The reorganization and reactivation of hippocampal maps predict spatial memory performance

Goal-related reorganization of hippocampal firing patterns

We developed a spatial memory task in which rats learned and subsequently recalled the locations of three hidden food rewards on a cheeseboard maze (Supplementary Fig. 1; Supplementary Methods). The learning session included 40 trials during which animals had to retrieve all hidden rewards before returning to the start-box to collect an additional reward. To prevent the use of odor cues, food dust was scattered across the maze and the board was rotated relative to the start-box (see Methods). The procedure required daily memory updates of goal locations because a new set of bait-locations was introduced every day. The animals’ memory performance was assessed by the number of crossings (Fig. 1b) and the time spent (Supplementary Fig. 2a,b) at goal areas (10 cm diameter circle around the learnt bait-locations) during the probe sessions in which rewards were no longer provided. These probe sessions were performed 2h after each daily learning session (i.e., post-learning probe or “post-probe”) and, on the following day, prior to the new learning session (i.e., pre-learning probe or “pre-probe”). Pre-probe sessions also served as a control: they were compared with post-probe sessions performed on the same day in order to assess changes to hippocampal network activity following learning of new bait-locations.

The learning performance of animals improved rapidly on each day: the distance travelled to complete a trial showed >50% reduction following the first trial and reached an asymptotic level within a few trials, showing that rats rapidly encoded and remembered the new bait-locations for the remaining trials (Fig. 1a, Drug-free). Rats also remembered these bait-locations in the subsequent post-probe session: they visited the newly-learnt bait-locations more frequently than those learnt the day before (Fig. 1b, Drug-free post-probe: P=0.0003; Supplementary Fig. 2b). These locations were remembered on the following day, as assessed during the next day’s pre-probe session (Fig. 1b, Drug-free pre-probe: P=0.010; Supplementary Fig. 2b).

To reveal how spatial memories for new goal locations are represented in the hippocampus during this task, we recorded the activity of multiple place cells and oscillatory field potential patterns using multichannel extracellular techniques ^13,35. The hippocampal representation of goal locations was quantified as the proportion of cells with a place field centre falling within a goal area (the place field center area was defined at pixels where the firing rate was larger than 80% of the peak rate; goal area was defined as a 10cm circular region around the food well; see Methods). Consistent with previous studies^14-16, we observed goal-related changes to hippocampal place maps (referred to as “goal-oriented remapping”). However this was region-specific: more CA1 cells represented newly-learnt locations in the probe session after learning than in the one before, while CA3 representations did not change (Fig. 2a-c, Drug-free; Supplementary Figs. 3 and 4a). The exclusion of SWR-related spiking activity did not change these results (CA1: pre-probe=0.090±0.015, post-probe=0.203±0.010, P<0.0001; CA3: pre-probe=0.099±0.025, post-probe=0.112±0.029, P>0.698; paired t-test). Moreover, the proportion of place cells representing the start-box did not significantly change (CA1: pre-probe=0.082±0.020, post-probe=0.115±0.016, P>0.263; CA3: pre-probe=0.077±0.037, post-probe=0.080±0.028, P>0.947; paired t-test). A population vector analysis ³⁶ used to quantify further the similarity of place-related assembly patterns (see Methods), showed that the CA1 population similarity score between probe sessions was lower than the baseline score calculated within sessions, while these scores remained similar in CA3 (Fig. 2f). Therefore, CA1 place-related assemblies present during probe sessions reorganized following learning, while CA3 assemblies remained stable.

We then tested whether the CA1 goal representation we observed during the post-probe session was established during learning per se. We detected learning-related reorganization of CA1, but not CA3, firing patterns: the proportion of CA1 place cells representing goal locations increased gradually over trials (Fig. 2d, Drug-free CA1 with r=0.370, P<0.00001; Supplementary Figs. 3 and 4b). Moreover, CA1 assembly patterns observed during the last 10 learning trials (referred to as “end of learning”) were more similar to assembly patterns during the probe session after learning than before (Fig. 2g, Drug-free; Supplementary Fig. 5). Hence, CA1 goal representations developed gradually during learning, and those representations present at the end re-emerged in the subsequent probe session.

