The entorhinal cognitive map is attracted to goals

See allHide authors and affiliations

Science  29 Mar 2019:
Vol. 363, Issue 6434, pp. 1443-1447
DOI: 10.1126/science.aav4837

Reward and the map in the brain

Recent findings suggest a more complex role of grid cells in the brain than simply coding for space. The grid map in the entorhinal cortex, which is responsible for encoding spatial information, is not as rigid as originally thought and can be distorted by environmental modifications (see the Perspective by Quian Quiroga). Butler et al. compared grid cell coding during a free-foraging task and a spatial memory task in rats. They discovered that entorhinal spatial maps restructure to incorporate the location of a learned reward. Boccara et al. tested the influence of behaviorally relevant information on the cognitive map that emerges from grid cell firing in the rat medial entorhinal cortex. They found that grid cells participate in neural coding of the goal locality, not the whole environment.

Science, this issue p. 1447, p. 1443; see also p. 1388


Grid cells with their rigid hexagonal firing fields are thought to provide an invariant metric to the hippocampal cognitive map, yet environmental geometrical features have recently been shown to distort the grid structure. Given that the hippocampal role goes beyond space, we tested the influence of nonspatial information on the grid organization. We trained rats to daily learn three new reward locations on a cheeseboard maze while recording from the medial entorhinal cortex and the hippocampal CA1 region. Many grid fields moved toward goal location, leading to long-lasting deformations of the entorhinal map. Therefore, distortions in the grid structure contribute to goal representation during both learning and recall, which demonstrates that grid cells participate in mnemonic coding and do not merely provide a simple metric of space.

To decide upon relevant behavior, individuals rely on dynamic neural representations of their world, computed from current and past experiences. The cognitive map formed by an extended network of specialized cell types coding for defined spatial features is essential for accurate navigation (1, 2). Hippocampal place cell activity is restricted to discrete, sparse place fields in specific environments, whereas parahippocampal grid cells present multiple firing fields arranged in regular hexagonal arrays that densely tessellate all environments (3, 4). This led to the hypothesis that grid cells provide a universal invariant metric for spatial cognition (5). As such, they were originally considered to have a narrower role than hippocampal place cells, which code for multimodal information beyond simple spatial representations (2, 6).

New data, however, suggest a more complex grid code (79). Nonspatial factors modulate local field firing rates without affecting the grid structure (10), and topographically organized auditory stimulus can drive grid-like structure (11). Furthermore, geometrical environmental features can influence the rigid grid structure, thus challenging the role of grid cells to provide invariant metrics (1214). However, grid distortions could also encode more complex behavioral information. We therefore tested the influence of behaviorally relevant information on the entorhinal cognitive map.

We trained rats to daily learn three new hidden reward locations on a cheeseboard maze while recording simultaneously from the medial entorhinal cortex (MEC) and the hippocampal CA1 region (Fig. 1A and fig. S1). This hippocampus-dependent task (15) consisted of three phases: pre-probe, learning, and post-probe, where the probes verified memory retention in the absence of food rewards (Fig. 1, B to D) (16). This paradigm led to daily changes in the cognitive valence of local points in an otherwise familiar environment, which in turn led to the accumulation of CA1 place fields around reward locations [i.e., goal remapping (15)]. This allowed us to test how rewards can be dynamically encoded in MEC neural representation during goal learning.

Fig. 1 MEC and CA1 spatial cells move toward newly learned goals.

(A) Dual recordings in CA1 (left) and MEC (right). Nissl-stained sections (red arrows denote electrode tracks) are shown with rate maps of simultaneously recorded place and grid cells [maximum firing rate in red and lowest in blue; peak firing rate (Hz) is shown at upper right] and LFP traces. (B) Behavioral sequence: pre-probe, rest 1, learning, rest 2, and post-probe. Bottom: animal’s path (gray) with goals (dots). (C) Average normalized learning curve (orange). (D) Memory retention test. Average time at goals: pre-learning (blue) and post-learning (red). P = 0.0053 (t test). Error bars indicate SEM across sessions. (E) Example grid cell showing spikes (colored dots) and field (circles) movement toward goals (black dots). (F) From top to bottom: Example maps of grid cells, MEC spatial cells, and CA1 place cells across paradigm. (G) Left: Proportions of spatial cells with fields moving significantly closer to goals. Significance was calculated against LNP data (solid bars) or downsampled data (hatched bars) (binomial test, all Ps < 0.0001; grid cells, 89%; MEC spatial cells, 84%; CA1 place cells, 79%). Right: Proportion of cells with their strongest fields at goals in pre-learning blue) and post-learning (red). Grid cells, P = 0.027, N = 56; MEC spatial cells, P = 0.0047, N = 157; CA1 place cells, P = 0.00018, N = 245 (Fisher exact test). Empty bars show control post-probe LNP data (all Ps < 0.018 against post-probe). Hatched bars show downsampled control data (all Ps < 0.00001). See supplementary figures for detailed legends. *P < 0.05, **P < 0.01, ***P < 0.001.

