea that vesicular loading from astrocytes is required [50]. Additionally, whether endogenous PAR1 activity can trigger SICs without exogenous experimental stimulation is unclear. Other pathways could thus be at play, involving metabotropic glutamate receptors, prostaglandin receptors or other astrocytic GPCRs, and relying on vesicular release [50,51].

Besides their mechanistic relevance, these observations are interesting because they open the possibility that extrasynaptic NMDARs may exist in two functionally distinct pools: (i) tonically activated receptors, which face suprathreshold amounts of glutamate and co-agonist, and (ii) overall silent extrasynaptic NMDARs, which can be acutely recruited by exogenous or endogenous (non-synaptic) glutamate release. If true, this could mean that glutamate availability is not homogeneous outside of synapses and could be spatially regulated, giving rise to different subcompartments. Another interpretation, however, could be that tonic and phasic extrasynaptic NMDAR responses are mediated by the same receptor population: tonic current could be carried out by these receptors under partial occupancy of their glutamate binding site, and additional (i.e. phasic) glutamate release would cause transient saturation and recruitment of supplementary receptors. Although this is purely speculative, we here propose that it could be relevant to consider the extrasynaptic space, at least from the point of view of glutamate, as comprising separate domains instead of one large homogeneous volume (figure 1). Understanding the rules that govern glutamate release, uptake and diffusion in the extracellular space would certainly shed light onto this situation and onto the control and roles of extrasynaptic NMDARs.

Figure 1. Simplified view of NMDAR organization at extrasynaptic sites. Astrocytic processes uptake glutamate through GLT1 and glycine through GlyT1 while they release d -serine ( d -ser) at synapses. Astrocytic GLT1 also lowers glutamate concentration near extrasynaptic NMDARs (ambient glutamate concentrations are depicted in fades of blue: dark blue represents higher concentrations of glutamate). This could allow acute activation of such receptors through vesicular (ves.) or Best1-mediated release of glutamate by astrocytes, upon activation of astrocytic GPCRs, causing SICs in the post synaptic neuron. On the other hand, extrasynaptic NMDARs that are not in close proximity to glial processes and/or GLT1 may face suprathreshold concentrations of glutamate and mediate tonic current. Literature has suggested that extrasynaptic NMDARs could interact with different intracellular partners (SAP102) than their synaptic homologues (PSD95), but this aspect is not completely clear. Similarly, the contribution of synaptic versus extrasynaptic NMDARs to excitotoxicity is still controversial. On the contrary, it seems that synaptic NMDARs contribute to LTP while both extrasynaptic and synaptic NMDARs are required for LTD. The presence of adhesion molecules (such as β-catenins) is depicted. NMDAR subunit composition was purposefully omitted to avoid adding confusion to the debate (see main text).

6. Co-agonist control of extrasynaptic NMDARs

As demonstrated more than 25 years ago, NMDARs have the unique property of requiring not one but two agonists to be activated [67]. The identity of the second agonist, termed co-agonist, was elusive until recently, even though glycine was originally shown to bind to the co-agonist site [67–69]. In the CA1 region of the adult rat hippocampus, it was demonstrated recently that d -serine is the endogenous co-agonist of synaptic NMDARs, but that glycine gates the extrasynaptic NMDARs that mediate the tonic current and contribute to the response induced by exogenous application of NMDA [19]. SICs were not investigated in this study and whether NMDARs that mediate SICs are also gated by glycine, or by d -serine, is still unknown. This is potentially of great interest, because there is no actual demonstration that SICs are caused by the release of glutamate. Instead, SICs could be generated by the release of d -serine or glycine by astrocytes, on NMDARs that face suprathreshold levels of ambient glutamate. This hypothesis, which has never been considered, fits with the body of data already available on SICs (but see below considerations about glycine concentrations in the extracellular space) and would explain why blocking glutamate transporters does not necessarily increase SIC amplitude or frequency, as if SIC-mediating NMDARs were already facing saturating levels of glutamate [50].