The performance improvement during learning is reflected by the development of stereotyped paths. Therefore, the reorganization of CA1 place cells might occur as a consequence of animals altering their foraging trajectory, and not because such reorganization is required for spatial memory. Such reorganization could be explained also by the presence of a reward or disproportionate dwell-time at reward locations. To test for these, we used a “Cued” version of the task in which food-wells were visually marked by intra-maze cues, such that animals did not have to remember the locations to gain the reward (Supplementary Fig. 1). During this cued learning, animals ate the same number of rewards; exhibited similar stereotyped movement paths; and spent similar time at goal locations as in the absence of cues (Fig. 3a,c; Supplementary Fig. 2). However, animals did not exhibit any spatial preference in subsequent probe sessions, indicating that they did not learn the visually-guided locations (Fig. 3b, Supplementary Fig. 2b). In these control experiments, we found that CA1 goal-oriented remapping did not take place, suggesting that such reorganization occurs when a map-based strategy was used to locate hidden rewards (Fig. 3d-h, Supplementary Figs. 4 and 6).

Effect of the NMDAR blockade

These results show that, in our task, new spatial memories were encoded by CA1 place maps representing goal locations during learning, and that these goal-oriented maps were reinstated as stable representations alongside successful memory recall. Next, we tested whether the acquisition, stabilization and/or reinstatement of such goal-related firing patterns could be observed under conditions of memory impairment. It has been shown that NMDARs are required for spatial memory ^3,31-34. Therefore, rats were injected with the NMDAR antagonist CPP (3-((R)-2-Carboxypiperazin-4-yl)-propyl-1-phosphonic acid; see Methods) after the pre-probe session ^32,37 in order to interfere with their spatial memory. We found that the learning performance of CPP-treated animals improved rapidly (Fig. 1a) and was comparable to those in the drug-free condition. However, CPP-treated animals subsequently failed to remember the newly-learnt locations in both probes when there was at least 2h gap between learning and recall (Fig. 1b, all Ps>0.339; Supplementary Fig. 2b; see Supplementary Fig. 14 for shorter delays).

In association with impaired memory for goal locations, we found that NMDAR blockade prevented the stabilization of hippocampal maps representing those locations. During learning, the proportion of CA1 place cells with goal-related firing increased gradually (Fig. 2d, CPP CA1 with r=0.334, P<0.00001; Supplementary Figs. 4b and 7), as in the drug-free condition. However, under CPP the similarity between reorganized assembly patterns present at the end of learning and those in the following probe session (Fig. 2g, CA1: P<0.0001, paired t-test) was reduced in comparison to the drug-free case. Moreover, assembly patterns remained similar across probe sessions (Fig. 2f), as did the proportion of place cells with goal-related firing (Fig. 2c, Supplementary Figs. 4a and 5). Therefore, goal-related CA1 place cell representations developed under NMDAR blockade during learning, in conjunction with animals’ learning of goal locations. However, these newly-acquired representations did not stabilize, in line with the fact that animals no longer remembered goal locations during subsequent probe sessions.

To test whether place cell representations are related to memory traces, we examined whether representations of goal locations predict spatial memory performance. We found that goal-oriented hippocampal maps during learning predicted the animal’s future memory performance as did the maintenance of these maps in the probe session: memory performance, as estimated by the number of crossings in goal areas, was correlated with the proportion of CA1 place cells representing goal locations at the end of learning (Fig. 2e, r=0.511, P=0.0014) and in the following probe session (Supplementary Fig. 8) but not at the beginning of learning nor in the probe session before (all Ps>0.131). This was the case neither under CPP nor in the Cued condition (all Ps>0.137).

These results show that NMDAR-dependent mechanisms are important for the stabilization of hippocampal goal-related firing patterns and for the recall of associated memories.

