Dissecting Spatial Knowledge from Spatial Choice by Hippocampal NMDA Receptor Deletion

Abstract

Hippocampal NMDA receptors (NMDARs) and NMDAR-dependent synaptic plasticity are widely considered as crucial substrates of long-term spatial memory, although their precise role remains uncertain. Here we show that GluN1^ΔDGCA1 mice, lacking NMDARs in all dentate gyrus and dorsal CA1 principal cells, acquired the spatial reference memory watermaze task as well as Controls, despite impairments on the spatial reference memory radial maze task. In a novel spatial discrimination watermaze paradigm, using two visually identical beacons, GluN1^ΔDGCA1 mice were impaired at using spatial information to inhibit selecting the decoy beacon, despite knowing the platform’s actual spatial location. This failure could suffice to impair radial maze performance despite spatial memory itself being normal. Thus, these hippocampal NMDARs are not essential for encoding or storing long-term, associative spatial memories. Instead, we demonstrate an important role for the hippocampus in using spatial knowledge to select between alternative responses that arise from competing or overlapping memories.

The ability to associate spatial locations with particular events is essential for the survival of all animals. It is widely believed that NMDAR-dependent synaptic plasticity contributes to the formation of associative memories¹, and hippocampal NMDARs, particularly those in the dorsal CA1, are thought to be important for associative, long-term spatial memory², although their precise role remains unclear. Indeed, as recently noted, “even though the synaptic plasticity and memory hypothesis is enshrined in most neuroscience text books, this issue is still far from resolved”³. Pharmacological studies with NMDAR antagonists are equivocal^4-6, and previous evidence from transgenic mice lacking NMDARs in dorsal CA1^7,8 is confounded by the spread of NMDAR deletion to principal neurons in cortex, as evidenced in recent studies employing this transgenic line^9-13.

Mice lacking NMDARs selectively in dentate gyrus (DG) granule cells, and hence long-term potentiation (LTP) in lateral and medial perforant path synapses, exhibit normal long-term spatial memory^14,15. We have now generated mice (GluN1^ΔDGCA1) in which NMDARs are additionally lacking in dorsal and, to a lesser extent, ventral hippocampal CA1 pyramidal cells, specifically in adulthood, thus providing an important new mouse model for assessing the hippocampal NMDAR and memory hypothesis. Here we investigated GluN1^ΔDGCA1 mice on tests of associative long-term spatial memory. Mice were assessed on reference memory versions of the watermaze and radial maze tasks, which both depend on dorsal hippocampus^16-21.

RESULTS

DG and CA1 specific NMDAR deletion in GluN1^ΔDGCA1 mice

We employed a mouse line (GluN1^ΔDGCA1, Fig. 1a) gene-targeted for loxP-tagged Grin1 alleles and carrying two transgenes Tg^LC1 and Tg^CN12, which enabled doxycycline (dox)-sensitive, Cre-mediated GluN1 ablation in excitatory hippocampal, but not cortical, neurons of adults^22,23 by use of a CaMKII promoter fused to a Grin2c silencer element²⁴. Selective hippocampal NMDAR removal required switching off Cre expression by dox during embryogenesis and nursing pups with mothers taken off dox post-natally. Delayed Cre expression, detected four weeks post-natally (Fig. 1b,c), eventually reached a peak in all DG granule cells, whereas in CA1, a Cre expression gradient formed along the dorso-ventral axis (Fig. 1d, Supplementary Fig. 1h). Specificity of cumulative Cre expression was assessed by X-gal staining in double transgenic Cre indicator mice (Tg^CN12/LC1/Gt(Rosa)26Sor) which demonstrated strong Cre-induced β-galactosidase activity throughout the entire DG and mossy fibers, as well as dorsal and, to a lesser extent, ventral CA1 (Fig. 1b,c, Supplementary Fig. 1d-g). All other parts of the hippocampal formation (CA3, lEnt, mEnt, Sub) exhibited negligible recombination (<1% co-labeling; Supplementary Fig. 1d-g), although β-galactosidase was observed in olfactory bulb granule cells and approximately 30% of layer II piriform cortex neurons (Fig. 1b,c, Supplementary Fig. 1d).

Figure 1. Removal of NMDARs in DG and CA1.