Interestingly, the difference in the co-agonist used by NMDARs at synaptic and extrasynaptic sites appears primarily to be based on the fact that d -serine is released at synapses and that the activity of glial glycine transporter 1 (GlyT1) prevents glycine from accessing the synaptic cleft [19]. Thus, the delineation between synaptic and extrasynaptic space, with regard to the co-agonist control of NMDARs, seems to be defined by d -serine and glycine abundance. It is unclear, however, whether the domains delineated by the co-agonists match the domains defined by glutamate spillover (see section ‘Defining the extrasynaptic space’).

Even more unclear is the origin of glycine available to extrasynaptic NMDARs. In structures where glycinergic innervation is abundant, such as in the retina and the spinal cord, glycine that serves as an endogenous co-agonist at NMDARs originates from glycinergic terminals [70,71]. In those preparations, it was demonstrated that glycine released at inhibitory synapses spills over and diffuses to bind to remote NMDARs. However, in the adult hippocampus, even though the existence of extrasynaptic glycine receptors (GlyRs) has been documented [72,73], the presence of glycinergic terminals has never been established. In fact, the inhibitory transmission is entirely abolished by GABA receptor antagonists in that structure [74], and the expression of functional GlyRs is thought to stop after birth [74–76]. Yet, microdialysis has revealed that the amounts of free glycine in vivo are as high as 10 µM in the hippocampus [77,78]. In vivo, a major source of extracellular glycine in the CNS could be the blood flow. Indeed, blood contains approximately 200 µM of glycine [79] and glycine is able to cross the endothelial wall of capillaries by mean of glycine transporters [80,81]. In slices, however, blood vessels are mostly emptied of their initial contents, narrowing down the source of glycine to the brain parenchyma itself. Because they contain 3–6 mM of glycine [82], glial cells could be a major source of glycine in brain slices. It was shown that potassium stimulates the release of glycine by glial cells in culture, through a calcium-independent mechanism [83–85]. In addition, GlyT1 activity can be reversed in certain conditions, leading to an efflux of glycine into the extracellular space [86,87]. For instance, in the cerebellum, if Bergmann glial cells are infused with 4 mM glycine, which is the estimated concentration in hippocampal astrocytes [82], and submitted to a mild membrane depolarization (approx. −50 mV), GlyT1 becomes responsible for an efflux of glycine. In the hippocampus, another noteworthy source of glycine could be GABA interneurons. Song et al. [88] indeed demonstrated in the CA1 region that GABA interneurons contain glycine and express glycine transporter 2 (GlyT2). They proposed that glycine could be co-released with GABA at inhibitory synapses similar to what occurs in the thalamus, the brainstem, the spinal cord and the cortex [89–94].

Regardless of glycine origin, in vivo microdialysis suggests that the amount of glycine available to extrasynaptic NMDARs should be high enough to be saturating [77,78], especially if those receptors are GluN2B = NMDARs (EC50 < 1 µM). In line with this idea, it was observed that exogenous applications of co-agonist do not enhance the amplitude of NMDAR-mediated tonic current [58]. If true, it means there is little room for regulatory mechanisms and modulation of NMDAR activity through the co-agonist-binding site at extrasynaptic locations, but this has never been addressed for SICs or any other manifestation of extrasynaptic NMDAR activity. This puzzling aspect thus needs further investigation.

7. Role of slow inward currents and tonic current in physiology and pathology

The roles played by extrasynaptic NMDARs are still very elusive, mostly because specific and reliable tools to investigate them have been lacking so far. Some studies have reported a role for non-synaptic NMDARs in various functions such as extrasynaptic inhibition [95], or dynamic range compression [96]. But, aside from presynaptic NMDARs [97,98], which are not addressed in this review, most studies have focused on the manifestations of extrasynaptic NMDAR activity such as tonic currents and SICs. It is accepted that the tonic current mediated by extrasynaptic NMDARs plays a role in the excitability of principal neurons and in modulating dendritic inputs [99]. It also becomes increasingly clear that NMDAR-mediated tonic current can be upregulated in pathological conditions, ranging from cocaine addiction [100] to Alzheimer's disease [101]. However, the relevance of such tonic current to neuronal physiology and the impact of its pathological disturbance need further investigation. In contrast, SICs have been quite well characterized (see above) and were shown to occur simultaneously in distinct neurons within 100 µm of each other [51]. This fits with the idea that SICs are generated by glutamate (or co-agonist) release from astrocytes whose anatomical territory spans an area approximately 50–100 µm in diameter. Based on these observations, SICs have been proposed to play a role in neuronal synchrony and network excitability [51,52]. However, they have only been observed in slices so far, often under non-physiological conditions (low magnesium, GLT1 blockers). Whether SICs represent an important feature of neuron–glia signalling, that also occurs in vivo, remains to be established, but it is worth mentioning that SIC amplitude was demonstrated to be enhanced in animal models of inflammation [61].