eSWR network responses during learning

In the hippocampus, SWR events have been suggested to assist the stabilization of memory traces ^{7,21,22,28,35}. Although SWR activity has been traditionally described during off-line sleep/rest (sSWRs) including periods of slow-wave sleep and long waking immobility, they are also present during on-line exploratory periods (eSWRs) ³⁵. These eSWRs have been suggested to strengthen place cell representations and therefore they may assist in stabilizing newly-formed goal representations ³⁵. We tested whether eSWR network responses predict memory performance; and whether they are sensitive to NMDAR blockade during the on-line stage of memory trace formation/stabilization. During learning, we observed eSWRs at the bait-locations (Fig. 4a,b). The number of goal-associated eSWRs was correlated with memory performance in the drug-free condition (r=0.524, P=0.0010) suggesting that eSWRs were associated with on-line memory trace formation. However, the number of SWRs that occurred during longer (>2.4s) immobility periods (iSWRs) ³⁵ during learning did not predict memory performance (P>0.722). Moreover neither eSWR nor iSWR numbers predicted memory performance under CPP (all Ps>0.410). Ongoing place-related activity is supplemented by increased network activity during eSWRs, strengthening the synchronized firing of cells encoding the same location and consequently promoting synaptic plasticity amongst them ³⁵. Since eSWRs tended to occur at reward-locations, we checked whether cells representing these places strengthened their firing synchronization during eSWRs. We compared the eSWR firing rate histograms of CA1 place cells representing goal locations to those representing the start-box (referred to as “goal-centric” and “start-box” cells respectively). In the drug-free condition, the infield firing rate of “goal-centric cells” during eSWRs was higher than that of “start-box cells” (Fig. 4c, 5.81±0.29Hz versus 4.22±0.36Hz, P<0.002). This was abolished under CPP (4.42±0.24Hz versus 3.98±0.35Hz, P>0.319) and was not present in the Cued condition (Supplementary Fig. 9a). Moreover, the strength of eSWR network responses measured at the end of the learning was associated with memory recall: network participation (i.e., synchrony) of CA1 place cells in eSWRs was correlated with memory performance (Fig. 4d, r=0.418, P=0.011) even when controlled for the eSWR firing rate or the proportion of place cells at goal locations (r=0.448, P=0.016; r=0.364, P=0.023 respectively, partial correlation). This was the case neither under CPP nor in the Cued version (all Ps>0.318).

Reactivation of waking firing patterns during rest

During off-line rest periods, sSWRs have been suggested to promote memory consolidation because waking activity patterns of place cells are reactivated during these network events ^26-28,35,38. To test whether reactivation occurs following spatial learning, we examined whether cells encoding similar places during learning fire together in subsequent sSWRs ²⁸. The place field similarity (PFS) of cell pairs (measured for place fields present at the end of learning) exhibited significant correlations with their joint firing tendency (cofiring) during sSWRs (Fig. 5a; Supplementary Figs. 10 and 11). For both goal-centric and start-box cells, cofiring in sSWRs after learning correlated more strongly with PFS than did cofiring before learning (Fig. 5a). However, reactivation of goal locations was even stronger than the reactivation of the start-box location (Fig. 5a, Drug-free). While CPP application did not prevent reactivation per se, such an enhanced reactivation of goal locations was abolished (Fig. 5a, CPP; Supplementary Figs. 10 and 11). The enhanced reactivation of goal locations was not observed in the Cued condition (Supplementary Fig. 9b). Since CPP caused a memory deficit in our task, we hypothesized that conflicting representations were reactivated under CPP. Because place representations reorganized during learning, we compared the reactivation of goal-related firing patterns that were present at the beginning and the end of the learning session. Under NMDAR blockade, but not in the drug-free condition, firing patterns from the beginning of learning were still reactivated strongly in the subsequent rest period (Supplementary Fig. 12). Thus, NMDAR blockade prevented the boosted reactivation of new goal-related patterns, allowing the recurrence of old, conflicting representations.

Finally, we tested whether reactivation of goal-related CA1 assembly patterns predicts memory performance. For each sSWR we calculated a “reactivation map” in order to determine which locations were represented by the sSWR assembly pattern (see Methods). These reactivation maps quantified how similar the sSWR assembly pattern was to waking assembly patterns representing different locations. Each pixel on these reactivation maps reflects the similarity of the activity pattern in the sSWR to the assembly activity pattern derived from the combined place-rate maps seen at that location during exploration. Thus, the peak of the reactivation map marked the location which the sSWR population activity represented the best. Either rate maps from the end of learning or from the following probe period were used to measure the similarity of assembly patterns (see Methods). For most sSWRs, the resulting reactivation maps highlighted one of the bait-locations (Fig. 5b). Moreover, the proportion of sSWRs representing goal locations predicted subsequent memory performance when reactivation maps were created using either end of learning (Fig. 5c, r=0.620, P=0.00005) or post-probe (r=0.362, P=0.028) representations. In contrast, this was not seen if sSWRs were taken from the rest session before learning (all Ps>0.193). This correlation was significant even when we controlled for the proportion of place cells at goal locations (end of learning: r=0.502, P=0.0007, post-probe: r=0.317, P=0.038, partial correlation). These relationships were not found under CPP (all Ps>0.412).

Subjects and electrode implantation

All procedures were carried out in accordance with the Animals (Scientific Procedures) Act, 1986 (UK) and associated procedures under an approved project license. We implanted seven adult male Long-Evans rats with 16 independently movable wire-tetrodes that were positioned above the right dorsal hippocampus (see Supplementary Information). Following a one-week postoperative recovery period, rats were reduced to and maintained at 85% of their age-matched preoperative weight. Water was available ad libitum. During this period, tetrodes were lowered to the CA1 and CA3 regions of the dorsal hippocampus.