(a) Genes to generate conditional NMDAR^−/− mice (GluN1^ΔDGCA1) by postnatal dox-controlled activation of Tg^CN12/LC1-encoded Cre in Grin1^tm1Rsp mice (GluN1^2lox). LoxP sites are depicted by black triangles. Exons appear as boxes, and when black, encode membrane-spanning segments. (b) Anti-Cre immunostainings of brain sections from dox treated (till P0) Tg^CN12/LC1 mice at postnatal days P28 and P175, and β-galactosidase stainings (blue) of sections from Tg^CN12/LC1/Gt(Rosa)26Sor mice (mf, mossy fibers; pi, piriform cortex with ~30% of layer II neurons expressing Cre). (c, d) Dox-regulated Cre expression in the brains of Tg^CN12/LC1 mice. (c) Expression pattern of the nuclearly localized Cre in CA1 pyramidal cells when dox had been removed from the drinking water of mothers at the day of delivery. At P28 (n=2 mice), P45 (n=2) and P150 (n=6), the extent of Cre expressing pyramidal neurons in the dorsal CA1 layer was estimated by α-Cre immunostains (green), counterstained with α-NeuN (red). The inset shows Cre-activated β-galactosidase expression in CA1 and DG of the hippocampus and the olfactory bulb (OB) in a brain section from Tg^CN12/LC1/Gt(Rosa)26Sor mice. (d, left), Horizontal brain sections of Tg^CN12/LC1 mice immunostained with α-Cre (green) and α-NeuN (red) at three different levels from bregma (– 1.4, – 2.8, and – 4.0 mm) showed a yellow gradient of Cre/NeuN immunoreactivity in the dorsal to ventral str. pyramidale of CA1, but not in the DG. The percentage of Cre positive cells among >1,000 NeuN positive neurons was evaluated at various depths in dorsal and ventral hippocampus, in slices from four P260 GluN1^ΔDGCA1 mice, and is listed as mean ± SD. (d, right) High magnification examples of α-Cre/α-NeuN co-labeled CA1 and DG neurons (yellow) from sections given on the left. The transition between dorsal and ventral hippocampus in the rostro-caudal axis had the highest variance of Cre expressing NeuN positive cells (middle insets). Other parts of the hippocampal formation feature negligible numbers of Cre-positive cells (<1% co-labeling). (e) In situ hybridization with GluN1 and GluA1 probes. Scale bars, 2mm. (f) α-NeuN and α-Calbindin immunostainings.

Efficient loss of GluN1 expression in dorsal CA1 and DG was confirmed by in situ hybridization, which also revealed reduced expression of the AMPAR subunit GluA1 in the DG granule cell layer (Fig. 1e, Supplementary Fig. 2). Further analysis indicated that in older mice the volume of the DG was reduced (~50% at age >P365). NeuN, Calbindin, GFAP and Parvalbumin expression patterns showed that the upper plate of the DG granule cell layer was more compromised than the lower plate, exhibiting gliosis in one-year old mice (Fig. 1f, Supplementary Fig. 3). Thus, chronic removal of NMDARs in these hippocampal areas resulted in thinning of the stratum granulare in DG, which might, in part, be related to a recently delineated role for NMDARs in the adult DG²⁵.

GluN1^ΔDGCA1 mice appeared normal and were visually indistinguishable from their littermate controls. However, they were hyperactive when placed in a novel environment (Supplementary Fig. 4).

Electrophysiological analysis

As expected, no NMDAR-mediated responses could be evoked at CA3-to-CA1 synapses in dorsal hippocampal slices from adult GluN1^ΔDGCA1 mice (Fig. 2a).

Figure 2. Loss of functional NMDA receptors at CA3-to-CA1 synapses in the dorsal CA1 region.

