Subjects and surgical procedures
Subjects were five male Lister hooded rats purchased from Charles River Laboratories and aged between 9 and 12 months at the time of electrophysiological recordings. Animals were food deprived to ~85–90% of their baseline weight. All animal experiments were carried out in accordance with UK Home Office regulations (UK Animals Scientific Procedures Act of 1986; project license PPL PD8CBD97C). Study protocols were in accordance with the terms of the project license, which was reviewed by the Animal Welfare and Ethical Review Board at University College London. No statistical methods were used to pre-determine sample size. Our study did not use separate groups, so neither randomization nor blinding was used.
Surgery and recording
Rats were anaesthetized with 0.5–1.5% isoflurane. Craniotomies were made bilaterally over the dorsal hippocampus (4.2 mm posterior to the bregma, ±3 mm lateral to the midline). The electrode array, containing 32 tetrodes (16 tetrodes per hemisphere), was implanted onto the surface of the cortex, and electrodes were turned 750 µm into the brain. One bone screw attached to the skull over the frontal cortex served as ground and reference. Tetrodes (nichrome, ¼ Hard Pac coating, 0.0005 inches in diameter; Kanthal, PF000591) were gold plated to <150 kΩ before implantation. Tetrodes were lowered to the dorsal CA1 over 2 weeks, and rats continued to run daily training sessions on the maze. Data were acquired using an Intan RHD USB interface board and RHD headstages.
The honeycomb maze consists of 61 tessellated hexagonal platforms (11.5 cm each side) in an overall hexagonal configuration (total maze width of ~200 cm). Each platform can be raised or lowered independently on a linear actuator, with the raised position ~30 cm higher than the lowered position. Platforms were controlled with digital pulses generated in custom-written software in LabView. The presence of an animal on a given platform was detected using load cells (RobotShop, RB-Phi-117) on which the platforms were affixed. The load cell signal was amplified with a custom-made circuit and input into our custom LabView software.
Animals were initially trained to consume the food reward (honey-flavoured corn flakes) in their home cage. Once they were consuming the food in their home cage, they were brought onto the maze, placed on the reward platform and given a food reward. Once they were consuming the food on the reward platform without hesitation (after 1 or 2 days), we began to run task trials, initially running small numbers of trials and increasing the number gradually in preparation for the recording sessions. The task was run as follows. A list of 13 start platforms was created in MATLAB by randomly choosing a single platform from each of 13 maze subsections. The first start platform was raised, the animal was manually placed on it and the trial was started from the custom LabView software by the experimenter. Two adjacent platforms were then pseudo-randomly selected by the software with two stipulations: first, that at least one of the platforms provided a position closer to the goal than the animal’s current position and, second, that previously unused platforms were selected from first, as long as the first stipulation could be met. The animal’s choice was registered once the load cell system had registered its presence on one of the two choice platforms for a continuous 5 s. This triggered the lowering of the two other platforms, and, after a delay of 4–10 s, another choice sequence commenced. The choice was registered as correct if the animal chose the platform closer to the goal. In some choices, the two choice platforms were the same distance from the goal; these choices were not included in the analysis of behavioural performance. Once the animal reached the goal platform, a food reward (honey-flavoured corn flakes) in a small metal bowl was placed next to the animal after a short delay. Once the animal had finished consuming the reward, the experimenter placed the animal on a holding pedestal next to the maze. Every four trials, the maze was wiped down with 70% ethanol and rotated 30° (alternating between clockwise and anti-clockwise rotations) on a bushing located under the maze to prevent the animal from following scent trails to the goal.
The animal’s ability to correctly navigate the maze was confirmed using the binomial test29 to determine whether the number of correct choices was greater than would be expected by chance given a 0.5 probability of correct choices.
In the first recording sessions (Figs. 1, 3 and 4), animals ran 13 trials of the task, followed by open-field foraging and then a second set of task trials that varied in number among the five animals (13 additional trials for rats 1, 4 and 5 and 7 additional trials for rats 2 and 3). In the second recording sessions (Fig. 2), animals ran 13 trials to goal 1 followed by 13 trials to goal 2; there was a set of ‘learning’ trials interleaved between the two sets, detailed below in the ‘Goal switch training’ section.