Training

Each daily experiment consisted of a sequence of five recording sessions in the following order: a probe test (“pre-probe”), an immobility/sleep rest session (“pre-rest”), a learning session, an immobility/sleep rest session (“post-rest”) and a probe test (“post-probe”) (see Supplementary Fig 1b). Hippocampal neuronal assembly activity was continuously monitored during these sessions. The two probe sessions (~25 min) were never rewarded. After both the pre-probe and the learning sessions, rats were allowed to settle down within the start box for the rest sessions (~25 min). During the learning session, rats were given successive trials (~40 trials) to locate a new set of 3 hidden rewards placed in randomly selected food-wells every day. As these baited locations changed from day-to-day but stayed fixed within a given day, this “matching-to-multiple-places” procedure required frequent updating of memory for goal locations in an otherwise unchanging environment. The same paradigm was used for the NMDAR blockade experiments during which rats were injected with the 3-((R)-2-Carboxypiperazin-4-yl)-propyl-1-phosphonic acid [(R)-CPP, 10 mg.kg⁻¹, i.p., Tocris Cookson Ltd] after the pre-probe ^32,37 (see Supplementary Methods). For the Cued version of the task, three identical objects (Falcon tube 50 ml, Greiner, see Supplementary Fig 1a) were placed near the baited food-wells during the learning trials and rats were trained to retrieve the hidden rewards from those visually marked locations. In this circumstance, the task was solved by using a guidance strategy that consisted of moving towards intra-maze cues identified to be closely associated with the goals. These intra-maze cues were removed for the two probe tests. Three rats were tested in all three experimental conditions. To prevent the use of an odor-guided search strategy during these experiments, food pellet dust was scattered across the maze before each experiment, the board was periodically wiped (using the towel used to handle the rat daily) and the board was rotated relative to the start-box between learning trials and between rest and probe sessions. In all a total of 12, 8 and 9 sequences of probe-rest-learning-rest-probe recorded in the familiar recording room were analyzed for the drug-free, CPP and Cued conditions respectively for a total of 2040 CA1 pyramidal cells (drug-free=1074, CPP=612, cued=354) and 690 CA3 pyramidal cells (drug-free=257, CPP=221, cued=212) included in the analysis (see Supplementary Methods).

Data acquisition

Wide-band (0.1Hz–5kHz) recordings of local field potentials and multiple-unit activity were amplified 1,000-fold using a 64-channel amplifier (Sensorium) and continuously digitized at 20 kHz using a 64-channel analogue-to-digital converter computer card (United Electronics Industries). Two 32-channel unity-gain preamplifier headstages (Axona Ltd) were used to reduce cable movement artefacts. An array of three LED clusters mounted on the preamplifier headstages was used to track the location of the animal (25 frames per s) via an overhead video camera (Sony). The animal tracking was synchronized with the electrophysiological recording. The animal’s location was constantly monitored throughout the daily experiment. The data were analyzed off-line using custom made software, including all unit isolation and field analysis, and occasionally the STAT 5.4 UNIX software package (Perlman G., 1980) and the R software environment (http://www.r-project.org/).

Behavior

Behavioral performance was calculated off-line using the animal’s position records in the tracking data. Learning performance was assessed by calculating the distance travelled to retrieve all three rewards during each trial. Memory retention performance was assessed during the first 3 minutes of each probe by both scoring the number of crossings (Figs. 1 and 3) and the time spent (Supplementary Fig. 2) in the goal areas (10 cm in diameter centered on the learnt baited locations). The pre-probe was used to evaluate the memory for the baited locations learnt the day before (i.e., 24h before learning); the post-probe was used to assess the memory for the newly learnt locations (2h after learning). Differences between groups were analyzed using a Student’s t test or an ANOVA which was followed by a post hoc comparison using the Tukey’s HSD post hoc as appropriate. Memory performance (number of crossings) from the probe session following learning was used for correlation analysis.