(a) Absence of NMDA responses in field recordings in acute brain slices from 10 to 12 month old GluN1^ΔDGCA1 mice: Blue traces show the average (five repetitions at 0.1 Hz) of fEPSPs elicited by three synaptic activations at 100 Hz in the presence of the AMPA receptor blocker DNQX (20μM) in control (upper panel) and GluN1^ΔDGCA1 (lower panel) mice. Red traces are the corresponding responses following subsequent superfusion with AP5 (50μM) to block NMDA receptors. Mean representative traces (averaged across 5 stimulations both before and after AP5) from a single slice from both a control and a GluN1^ΔDGCA1 mouse. Experiments were repeated in 3 control (n=12 slices) and 2 GluN1^ΔDGCA1 mice (n=8 slices). (b) Absence of LTP: Normalized and pooled fEPSP slopes evoked at CA1 radiatum and oriens synapses in slices from both control (open symbols) and GluN1^ΔDGCA1 mice (filled symbols). Forty to 45 min after the last tetanization of the afferent fibres in the stratum radiatum in slices from five Controls, the average slope of the field EPSP (open circles) was 1.60 ± 0.07 (mean ± SEM; n=18 slices) of the pre-tetanic value, whereas the untetanized oriens pathway (triangles) was unchanged (0.97 ± 0.04). In slices from four GluN1^ΔDGCA1 mice LTP was completely abolished (0.99 ± 0.03, n=14 slices), and evoked responses were not significantly different from those in the untetanized pathway (1.03 ± 0.04; p = 0.35). (c) LTP at CA3-CA3 and CA3-CA1 synapses: normalized and pooled fEPSPs before and after LTP induction at CA3-CA3 (open circles) and CA3-CA1 (filled circles) radiatum synapses. In slices from three control mice (upper panel), LTP was well-developed 40 – 45 min post-tetanization, with similar magnitudes in the two regions (CA3; 1.33 ± 0.12, CA1; 1.37 ± 0.11; n=11 slices; p = 0.77). The same experimental paradigm performed on slices from three GluN1^ΔDGCA1 mice (lower panel) showed preserved LTP in CA3 whereas LTP in CA1 failed to develop (CA3; 1.20 ± 0.07, CA1: 0.94 ± 0.19; n=8 slices; p = 0.01). All means ± s.e.m. Arrows at the abscissa indicate the time points of tetanic stimulation. Insets show means of six consecutive synaptic responses in the tetanized pathway before and 45 min post-tetanization.

Extracellular field LTP experiments confirmed the loss of functional NMDARs in the Schaffer collateral/commissural pathway in dorsal CA1. In GluN1^ΔDGCA1 mice, LTP failed to develop after four tetanizations (Fig. 2b).

In subsequent experiments, a stimulation electrode was placed in stratum radiatum, at the border between CA1 and CA3. Stimulation at 0.1 Hz elicited synaptic responses recorded simultaneously by two glass electrodes localized to stratum radiatum in CA1 and CA3. In Controls LTP was simultaneously well developed in both pathways 40–45 min after repeated tetanization, with a similar magnitude in the two subfields (Fig. 2c, upper panel). In GluN1^ΔDGCA1 mice LTP in the CA3 subfield was preserved, whereas CA3-to-CA1 LTP was absent (Fig. 2c, lower panel), providing direct evidence for loss of NMDAR-mediated function in CA1. Thus, GluN1^ΔDGCA1 mice provide a novel means of assessing the hippocampal NMDAR-dependent synaptic plasticity and memory hypothesis. Associative, long-term spatial memory was therefore assessed in GluN1^ΔDGCA1 mice using two classic, hippocampus-dependent tests (Supplementary Fig. 5).

Spatial reference memory in the Morris watermaze

Mice with hippocampal lesions are unable to acquire the fixed location, hidden platform version of the watermaze (Supplementary Fig. 5a-c). In contrast, experimentally naïve GluN1^ΔDGCA1 mice learned the spatial location of the platform as well as Controls (Fig. 3). Mice exhibited reduced pathlengths to find the platform as training proceeded but there was no impairment in GluN1^ΔDGCA1 mice (ANOVA; main effect of block - F (8,176) = 62.68; p <0.0001; no main effect of genotype - F < 1; no genotype by block interaction - F (8,176) = 1.64; p >0.10; Fig. 3a, Acquisition). During probe tests in which the platform was removed from the pool, hippocampal lesioned mice characteristically searched equally across all four quadrants of the pool (Supplementary Fig. 5c), but in the first probe test (24 h after trial 24) both Controls and GluN1^ΔDGCA1 mice exhibited strong spatial memory and searched selectively at the training platform location (Fig. 3b). ANOVA revealed only a significant main effect of quadrant (F (2,66) = 49.66; p <0.05). In the second probe test (24 h after trial 36), GluN1^ΔDGCA1 mice showed stronger spatial memory, and spent significantly more time in the training quadrant than Controls (Fig. 3c). ANOVA revealed a main effect of quadrant (F (2,66) = 68.74; p <0.05), and a genotype by quadrant interaction (F (2,66) = 3.27; p <0.05). There was a significant group difference in time spent in the training quadrant (t-test; t (22) = 2.24; p <0.05).

Figure 3. Associative long-term spatial reference memory in the Morris watermaze.