8. Long-term potentiation and long-term depression: on the importance of subunit composition and/or localization?

Using both genetic and pharmacological approaches, there has been intense speculation and many demonstrations that specific NMDAR subtypes are responsible for selectively inducing LTP or LTD [102–107]. It was proposed that GluN2A-NMDARs would preferentially trigger LTP, whereas GluN2B-NMDARs are preferentially associated with LTD. This dichotomy, however, has been significantly and repeatedly challenged [106,107] (table 1), and the idea that a particular subtype of NMDAR is specifically involved in inducing potentiation or depression at glutamatergic synapses seems over-simplifying, as reviewed in great detail by Paoletti et al. [9]. This raises the question of whether NMDAR location, rather than subunit composition, could be the determining factor to the direction of synaptic plasticity [117]. Unfortunately, very few studies have directly addressed this aspect. Instead, it seems that the role of synaptic and extrasynaptic NMDARs in plasticity have mostly been inferred from results obtained on the role of NMDAR subunit composition (in particular, GluN2B-NMDARs) in LTP and LTD. However, there seems to be no simple relationship between cellular location and subunit composition in particular at ages most often used in those studies (approx. P20–P25, see discussion above). It is thus notoriously hazardous to form conclusions from works studying the involvement of extrasynaptic NMDARs in synaptic plasticity (and, in fact, in any specific functions) based on their GluN2B pharmacological profile. Activating or silencing specifically synaptic or extrasynaptic NMDARs thus requires other types of approaches that do not rely on subunit composition. The main alternative is the use of the open-channel blocker MK801, which allows inactivation of NMDARs that are recruited while leaving the silent/not recruited receptors intact. When combined with low-frequency stimulation of afferent fibres, this method allows the selective blockade of synaptic NMDARs and leaves intact most of their extrasynaptic counterparts [19,26,27,118,119]. Conversely, this approach also allows the blockade of extrasynaptic NMDARs mediating tonic currents, while leaving intact unstimulated synaptic NMDARs [58]. Either way, this method presents the advantage that it only relies on whether receptors are active or not during MK801 application, and therefore circumvents caveats associated with the use of subunit-selective reagents or immature animals. Interestingly, using such an approach, Liu et al. [120] found that selective stimulation of extrasynaptic NMDARs triggers LTD in CA1 neurons. Alternatively, it was recently demonstrated that d -serine and glycine gate synaptic and extrasynaptic NMDARs, respectively [19]. Taking advantage of this segregation, the authors were able to show in adult rat hippocampal slices that synaptic, but not extrasynaptic, NMDARs are required for LTP induction, whereas activation of both types of receptors is required for LTD. These findings fuel the idea, proposed by Rusakov et al. [117], that what matters in LTP and LTD induction may be the location, rather than the subtype, of NMDARs recruited.

Table 1. Role of synaptic and extrasynaptic NMDA receptors in excitotoxicity in cell culture, brain slices and in vivo. Studies were organized by the date of publication. OGD, oxygen glucose deprivation NVP, NVP-AAM077, a NMDAR antagonist with mild preference for GluN2A-NMDARs DL-TBOA, blocker of the glutamate transporter 1 (GLT1) 4-AP, 4-aminopyridin, blocker of voltage-gated potassium-channels.

a Katsuki et al., as well as Shleper et al., did not specifically investigate the role of synaptic and extrasynaptic NMDARs, but rather the role of d -serine-gated NMDARs in excitotoxicity.