Spatial firing rate maps

The x-y plane of the cheeseboard was divided into bins of 5×5 cm² and hippocampal place rate-maps were calculated during exploratory epochs (speed>5cm/s) by a kernel-based method in which both the firing rate and occupancy maps were smoothed with a Gaussian kernel function ^28,35. Hippocampal place cells were screened for their spatial tuning using a coherence value of at least 0.6 and a sparsity value of no more than 0.3. Coherence reflects the similarity of the firing rate in adjacent bins, and is the z–transform of the correlation between the rate in a bin and the average rate of its eight nearest neighbors ⁴⁷. Sparsity corresponds with the proportion of the environment in which a cell fires, corrected for dwell time ⁴⁸, and is defined as (ΣPiRi)²/ΣPiRi², where Pi is the probability of the rat occupying bin i, Ri is the firing rate in bin i. For each place cell, the place field center was defined as the x-y locations where the cell fired >80% of its peak firing rate. The hippocampal representation of goal locations was then quantified as the proportion of cells with a place field centre falling within <10cm circular region from the center of a baited well. The similarity score of place-related assembly patterns was estimated using a population vector analysis ³⁶: rate maps were stacked into three-dimensional matrices for each waking period (the two spatial dimensions on the x and y axis, the cell identity on the z axis; see schematic in Figure 2f); population vectors were calculated at each x-y bin; these were then correlated between periods and averaged across all spatial bins. To avoid artefacts in the correlation measure, x-y locations visited less than 100 ms in either waking period were not considered for analysis. Analyses using the end of learning were performed using the last 10 learning trials. Analyses over the course of learning trials were performed by calculating firing rate maps using a successive window that encompassed 5 trials altogether.

Reactivation

Reactivation of place-related waking patterns was assessed by testing whether the tendency of pairs of cells to fire together (cofiring) during the rest session after learning reflected the degree to which their place-fields overlapped (place field similarity, PFS) as previously described ²⁸. The cell pairs were taken from hippocampal place cells with the peak firing rate pixel of their place map within 30 cm from either one of the food-wells or the start-box, and referred to as “goal-centric cells” or “start-box cells”, respectively. This analysis included a total of 17,980 and 10,369 goal-centric cell pairs in the drug-free and CPP conditions respectively, and of 2,979 and 1,734 start-box cell pairs in the drug-free and CPP conditions respectively. The PFS was established by calculating the correlation coefficient of their place rate-maps during learning. To measure the cofiring during rest, we first established for each cell its instantaneous firing rate counts (IFRC) in 100 ms windows centered on the peak of sSWRs and then calculated the correlation coefficient between the IFRCs for each cell pair. Since the rat explored the maze before both rest periods, reactivation in the rest after learning might not only reflect associations from learning but also activity from the pre-probe session of the day. In order to control for associations that may reflect exposures prior to learning or baseline correlations we used a partial correlation to assess reactivation of place-related waking patterns, correlating learning-related PFS with rest cofiring after learning while controlling for rest cofiring before learning^26,49,50. In order to verify the degree to which baseline correlations existed in the rest session before learning, the correlation was reversed, and the correlation was controlled by the cofiring after learning^49,50.

Reactivation maps were computed to observe whether CA1 assembly patterns present at the end of learning period or reinstated during the post-probe were present during the rest periods. We compared sSWR firing patterns from the rest periods to waking population activity from each spatial bin in the maze. Rest population firing vectors were established for each sSWR separately by calculating the IFRCs for all cells. Waking population activity was calculated as described above using the population vector analysis (see schematic in Figure 2f). Then we calculated the correlation coefficient between each sSWR-population vector and the waking population vectors from each x-y location in the maze. Peaks in these correlation maps are expected to correspond to regions in space where the waking population activity most strongly resembles the sSWR-population activity and thus presumably reflects the rat’s internal representation of its position at the time scale of a single ripple event. To test whether the relative strength of the reactivation of goal-related assembly patterns predicted subsequent memory performance, we used a partial correlation to calculate the similarity between each sSWR-population vector and the waking population vectors from either the end of learning or the post-probe, controlled for the corresponding waking population vectors from the pre-probe. Thus, for each sSWR event, each spatial bin was represented by a partial coefficient reflecting the similarity between the rest assembly patterns at that moment and the learning-related assembly patterns at that location. The reactivation index of goal-related assembly patterns corresponded to the proportion of sSWRs representing a goal location as indicated by the highest positive correlation coefficients on the map. The frequency of sSWRs did not differ significantly between the rest periods before and after learning and between the drug-free and the CPP conditions (see Supplementary Fig. 13).

eSWR responses

The eSWR network response of spatially active pyramidal cells and pyramidal cells that fire outside their place field was calculated as described before ³⁵. Briefly, eSWR firing rate histograms of CA1 pyramidal cells were calculated in 20 ms bin in reference to the eSWR peak (i.e., peak of ripple-band power) separately for their infield firing (i.e. inside the place field) and outfield firing (i.e. outside the place-field) firing ³⁵.