(a) Similar pathlengths for Controls (n=12) and GluN1^ΔDGCA1 mice (n=12) during Acquisition of the fixed location, hidden escape platform watermaze task (Acquisition; geno – F<1, geno x block – F(8,176)=1.64; p>0.10; 4 trials/block). When the platform was moved to the opposite quadrant of the pool (Reversal), the pathlengths of GluN1^ΔDGCA1 mice were significantly longer than Controls (geno – F(1,22)=13.41; p<0.005). All means ± s.e.m. (b, c) Long-term memory performance during probe trials (Transfer tests) after 6 and 9 training blocks. Percentage time spent (left panel) and numbers of annulus crossings (right panel) in each quadrant (left to right; adjacent left, goal (G), adjacent right, opposite) for Controls and GluN1^ΔDGCA1 mice. In transfer test 2 (TT2), GluN1^ΔDGCA1 mice searched longer in the training quadrant than Controls (t(22)=2.24; * p<0.05). Broken line: chance levels of performance. In addition, the numbers of annulus crossings in each quadrant are given.

On the next day the platform was moved to the opposite quadrant of the pool for a further 3 days of training. Whereas initial acquisition had been normal in GluN1^ΔDGCA1 mice, during reversal learning they were significantly impaired, taking longer paths to the new platform position (ANOVA; main effect of genotype - F (1,22) = 13.41; p < 0.005; main effect of block - F (2,44) = 36.54; p < 0.0001; no genotype by block interaction - F (2,44) = 2.24; p > 0.10; Fig. 3a, Reversal).

Spatial reference memory on the radial maze

Although GluN1^ΔDGCA1 mice displayed normal acquisition in the watermaze, they were, like mice with hippocampal lesions (Supplementary Fig. 5d), substantially impaired on acquisition of a radial maze task, in which they had to discriminate between three always-baited arms and three never-baited arms (Fig. 4). GluN1^ΔDGCA1 mice made more reference memory errors (ANOVA; main effect of genotype - F (1,21) = 30.42; p< 0.0001; main effect of block - F (11,231) = 16.08; p<0.0001; no genotype by block interaction - F (11,231) = 1.03; p>0.40; Fig. 4a). The number of errors into the single, non-rewarded arm and the pair of adjacent, non-rewarded arms (allowing for the number of arms) were also compared (Fig. 4b). Mice were more likely to make errors into the single, non-rewarded arm than into one of the adjacent, non-rewarded arms (ANOVA; main effect of error type – F(1,21) = 7.06; p < 0.02). However, this pattern was the same for both genotypes (genotype x error type interaction – F < 1; p > 0.90).

Figure 4. Associative long-term spatial reference memory on the radial maze task.

(a) Inset: Mice were trained to discriminate between three rewarded (+) and three un-baited (−) arms. Errors per trial for Controls (n=12) and GluN1^ΔDGCA1 mice (n=11) across 12 blocks of testing (4 trials per block). GluN1^ΔDGCA1 mice were impaired at discriminating between the baited and non-baited arms, making significantly more spatial reference memory errors (F (1,21) = 30.42; p<0.0001). (b) Error types during spatial reference memory acquisition. Total number of errors into the single, non-rewarded arm (Sin) and the total number of errors into the pair of adjacent (Adj/2), non-rewarded arms. The number of errors into the pair of adjacent, un-baited arms was divided by two to correct for the number of arms. All means ± s.e.m.

The dissociation between these two classic spatial reference memory tests (watermaze vs. radial maze) did not reflect differences in sensorimotor, motivational or arousal demands of these tasks²⁶. The behavioral difference between genotypes was also present in both appetitive and aversive (swim/escape) versions of a three-arm radial maze (Y-maze) reference memory task. In both versions of this Y-maze task GluN1^ΔDGCA1 mice were impaired during acquisition (Supplementary Fig. 6). Furthermore, both groups acquired an appetitively-motivated, visual discrimination task (grey vs. black/white striped goal arms) at a comparable rate (Fig. 5). ANOVA revealed a main effect of block (F (8,176) = 36.18; p < 0.0001), but no main effect of genotype (F < 1), and no genotype by block interaction (F (8,176) = 1.07; p > 0.30). They were also indistinguishable during subsequent reversal of this non-spatial task (Supplementary Fig. 7).

Figure 5. Preserved visual discrimination learning in GluN1^ΔDGCA1 mice.