Spikes were automatically sorted using KiloSort30 followed by manual curation using Phy31, which consisted mainly of merging and deleting clusters, using autocorrelations and cross-correlations as a guide. Cells with greater than 1% of spikes within the first 2 ms of spike autorcorrelation were excluded from further analysis. Cells were classified as pyramidal or interneuron cells or were left unclassified (excluded from further analysis) on the basis of spike width, mean rate, burst index (n spikes from 0–10 ms of the autocorrelation/n spikes from 40–50 ms of the autocorrelation) and oscillation score32 using a principal-component analysis. Principal components were calculated from the four variables, and the first two principal components were plotted as a scatterplot. Pyramidal cells tended to cluster together, while interneurons were scattered outside the main cluster; the experimenter selected the cells within the cluster by manually drawing a boundary, followed by visual verification of the waveforms.
Video was recorded in custom LabView software using a monochrome USB camera at a frame rate of ~25 frames per second (Imaging Source). Tracking was performed offline using DeepLabCut33. In DeepLabCut, we trained the network to identify two infrared LEDs positioned on top of the animal’s implant, as well as dark fur patches on the shoulders and back. The animal’s head position was taken as the midpoint between the two LEDs, and its head angle was the angle of the line between the LEDs.
Marking lesions were made using 20 µA of anodal current for 10 s. Animals were transcardially perfused with PBS followed by 4% paraformaldehyde (PFA), brains were cryoprotected in 30% sucrose in 4% PFA, and frozen slices of 40 µm were cut. Slices were stained with Cresyl Violet.
Preprocessing of spike data
Hippocampal place cell data are typically velocity thresholded because place cells lose some place tuning when the animal stops moving. In our task, because the animal was not able to move freely around the maze, this seemed an inappropriate approach. Instead, we focused on excluding sharp wave ripple events by using three criteria that had to be met simultaneously: (1) theta power (6–12 Hz) below the mean; (2) population firing rate 2 s.d. above the mean; and (3) ripple power (100–250 Hz) 2 s.d. above the mean. Spectral power was computed by taking the absolute value of the output of the continuous one-dimensional wavelet transform (MATLAB function ‘cwt’). Data were excluded only if all three criteria were met for a minimum duration of 50 ms. Spike data from a given cell were only used for analysis during any of the conditions (honeycomb task, forage, goal 1, goal 2) if at least 500 spikes were fired in the relevant condition after this exclusion.
Relative direction analysis
The field of view was tiled with potential ConSinks, arranged along the x and y axes at ~7-cm intervals (34 × 29 total positions). The head directions relative to each potential sink were then calculated for each spike by subtracting the angle of the vector from the animal’s position to the sink location from the animal’s allocentric head direction. Thus, if the animal was facing the sink, these two directions were equal and the relative direction was 0°. If the animal was facing in the opposite direction, the relative direction was 180°. The convention used in this paper is that positive relative directions indicate that the animal’s head direction was to the left of a line from the animal to the sink (that is, the sink is to the animal’s right) and negative directions indicate a rightward head direction relative to the sink. For a given cell, a binned distribution of relative directions could then be calculated (24 bins spanning −180° to 180°).
This distribution was then corrected for uneven sampling of relative directions by the animal. Because of the potential for differences in sampling of relative direction across the maze, we calculated control distributions of relative direction at each platform (61 platforms in total) using all video frames in which the animal occupied a given platform (an animal was deemed to be occupying a platform if its torso was within the platform’s perimeter). For each cell, the distributions were scaled according to the number of spikes the cell fired on each platform. After scaling, the distributions were summed across platforms. Finally, the cell’s relative direction distribution was divided by the summed control distribution, providing a corrected distribution taking into account any uneven sampling of relative direction by the animal across positions at which the cell fired spikes. From this corrected distribution, using the CircStat toolbox34, we computed the Rayleigh test for non-uniformity of circular data (all ConSink cells were significantly non-uniform) and calculated the mean direction and the MRL. Thus, each cell had an MRL value associated with each potential sink; the candidate sink was taken as the potential sink with the highest MRL. Note that the same correction was also performed for calculation of allocentric head direction tuning.
To test whether a cell was significantly tuned to direction relative to the candidate sink, we shuffled the cell’s head directions such that the head directions were no longer associated with actual positions on the maze. Distributions were calculated as above, yielding MRL values for each xy position. For each of the 1,000 shuffles, the maximal MRL value across all xy positions was used to make a distribution of MRL values. A cell was deemed to be significantly modulated by relative direction if its MRL was greater than the 95th percentile of the control distribution (Extended Data Fig. 5).
To account for disruption of the temporal structure of spiking caused by our shuffling procedure, we performed a second test in which we shifted the spike trains in time (minimum shift of 60 s), recalculating the strength of tuning to the sink position using the shifted positional and head direction values. A control distribution of MRL values was created from 1,000 repeats of the shifts, and cells were discarded if their real MRL was less than the 95th percentile of this control distribution.