The majority of grid cells (80 to 90%) had at least one of their firing fields significantly moving toward a goal (Fig. 1, E to G, binomial test, P < 0.00001; see figs. S2 and S3) (16). MEC nonperiodic spatial and CA1 place cells showed a similar behavior (Fig. 1, F and G, and fig. S3, binomial test, P < 0.0001). This led to the accumulation of entorhinal and hippocampal strongest firing fields at goal locations (Fig. 1G). Because successful learning led to a higher number of visits to reward locations (Fig. 1D), we used a linear-nonlinear Poisson (LNP) spiking model to verify that the reorganization of firing fields was independent of variations in trajectories, speed or heading between pre- and post-probe by comparing real data to maps generated with this method (17). We also performed an additional control by downsampling the pre- and post-probe map to match occupancy in each spatial bin (Fig. 1G) (16). Conjunctive head-direction coding did not influence field movement to goal (fig. S4).

The proportion of cells with their strongest fields within goal locations increased progressively during learning in both entorhinal and CA1 spatial cells (fig. S5, A to C). To test whether these changes were long-lasting, we performed the same analyses on pre-probe, this time using the previous day’s goal locations. MEC cells retained the accumulation of firing fields around previous goal locations (fig. S6, A and B). Unexpectedly, CA1 changes were more transient (fig. S6C), suggesting differences in memory trace lability between those two regions, in line with models arguing for faster plasticity in CA1 than in MEC (18). We also observed faster CA1 plasticity during learning by scoring the development of goal representation as a function of firing related to goal vicinity (fig. S5, D to G) (16).

Because the pre-probe grid map maintained an accumulation of firing fields at the previous day’s goals, we examined the effect of the local change in cognitive valence on the grid structure itself. For three animals we added the exploration of a familiar open field, which acted as a control environment without valence bias (16). Grid scores were significantly higher in the open field relative to both probes [Fig. 2, A and B; one-way analysis of variance (ANOVA), P < 0.00001], independently of differences in spatial sampling, trajectories, speed, or heading (Fig. 2B; one-way ANOVA, P < 0.00001) (16).

Fig. 2 Grid score degradation on the cheeseboard.

(A) Example of MEC grid cell, which exhibits its highest grid score on the open field and a degraded score on pre- and post-probes. Top: Rate maps with maximum firing rate in red and lowest in blue; peak firing rate (Hz) is shown at upper right. Bottom: Corresponding spatial autocorrelogram maps, ranging from +1 (red) to –1 (blue), with grid score shown at upper right. (B) Average grid score (±SEM) across MEC cells in open field (OF; dark cyan), pre-probe (Pre; blue), and post-probe (Post; red) (P < 0.00001, one-way ANOVA). Hatched bars show downsampled control data (P < 0.00001, one-way ANOVA). Light cyan and empty bars show control data obtained with LNP spiking model in open field (LNP-OF) (P < 0.00001, one-way ANOVA). ***P < 0.001.

To test whether the grid score drop resulted from local map distortions around goals, we used a Laplacian of Gaussian (LoG) filter to detect individual fields independently of their peak firing rate (fig. S7) (9, 10, 16). Although there were no significant differences in the number of fields, nor in their size, spacing, or ellipticity after learning (fig. S8), we found a significant decrease in the mean distance between field center and closest reward [Fig. 3, A and B; Kolmogorov-Smirnov (KS) test, P = 0.01472] and an increase in the number of grid fields near goals (Fig. 3C; Fisher exact test, P = 0.0145).

Fig. 3 Stronger attraction of MEC fields closest to goal locations.