In the visual discrimination task both Controls (n=13) and GluN1^ΔDGCA1 mice (n=11) learned to associate floor/wall color (grey vs. black/white stripes) with a reward (+) independent of its spatial position in the arms of the maze. The top part of the figure depicts the T-maze with the start arm (S) and both the rewarded (+) and non-rewarded (−) goal arms, which can be arranged in either of two possible configurations. There were 9 blocks of training (10 trials per block). Mean ± s.e.m. percent correct choices for Controls and GluN1^ΔDGCA1 mice. Broken line: chance levels of performance.

Spatial discrimination using a beacon watermaze task

Instead, the dissociation likely reflects the different psychological processes involved in these watermaze and radial maze tasks. On the radial maze, all arms have the same physical appearance (the intramaze cues are constant). They have the same Perspex door, the same grey floor and sidewalls, and the same foodwell. When mice find food at the ends of the arms these common features become associated with reward, so mice will tend to run down arms expecting food. However, these common intramaze cues are present in both rewarded and non-rewarded arms, and are therefore ambiguous. In order to inhibit the tendency to run down non-rewarded arms, and so discriminate successfully between baited and never-baited arms, the mice must therefore use the arm-specific, extra-maze spatial cues to select the correct response (run vs. don’t run) for each arm. We hypothesized that it is this ability to use spatial information to disambiguate between competing or overlapping memories (a form of pattern separation), and thereby behaviorally inhibit inappropriate associative, conditioned responses that is impaired in GluN1^ΔDGCA1 mice on this task.

To test this possibility, we trained experimentally naïve mice on a watermaze paradigm in which they were required to discriminate between two visibly identical beacons (black spheres), depending on their spatial locations (Fig. 6a)^{27, 28}. One beacon was located over the platform. The other, “decoy” beacon was at the point-symmetrically opposite location in the pool and provided no escape. Both beacons remained in fixed spatial locations. Mice were placed into the pool, either close to the correct beacon (S⁺ position), close to the decoy beacon (S⁻ position), or equidistant between the two beacons (Fig. 6b).

Figure 6. Spatial memory and choice performance on the spatial discrimination beacon watermaze task.

(a) Schema and horizontal side view of the spatial beacon watermaze task. (b) Top view indicating the six possible start sites and the positions of the beacons (broken line = + Platform; shaded circle = – platform). (c) Both Controls (n=11) and GluN1^ΔDGCA1 mice (n=12) learned to choose the correct beacon in terms of first choice accuracy: geno – F(1,21)=2.88; p>0.10, geno x blocks – F<1; 24 trials/block). (d) When probe tests (TT1 and TT2; beacons and platform removed) were conducted, both groups showed a strong preference for the goal quadrant (adjacent left, goal (G), adjacent right, opposite). (e) GluN1^ΔDGCA1 mice were impaired at choosing the correct beacon when starting from a point close to the incorrect beacon (S⁻), but not when starting from either of the other two start locations; (S⁺) or equidistant; geno x start position - F(2,42)=4.24; p<0.025, simple main effects group difference *p<0.025). (e, right) Mean ± s.e.m. percent correct choices across 5 blocks of training (8 trials/block) on trials starting from the S⁺, the equidistant and the S⁻ position. Broken line, chance levels of performance. All means ± s.e.m.

GluN1^ΔDGCA1 mice were impaired in their spatial choices between the two beacons, despite knowing the platform’s spatial location. During pre-training with a single beacon in a variable spatial location, all mice rapidly learned to swim towards the beacon and escape onto the platform. During subsequent spatial discrimination training both groups then learned to select the correct beacon according to its spatial location as shown by the gradual improvement in first choice accuracy (ANOVA; main effect of block - F (4,84) = 24.52; p <0.0001; no main effect of genotype - F (1,21) = 2.88; p >0.10; no genotype by block interaction - F < 1; p >0.70; Fig. 6c). Furthermore, when probe tests were performed, with both beacons and the platform removed from the pool, both groups equally preferred the quadrant where the platform had been, demonstrating spatial learning (Fig. 6d; probe test 1 ANOVA; effect of quadrant - F (2,63) = 19.29; p < 0.05; no genotype by quadrant interaction - F < 1; p >0.80; time spent in the training quadrant only – t-test; t < 1; p > 0.80; probe test 2 ANOVA; main effect of quadrant - F (2,63) = 19.90; p < 0.05; no genotype by quadrant interaction - F (2,63) = 2.07; p >0.10; comparison of training quadrant times – t-test; t (21) = 1.75; p = 0.09). Thus, GluN1^ΔDGCA1 mice learned the location of the platform as well as the Controls, successfully replicating the finding of preserved associative spatial memory from our first experiment.

