Comparison of Head Direction Cell Firing Characteristics Across Thalamo-Parahippocampal Circuitry

Abstract

Head direction (HD) cells, which fire persistently when an animal’s head is pointed in a particular direction, are widely thought to underlie an animal’s sense of spatial orientation and have been identified in several limbic brain regions. Robust HD cell firing is observed throughout the thalamo-parahippocampal system, although recent studies report that parahippocampal HD cells exhibit distinct firing properties, including conjunctive aspects with other spatial parameters, which suggest they play a specialized role in spatial processing. Few studies, however, have quantified these apparent differences. Here, we performed a comparative assessment of HD cell firing characteristics across the anterior dorsal thalamus (ADN), postsubiculum (PoS), parasubiculum (PaS), medial entorhinal (MEC) and postrhinal (POR) cortices. We report that HD cells with a high degree of directional specificity were observed in all five brain regions, but ADN HD cells display greater sharpness and stability in their preferred directions, and greater anticipation of future headings compared to parahippocampal regions. Additional analysis indicated that POR HD cells were more coarsely modulated by other spatial parameters compared to PoS, PaS, and MEC. Finally, our analyses indicated that the sharpness of HD tuning decreased as a function of laminar position and conjunctive coding within the PoS, PaS and MEC, with cells in the superficial layers along with conjunctive firing properties showing less robust directional tuning. The results are discussed in relation to theories of functional organization of HD cell tuning in thalamo-parahippocampal circuitry.

Introduction

Head direction (HD) cells fire whenever an animal points its head in a particular direction, independent of ongoing behavior (Taube et al., 1990a; 1990b; Taube, 1995; 2007). The direction of maximum response (referred to as the cell’s ‘preferred direction’) varies from cell to cell, and firing characteristics remain stable across recording sessions many days apart. The cell’s preferred direction can be controlled by environmental cues (Taube et al., 1990b; Taube, 1995; Yoder et al., 2011b), but is generated and maintained by self-motion cues (e.g., vestibular, proprioceptive, optic flow) through angular path integration (Arleo et al., 2013; Clark & Taube, 2012; Taube and Burton, 1995; Goodridge et al., 1998; Yoder et al., 2011a). HD cells were first described in the postsubiculum (PoS) (Taube et al., 1990a), but have subsequently been identified in several limbic regions comprising the classic “Papez” circuit including the anterior dorsal thalamic nucleus (ADN) (Taube, 1995), lateral mammillary nuclei (Blair et al., 1998; Stackman and Taube, 1998) and retrosplenial cortex (Chen et al., 1994; Cho & Sharp, 2001; Jacob et al., 2017). More recently, cells with directional correlates have been identified in parahippocampal regions including the medial entorhinal cortex (MEC) (Sargolini et al., 2006), parasubiculum (PaS) (Boccara et al., 2010), and postrhinal cortex (POR) (Gofman et al., 2019; LaChance et al., 2019).

Models of HD signal processing have proposed that the directional signal is generated within a midbrain-diencephalic circuit, while the directional signal is utilized within thalamo-parahippocampal connectivity for navigational functions (Fig. 1A; Clark & Taube, 2012; Taube, 2007; Yoder et al., 2015). In support of this framework, studies have shown that damage to subcortical structures including the ADN can abolish HD cell tuning in “downstream” parahippocampal regions, while damage to cortical regions do not abolish HD signaling in “upstream” subcortical regions (reviewed in Clark & Taube, 2012; Taube, 2007). Instead, projections stemming from the PoS and retrosplenial cortex stabilize the ADN HD signal with respect to environmental landmarks (Clark et al., 2010; Goodridge & Taube, 1997). Regional differences in basic directional firing characteristics have been reported, which suggests functional specialization within the thalamo-parahippocampal circuitry. For instance, ADN cells exhibit higher peak firing rates and convey more HD information relative to PoS (Peyrache et al., 2015; c.f. Blair and Sharp, 1995; Taube & Muller, 1998), and can more accurately decode the animal’s current HD compared to PoS, PaS, and MEC HD cells (Viejo et al., 2018; Xu et al., 2019). In addition, ADN and PoS cells differ with respect to their anticipatory firing characteristics (Blair & Sharp, 1995; Blair et al., 1997; Taube & Muller, 1998). Specifically, during head movements, HD cells recorded in ADN anticipate, or fire maximally, just before the animal’s head reaches the cell’s preferred firing direction. In contrast, PoS HD cells tend to fire maximally at, or shortly after, the animal’s head passes through the cell’s preferred direction.

Figure 1.