Place cells frequently fired in bursts as the animal scanned the environment, causing a smearing of head direction that could potentially lead to false negatives in our search for ConSink cells. Thus, we repeated our search for ConSink cells using only bursts and averaged relative direction and position within each burst to eliminate the smearing effect. Bursts were defined as trains of at least ten spikes fired with interspike intervals of less than or equal to 0.25 s. If two bursts were separated in time by less than 0.5 s, they were combined. If a cell was significant in both analyses (15/77 cells), only the data from the burst analysis, which produced the greatest tuning in all cases, were carried forward into subsequent analyses. In these 15 cells, we confirmed that both analyses identified the same ConSink positions (distance between ConSinks in the same cells = 10.3 cm and in different cells = 85.2 cm, P < 0.001; median difference in preferred relative direction, in the same cells = 3.3º and in different cells = 74.9º, P < 0.001).
To make the vector fields for individual ConSink cells (for example, Fig. 1d–g), the field of view was binned into 20 × 16 spatial bins and a mean head direction value was calculated for each bin with more than 20 spikes.
To confirm the validity of our methodology, we recalculated sink positions and preferred directions using a downsampling method (Extended Data Fig. 5d–i). For each cell on each platform, the spikes were binned according to allocentric head direction (24 bins of 15º in width). Spikes within each bin were then downsampled according to the total directional occupancy in that bin; that is, if the animal spent relatively more time facing a particular direction, the spikes fired in that direction were downsampled by a proportionate amount. For a given cell, spikes were then summed across platforms and the sink position, preferred relative direction and strength of tuning (MRL) were calculated. This was repeated for each cell 1,000 times, and the mean values were calculated.
For population vector fields, bins instead corresponded to maze platforms. For each cell on each platform, we calculated the cell’s mean allocentric head direction. We then created a unit vector from this head direction value and scaled it by multiplying it by both its platform-associated mean firing rate and MRL. Finally, these scaled vectors were summed across all ConSink cells that fired spikes on the platform, and a direction was calculated from the resulting vector. The population sink position was calculated using the same search across xy positions as for individual cells, converting each platform-associated head direction to a relative direction whose contribution was scaled according to the length of its associated vector and calculating the MRL, taking as the sink the position with the maximal MRL value.
To determine what platform an animal was on for post hoc analysis, we tracked the position of the animal’s torso using a dark fur patch behind the shoulders in DeepLabCut33. The animal was considered to be on a particular platform if the torso position was within the platform’s perimeter. For the analysis of correct and error choices (Fig. 4), wait period 2 was defined in relation to the time when the animal’s torso moved onto the chosen platform and was taken as the 4-s window starting 5 s before this transition; the 1-s gap between the end of wait period 2 and the transition to the new platform ensured no contamination of wait period 2 by the transition itself. Wait period 1 was defined as the time after the previous subtrial start when the unchosen platforms had lowered and before the next choice platforms started to be raised.
Goal switch training
In the goal switch trials, all animals ran 13 trials to goal 1. Once these trials were completed, it was necessary to teach the animals that the goal position had switched. To do this, we ran a number of ‘easy’ trials to goal 2. These easy trials were characterized by choice platforms that all led the animal closer to the new goal, such that the animal would arrive at the new goal through no real choice of its own. Once the animal arrived at the new goal, it was rewarded as normal. These trials were interleaved with easy unrewarded trials to goal 1.
The training sequences for each animal were as follows. Rat 1 ran four easy trials to goal 2, followed by a normal trial that it was not able to complete successfully. It then ran two easy unrewarded trials to goal 1, followed by a single easy rewarded trial to goal 2. It subsequently ran 13 normal trials to goal 2, all of which are included in the presented analysis.
The data presented for rat 2 are from the second goal switch session. The first goal switch session was not completed successfully owing to this rat’s inability to learn the new goal location. We subsequently ran brief sessions across 6 days of 1–4 trials to this new goal location in which the rat demonstrated clear learning of this new goal location. We then ran a second goal switch session, using goal 2 from the failed goal switch session as goal 1. After running 13 trials to goal 1, we switched the goal and the animal ran three easy rewarded trials. He subsequently ran 15 normal trials to goal 2. The final 13 of these trials were used in all analyses except for the analysis examining movement of the sinks across the two halves of the goal 2 epoch (Supplementary Fig. 5), which used the first 13 trials.
Rat 3 ran 13 trials to goal 1, followed by 8 easy trials alternating between rewarded trials to goal 2 and unrewarded trials to goal 1. This was followed by 17 normal trials to goal 2. The final 13 of these trials were used in all analyses except for the analysis examining movement of the sinks across the two halves of the goal 2 epoch (Supplementary Fig. 5), which used the first 13 trials.