However, when first choice accuracy was examined in terms of starting position, GluN1^ΔDGCA1 mice were impaired at choosing the correct beacon when started from close to the incorrect/decoy beacon (S⁻), but not when starting from either of the other start locations (S⁺ or equidistant; Fig. 6e). Not surprisingly, given that they had learned to swim towards the beacon as a means of escape during pre-training, all mice initially approached the first beacon that they encountered which resulted in sub-chance performance on S⁻ trials and above-chance performance on S⁺ trials. Mice then gradually learned to inhibit this approach response depending on the spatial location of the beacon. ANOVA of first choice accuracy also revealed a main effect of start position, which was due to poorer performance when mice started close to the S⁻ beacon and better performance when they started close to the S⁺ position (F (2,42) = 233.50; p <0.0001). There was also a significant genotype by start position interaction (F (2,42) = 4.24; p <0.025). Further investigation, using analysis of simple main effects, confirmed that GluN1^ΔDGCA1 mice were impaired when trials started from close to the S⁻ beacon (F (1,21) = 5.92; p <0.025), but not when trials started close to the S⁺ beacon (F < 1; p > 0.80), nor from equidistant between the beacons (F (1,21) = 1.86; p> 0.10). Separate ANOVAs revealed significant improvements with training for each start position (main effect of block; S⁻ trials - F (4,84) = 8.24; p <0.0001; equidistant trials - F (4,84) = 24.50; p < 0.0001; S⁺ trials - F (4,84) = 5.63; p <0.0005). Importantly, by the end of training, the control mice were exhibiting above chance levels of performance from all start positions (e.g. S⁻ trials only; one group t-test - p <0.05).

Because mice sometimes swam under the S⁻ beacon and then re-emerged before, again, swimming under the S⁻ beacon, we also analyzed total errors in which each repeated choice added to the cumulated score for the trial. This revealed a main effect of genotype (F (1,21) = 18.53; p <0.0005) and an interaction between genotype and starting position (F (2,42) = 5.18; p <0.01; Supplementary Fig. 8). Subsequent analysis of the genotype by start position interaction using simple main effects confirmed that GluN1^ΔDGCA1 mice were impaired on S⁻ trials (F (1,21) = 17.02; p <0.001), and unimpaired on S⁺ trials (F <1; p > 0.30). On the total errors measure they were now significantly impaired on equidistant trials as well (F (1,21) = 5.39; p <0.05), presumably because if they did approach the wrong beacon they were then more likely than Controls to persist in that response.

Non-spatial discrimination using a beacon watermaze task

GluN1^ΔDGCA1 mice performed as well as Controls on an analogous, non-spatial version of the task in which they had to choose between two visually distinctive beacons²⁷ (grey funnel vs. black/white striped cylinder; Fig. 7a,b). Mice acquired the visual discrimination task, displaying increased first choice accuracy for the S⁺ beacon and reduced total errors as training proceeded, but there were no group differences (Fig. 7c,d, Supplementary Fig. 9). As in the spatial discrimination task, mice found the task harder (they were less accurate with their first choices and made more errors), when starting from close to the S⁻ beacon. Conversely, they performed more efficiently when trials started from close to the S⁺ beacon. ANOVA of first choice accuracy revealed a main effect of block, demonstrating acquisition of the task (F (5,75) = 27.80; p <0.0001; Fig. 7c). There was also a main effect of start position which was due to mice being more likely to choose incorrectly when starting from close to the S⁻ beacon, and more likely to choose correctly when starting from closer to the S⁺ beacon (F (2,30) = 48.20; p <0.0001; Fig. 7c right, 7d). Importantly, however, there was no main effect of genotype or significant interaction involving genotype (ANOVA; genotype - F < 1; p > 0.60; genotype by block, genotype by start position - both F < 1; p > 0.60; genotype by block by start position - F (10,150) = 1.09; p > 0.30). Similar results were obtained when analyzing total errors (Supplementary Fig. 9). Finally, there was no impairment in GluN1^ΔDGCA1 mice during subsequent reversal of the non-spatial beacon task (Supplementary Fig. 10).

Figure 7. Normal performance of GluN1^ΔDGCA1 mice on the visual discrimination beacon watermaze task.