(A) Example of grid field movement from pre-learning (blue) to post-learning (red); overlay (right). Top: animal path (gray), spikes (colored dots), goals (black dots). Bottom: field (colored circles) detected by LoG filter; grayscale rate maps. (B) Density of fields as a function of distance to closest goal in pre- and post-probe; P = 0.01472 (KS test). (C) Proportion of fields at goals in pre- and post-probe; P = 0.0145 (Fisher exact test). (D) Scatterplot of attraction strength versus pre-probe goal distance. Purple line, running average; shaded region, SEM. Dashed line denotes significance threshold (P < 0.05, t test on each 10-cm window below 29 cm). (E) Running average of mean inferred attraction as a function of distance to closest goal, r = –0.192, P = 0.01225 (Spearman correlation analysis). Shaded region denotes SD. (F) Memory retention (time at goals) against normalized mean firing rate increase at goals. One dot per session; line denotes regression analysis, r = 0.7, P = 0.011. (G) Grid score in open field (left) and distorted open field (right, generated following real-data field movement distribution). P = 0.0003 (one-way ANOVA). Error bars represent SEM. *P < 0.05, ***P < 0.001.

To determine which parameters contributed to grid field movement, we calculated the “strength of attraction” of each field toward a goal and correlated it against spatial parameters and behavioral performance (Fig. 3, D and E, and fig. S9). The strongest correlation was found with the pre-probe distance to goal location: The closest fields were generally subjected to a strong attraction, whereas fields over a distance of ~30 cm showed little detectable attraction (Fig. 3, D and E). Most of the attracted fields moved toward the closest goals, and the most visited post-probe goal was pulling the strongest attraction (fig. S9). Different fields of a given cell could be attracted to different goals, depending on their relative position to the goals (fig. S7, H to K). We observed a weak increase in the peak grid field firing rate during post-probe; however, this increase was not associated with goal locations, and there was no correlation with strength of attraction (fig. S10). Finally, the increase of activity at goals correlated with memory retention (correlation analysis, r = 0.77, P = 0.002; Fig. 3F). There was no significant difference of in the strength of attraction between CA1 and MEC fields, although the pre-probe field distance influenced only MEC fields (fig. S11). To test whether the local movement of firing fields toward the closest goals could explain the grid score drop in probe sessions, we applied a movement toward imaginary goals on grid fields recorded in the open field, following the distribution of movements determined after learning. This resulted in a grid score drop was analogous to that observed between open fields and cheeseboard environments (Fig. 3G; P = 0.0003, one-way ANOVA).

We subsequently examined the reorganization of place-related assemblies at the population level (Fig. 4 and figs. S12 and S13). Both MEC and CA1 cell assemblies showed a significant reduction in vector similarity between pre- and post-probe relative to the intrinsic variability in comparing the two halves of the pre-probe (fig. S12B; all Ps < 0.0001, t test). This was independent of changes in spatial sampling (fig. S12B; all Ps < 0.0001, t test). MEC population vector similarity across pre- and post-probe sessions were weaker around goal locations than away from them (fig. S12C; P = 0.0003, t test) and positively correlated with distance from the goal locations (Fig. 4E; P < 0.0001). In contrast, the reorganization of CA1 population vectors between pre- and post-probes did not exhibit significant positive correlation with goal distance (Fig. 4E; P = 1).

Fig. 4 Flickering in MEC and CA1 between coexisting representations.

(A) Schemas illustrating population-vector computation. (B) Example of progression of z-scored correlation coefficients of ensemble activity across paradigm with pre-probe (negative, blue) and post-probe (positive, red) population vectors in MEC and CA1. Yellow line, smoothed score progression; black dashed line, regression analysis. (C) Distribution of z-scores in MEC (green), CA1 (brown), and cell-ID shuffling (black). MEC versus shuffling, P < 0.00001; CA1 versus shuffling, P < 0.00001 (KS test). The shuffling distributions did not differ (P > 0.1, KS test). (D) Running average of population vector similarity as a function of distance to goal. Green, MEC; brown, CA1. One-sided t test regression analysis: MEC, r = 0.1181, P < 0.00001; CA1, r = –0.00227, P = 1. Spearman correlation analysis: MEC, r = 0.10415, P < 0.00001; CA1, r = –0.02331, P = 0.02369. Shaded regions denote SD.