Rat 4 ran 13 trials to goal 1, followed by 6 easy trials alternating between unrewarded trials to goal 1 and rewarded trials to goal 2 and 3 additional easy rewarded trials to goal 2. This was followed by 21 normal trials to goal 2. The final 13 of these trials were used in all analyses except for the analysis examining movement of the sinks across the two halves of the goal 2 epoch (Supplementary Fig. 5), which used the first 13 trials.
Rat 5 ran 13 trials to goal 1, followed by 8 easy trials alternating between unrewarded trials to goal 1 and rewarded trials to goal 2. This was followed by normal trials to goal 2, but the animal persisted in going to goal 1. A few attempts were made to ‘retrain’ the animal to goal 2 with additional easy trials followed by normal trials, but the animal stopped running the task before running the necessary number of trials to goal 2. This animal was therefore excluded from the main analysis. It was trained to goal 2 over several additional sessions, and data from the final session are presented in Supplementary Figs. 1, 3 and 4.
To determine the dependence of ConSink cell spiking on distance and direction to the sink, we used a technique developed to identify mixed selectivity in individual neurons by quantifying the dependence of spiking on all possible combinations of a set of variables35. A model, corresponding to a particular combination of variables, is trained by optimizing a set of parameters that convert animal state vectors corresponding to the variables of interest into firing rates, which are estimated as an exponential function of the sum of each variable value projected onto the set of parameters. The analysis uses tenfold cross-validation, splitting the data into ten equally sized partitions, training the model using nine of the partitions and testing on the held-out partition such that each partition is tested once. The LLH increase in spike prediction relative to a mean firing rate model is calculated for each model, and the simplest model (that is, the one with the fewest number of variables) that produces a significant increase relative to the mean firing rate model, as well as, in the case of multivariable models, a significant increase over any simpler models (that is, models with fewer variables) is selected as the model that best describes the neuron’s tuning. The significance is assessed using Wilcoxon signed-rank tests comparing the LLH increases for each test partition across the relevant models. We adapted the model to use three variables: (1) relative direction to the sink (RD), (2) distance to the sink (Ds) and (3) direction from the animal’s position to the sink (AD); this produced seven possible models (that is, a three-variable model, three two-variable models and three single-variable models). Firing rate and animal state vectors were constructed using 100-ms windows. Relative direction and direction from position were binned using 18 bins spanning −180° to 180° and 0° to 360°, respectively. Distance to sink was binned using 20 bins from 0 cm to the animal’s maximum distance from the sink.
To calculate the fantail plot (Fig. 3k) showing the population firing rates in the range of head directions relative to the goal, spikes within animals were combined across all ConSink cells. For each spike, the animal’s head direction relative to the goal was calculated in the same way as for head direction relative to the ConSink (see ‘Relative direction analysis’ above). Spikes were then separated according to the platform occupied by the animal during spiking, and, for each platform, its associated spikes were binned according to relative direction to goal (30º-wide bins). Similarly, for each platform, the total amount of time that the animal spent within each relative direction bin was determined. Finally, the spike counts in each bin were divided by the total time (in seconds) spent in each bin, to produce firing rates in each bin. These were then averaged across all platforms and divided by the total number of ConSink cells to generate a per-cell firing rate in binned direction relative to the goal; these values, z-scored within animals, are shown in Fig. 3i and are averaged across animals in Fig. 3k.
To assess remapping between the honeycomb task and open-field foraging, we created rate maps for all cells by partitioning the field of view into 1,280 bins (40 bins in the x direction, 32 bins in the y direction). To establish baselines for cell stability, to which we could compare our remapping data to assess significance, we created rate maps corresponding to the first and second halves of the task and open-field foraging epochs; specifically, for each spatial bin, we calculated the total occupancy (in seconds) and placed the data corresponding to the first and second halves in the corresponding rate maps. Population vector correlations were performed by constructing vectors for each bin using the firing rates of each cell at that bin and then calculating Pearson’s linear correlation coefficient for the two vectors being compared. Similarly, place field correlations were performed by linearizing the two-dimensional rate maps for a given cell and calculating the correlation between the two vectors.
Place field centres
Place fields centres were taken as the centre of mass of the cell’s rate map.
All statistical tests were two sided and non-parametric unless stated otherwise. In box plots, the central mark indicates the median and the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. The whiskers extend to the most extreme data points within 1.5 times the interquartile range from the bottom or top of the box, and all more extreme points are plotted individually using a plus symbol.
Further information on research design is available in the Nature Research Reporting Summary linked to this paper.