(a) Horizontal side view of the non-spatial beacon watermaze task. (b) Top view of the six possible start positions. (c) Both Controls (n=8) and GluN1^ΔDGCA1 mice (n=9) learned to choose the correct beacon. First choice accuracy improved with training for both groups across 6 blocks of training (24 trials/block; c, left). The first choice accuracy across all training trials from different starting positions showed that all mice were less accurate at choosing the correct beacon when starting from S⁻ (c, right). (d) Performance across 6 blocks of training on trials starting from S⁺, equidistant and S⁻ positions. All mice performed less accurately when starting from S⁻. There was no difference between Controls and GluN1^ΔDGCA1 mice across 6 blocks of training for any of the start positions (8 trials/block). Broken line: chance levels of performance. All means ± s.e.m.

DISCUSSION

The present results show that NMDARs on DG and dorsal CA1 principal cells are not essential neuronal underpinnings of hippocampus-dependent spatial memory acquisition or storage. Instead they may play a critical role in using spatial information to guide selection between alternative responses. Despite ablation of NMDARs from both DG granule cells and dorsal CA1 pyramidal cells, and additionally some DG granule cell loss, GluN1^ΔDGCA1 mice still acquired the hippocampus-dependent, spatial reference memory version of the watermaze as well as Controls. This was despite the loss of NMDAR currents and LTP in dorsal CA1 pyramidal cells in these animals. Importantly, although our manipulation was not confined to dorsal CA1 but also affected ventral CA1 and the entire DG, these additional consequences cannot explain the lack of a deficit on the standard watermaze task. Furthermore, while there is a gradient in ventral hippocampus of residual CA1 pyramidal cells that retain NMDARs, it is unlikely that these could support watermaze acquisition because lesion data from both rats and mice show that this ventral region alone is insufficient to support spatial learning on this task^18-20.

In contrast, GluN1^ΔDGCA1 mice were unable to acquire the hippocampus-dependent, spatial reference memory radial maze task. This dissociation between the two classic tests of associative long-term spatial memory is not due to different sensorimotor or motivational task demands, but instead reflects the different psychological processes involved. To investigate this dissociation further, a variant of the watermaze task was run in which mice were required to solve a spatial discrimination between two visually identical beacons. GluN1^ΔDGCA1 mice were less able than Controls to withhold responding to the incorrect, decoy beacon despite an equal knowledge of the spatial location of the platform. This did not reflect a generalized increase in approach tendencies. When two visually distinct, and hence unambiguous, beacons were used, GluN1^ΔDGCA1 mice were unimpaired. Thus, an account based on how likely mice were to swim to any beacon, or based on task difficulty, cannot explain these data. Instead the impairment in the spatial task arose specifically as a result of the ambiguity evoked from two identical beacons. While previous studies have revealed dissociations within the domain of associative, long-term spatial memory²⁹, here we show that even when spatial navigation is normal in GluN1^ΔDGCA1 mice, they are unable to use that spatial knowledge to select appropriately between competing response options.

The behavioral consequences in GluN1^ΔDGCA1 mice were distinct from those seen after NMDAR deletion restricted to just DG granule cells¹⁴. GluN1^ΔDG mice displayed normal long-term associative spatial reference memory on the radial maze (although they did exhibit spatial working memory impairments). GluN1^ΔDGCA1 mice in the present study were impaired on the associative, long-term reference memory radial maze task, and one might therefore conclude that it is the gene ablation in CA1 pyramidal cells that is specifically responsible, although we cannot rule out a cumulative phenotype in DG and CA1. Nevertheless, the very selective behavioral impairments exhibited by GluN1^ΔDGCA1 mice, combined with the preserved ability to learn the spatial location of the platform in the watermaze, provide key information about the algorithm being performed by the hippocampus.

Our data dispute previous claims that NMDARs in dorsal CA1 are essential for associative spatial memory formation in the watermaze⁷, based on conditional GluN1 knockout via the Tg29-1 transgenic mouse line, which was thought to be confined to dorsal CA1 pyramidal cells. Potential differences between the two studies include genetic background, temporal onset of the genetic deletion, the dorsal/ventral extent of GluN1 ablation within the hippocampus and the age of the animals. The most likely explanation, however, is the lack of hippocampal selectivity of the GluN1 gene knockout, which more recent studies of the Tg29-1 line have demonstrated spreads into cortical areas^9-13. Subsequent publications report a clear reduction in GluN1 levels, measured by in situ hybridisation and immunocytochemistry, as early as 2 months of age if not sooner^{11, 13}. Indeed, the Tg29-1 line expresses Cre in cortex as early as 6 weeks post-natally⁹. The evidence that Cre expression, and hence loss of NMDARs, occurs outside of the hippocampus with this transgenic line is now compelling (see also^{10, 12}). Notably, the Tg29-1 mice with floxed Grin1 alleles were also impaired on a visually cued version of the watermaze paradigm⁷, demonstrating a hippocampus-independent learning impairment and indicating an effect of GluN1 deletion outside the hippocampus.