Given that goal remapping of individual MEC and CA1 spatial cells occurred incrementally during learning (fig. S5), we next examined how the expression of assemblies dynamically shifted toward post-probe goal representation. We computed the Fisher z-scored correlation coefficients of the population activity in 125-ms time bins, with the population vectors representing the current location of the animal in the pre- and post-probes (see example session in Fig. 4B and fig. S13). As learning progressed, similarity to post-probe representation increased in both MEC and CA1 cell assemblies. Yet when examining the fine temporal structure of assembly expression during learning, we observed a fast-paced flickering between pre- and post-probe representations. The distribution of z-scores was then compared with a control distribution obtained by a cell-ID shuffling procedure (Fig. 4C). Significant differences were detected for both CA1 and MEC (P < 0.00001 for all combinations, KS test). Moreover, we observed a significantly heavier-tailed distribution of real z-scores relative to the control shuffled ones (P < 0.0001, binomial test). Real data therefore showed a stronger tendency to have extreme values, which suggests that there were no intermediate representations but rather a quick alternation between the two competitive representations of pre- and post-probes. The MEC flickering we observed was reminiscent of previously reported CA1 flickering (19, 20). Given that MEC vector correlation increased with goal distance (Fig. 4D), we restricted flickering analyses to goal locations. The distribution of flickering scores was significantly different and generally shifted toward more positive values (figs. S12 and S13; P < 0.0001, KS and Mann-Whitney U tests). We also observed interregional differences in goal flickering dynamics, with CA1 reaching a plateau faster than MEC during learning (figs. S12 to S14).

Goal learning can lead to the local and long-lasting distortion of the entorhinal spatial maps. This demonstrates the influence of nongeometrical cognitive factors on the grid structure itself. These findings support emerging hypotheses that the grid pattern carries a broader organizational role for both spatial and nonspatial information in more complex and naturalistic behaviors (21, 22). Grid structure distortions have recently been linked to the geometrical features of the recording environment (1214), which may reflect distorted perception of space. Here, we found evidence for a grid code at the structural level that goes beyond simple metrics: Individual grid fields moved toward newly learned goal locations, leading to the deformation of the grid map, independently from variations of spatial sampling, trajectories, speed, or heading inherent to our behavioral paradigm.

Field attraction strength to a goal was proportional to the original goal-field distance, locally constraining the deformation of the entorhinal spatial representation. While local remapping took place in MEC, CA1 reorganized through global remapping. Moreover, field reorganization toward goal location was maintained overnight for MEC but not for CA1, although both maintained reorganized fields after learning (up to 2 hours). This argues for a higher lability of CA1 spatial memory traces relative to MEC cells (23). The role of the hippocampus in goal encoding was recently highlighted by reports of a subpopulation of CA1 neurons with an angular tuning for goal direction (24, 25). Although our results are consistent with the role assigned to CA1 in computational models of goal-directed navigation (21, 26), differences in goal coding between CA1 and MEC require updates in the current models.

Finally, we showed that assembly expression of different goal-related context rapidly alternated in MEC during learning, similar to CA1 flickering (19, 20). The expression of the old and new MEC representations in the same trial periods suggests that multiple maps can be stored in MEC. The apparent absence of intermediate representations in MEC has implications for how new or modified maps could be dynamically encoded and (re)organized in the CA1-MEC circuits during learning. This may open new avenues of computational research as to the role of MEC in inferential reasoning and associative memory.

Supplementary Materials

Materials and Methods

Figs. S1 to S14

References (2734)

References and Notes

  1. See supplementary materials.
Acknowledgments: We thank D. Derdikman for comments on an earlier version of the manuscript. Funding: Supported by European Research Council consolidator grant 281511 and European Union Horizon 2020 grant 665385. Author contributions: C.N.B. and J.C. designed and implemented the study and wrote the manuscript; C.N.B. performed the experiments; J.O. helped with the implementation of the study; C.N.B. and M.N. performed cluster cutting; C.N.B., M.N., and F.S. planned the analyses; M.N. and F.S. analyzed the data; M.N. wrote the supplementary methods; and all authors discussed the results and contributed to the manuscript. Competing interests: The authors declare no conflicts of interest. Data and materials availability: Original data and programs were stored in the scientific repository of the Institute of Science and Technology Austria:

Stay Connected to Science

Navigate This Article