Combined with our current data, these results suggest that NMDARs, either elsewhere in the extended hippocampal formation, such as entorhinal cortex³⁰ or subiculum³¹, or across the wider cortical mantle, could underlie associative spatial memory performance in the watermaze. Notably, a similar conclusion arises from studies with GluN2B knockout mice. Whereas genetic deletion of the GluN2B subunit from both hippocampus and cortex impaired watermaze acquisition¹², deletion of the subunit specifically from dorsal CA1 pyramidal cells and DG granule cells had no effect on acquisition²². Thus, there is likely to be an important role for cortical NMDARs in spatial memory acquisition.

At first glance the current results also appear at odds with pharmacological studies in rats. Intracerebroventricular infusion of the NMDAR antagonist AP5 impaired acquisition of the reference memory version of the watermaze⁴. Of course, intracerebroventricular infusion will affect both hippocampal and extra-hippocampal NMDARs. Even so, subsequent experiments found that watermaze acquisition was not impaired by NMDAR blockade using AP5 if rats had first been pre-trained as normal animals on a spatial task in a different laboratory^{5, 6}. This is consistent with the present data in demonstrating that hippocampal NMDARs are not necessary (i) for spatial navigation per se, nor (ii) for forming a long-term association between a particular spatial location and the platform. In a further parallel with GluN1^ΔDGCA1 mice in the present study, despite spatial pre-training, AP5 treated rats were impaired during spatial reversal in the watermaze when the location of the platform was switched to a novel place within the same, familiar environment³².

Thus, the role of hippocampal NMDARs lies not in associative memory formation but elsewhere. Although NMDAR deletion in DG and dorsal CA1 principal cells does not prevent encoding of long-term spatial memories, it does affect the use of spatial information to disambiguate between overlapping or competing associative memories. It has been suggested that the hippocampus is a key component of a comparator system to detect mismatch or conflict^{33, 34}. Human fMRI experiments^35-37 and both electrophysiological unit recording^38-40 and lesion studies⁴¹ in rodents have implicated the hippocampus, and particularly the CA1 subfield, with a role in detecting and/or resolving associative mismatch and conflict, such as might occur between an expectation based on information retrieved from long-term memory and the current state of the perceptual world. This comparator function is not, however, a reward prediction error signal that determines the extent of associative learning⁴². Instead, one key output of this hippocampal mismatch detection system is behavioral inhibition of on-going activity^{33, 34}, modifying within-trial performance without affecting the progressive changes in associative learning from trial to trial. A failure in this process likely underlies the deficit in GluN1^ΔDGCA1 mice when the platform was moved to the opposite quadrant of the pool during the first watermaze experiment (Fig. 3a, Reversal; see also³²). Notably, GluN1^ΔDGCA1 mice were not impaired in either of two non-spatial reversal paradigms. This excludes an account based on a general tendency to perseverate all learned responses.

The present results also demonstrate a novel but related role for hippocampal NMDARs in resolving interference between competing long-term memories, so enabling behavioral inhibition of inappropriately cued or conditioned responses on the basis of spatial or contextual information. This key deficit in GluN1^ΔDGCA1 mice manifested as an inability to withhold responding to the decoy beacon, despite knowing the spatial location of the platform. Critically, this did not reflect a general inability to inhibit all approach responses, as evidenced by the normal performance of GluN1^ΔDGCA1 mice on the visual discrimination beacon task, irrespective of start position. Instead, the deficit occurred when a conflict arose as the result of ambiguity between competing long-term memories associated with two visibly identical beacons. This same account can explain the radial maze deficit in GluN1^ΔDGCA1 mice, which results from the ambiguity that arises from the identical physical appearance of all 6 arms of the maze.

It has repeatedly been suggested that an important role of the hippocampus is pattern separation to disambiguate between overlapping inputs^43-45. On this occasion the ambiguity arises, not from overlapping extramaze spatial cues, but from the similarity of the maze arms or the visibly identical beacons, and the fact that these cues are associated with conflicting outcomes. Thus, at a psychological level, NMDARs in the hippocampus may serve to resolve conflict or ambiguity as a result of detecting mismatch or uncertainty rather than sub-serving learning by detecting coincidence.