A. Circuit diagram showing the principal connections (arrows) between the anterodorsal thalamus (ADN), postsubiculum (PoS), parasubiculum (PaS), medial entorhinal cortex (MEC), and postrhinal cortex (POR). B. Representative electrode tracks in Cresyl violet stained sections for each region. C. Ten HD cell tuning curve plots (firing rate x HD) are shown for each region. Cells were selected based on their mean vector length: the 5 largest and 5 lowest mean vector length values are shown. Peak firing rate (top) and mean vector length (bottom) are reported above each tuning curve. D. Heat maps for each region show the population of classified HD cell tuning curves sorted by preferred firing direction (bin with maximum peak firing rate). Rows show the individual tuning curve of a cell with binned firing rates normalized on a scale from 0 to 1. E. Angular difference (degrees) between the preferred firing directions of HD cells recorded on the same tetrode in the same recording session. F. Representative tuning curves of bidirectional HD cells are shown for each region. Peak firing rate (top), mean vector length (middle), and bidirectional index (bottom) are reported above each tuning curve.

Recent evidence also suggests that the firing characteristics of HD cells can be organized based on regional topography within the parahippocampal cortex and can vary with conjunctive firing characteristics (Caccuci et al., 2004; Giocomo et al., 2014; Kornienko et al., 2018; Zutshi et al., 2018). For instance, Giocomo et al. (2014) reported that while HD cells located in the dorsal portions of MEC layer III exhibit sharp directional tuning, HD cells in ventral MEC layer III are more coarsely modulated by HD. Consistent with this finding, Zutshi et al. (2018) reported a distinct population of broadly tuned HD cells in the superficial layers of MEC. Finally, Kornienko et al. (2018) identified two distinct populations of HD cells in the mouse MEC: a class of theta-modulated HD cells that were more likely to maintain their fixed orientation relative to other HD cells during environmental cue manipulations, and a class of non-theta modulated HD cells that uncouple their orientation from other HD cells. Together, these findings support the hypothesis that parahippocampal regions contain more than one directional representation, possibly guided by distinct types of sensory information (Dudchenko et al., 2019; Munn & Giocomo, 2020; Taube, 2017). Currently, however, no study has quantified these apparent differences amongst different HD cell populations across multiple brain areas.

The aim of the present study was to perform a comparative assessment of HD cell firing characteristics within the thalamo-parahippocampal HD cell circuit. We therefore analyzed recordings from HD cells in the ADN, PoS, PaS, MEC, and POR while rats foraged for randomly scattered food pellets in a circular or square open fields. From these recordings we compared directional firing characteristics across the five regions and across laminar position within parahippocampal cortices. We also identified and then analyzed “conjunctive” HD cells that were co-modulated with other variables such as spatial location (grid, border, place cell activity) (Boccara et al., 2010; Peyrache et al., 2017; Sargolini et al., 2006; Solstad et al., 2008), egocentric orientation or distance relative to the center of the environment (Gofman et al., 2019; LaChance et al., 2019; Peyrache et al., 2017), linear or angular head velocity (Chen et al., 1994; Hinman et al., 2016; Kropff et al., 2015; Sharp, 1996; Solstad et al., 2006; Taube, 1995; Taube & Muller, 1998), and locking to theta frequency (Boccara et al., 2010; Brandon et al., 2013). Non-conjunctive and unidirectional cells were classified as “pure” HD cells. Here, we report three general findings: (1) ADN and POR HD cells represent two distinct and different sets of firing characteristics within thalamo-parahippocampal circuitry (sharp and coarse HD cell coding, respectively), (2) PoS, PaS, and MEC HD cells are similar across a broad range of HD firing measures, and (3) HD tuning varies with laminar position and conjunctive coding within the PoS, PaS and MEC.

Materials & Methods

Subjects

Some of the data analyzed in the present study are from previous studies (Butler & Taube, 2017; Winter et al., 2015; LaChance et al., 2019; 2022). All recordings were from female Long-Evans rats (n = 37). Rats weighed 240–340 g at the beginning of testing and were housed singly in Plexiglas cages and maintained on a 12hr light/dark cycle. Water was provided ad libitum while access to food was restricted as necessary to maintain the animal’s body weight in the range of 85–90% of its free-feeding weight. All procedures involving the rats were performed in compliance with institutional standards as set forth by the National Institutes of Health Guide for the Care and Use of Laboratory Animals.

Electrode Implantation Surgery

Rats were surgically implanted with a moveable microdrive containing four tetrodes that was similar in design to those described in previous studies (LaChance et al., 2019; Winter et al., 2015a; 2015b) and targeted the PoS (n =4), PaS and MEC (n = 15), or POR (n = 7). Rats in the ADN group (n = 11) were implanted with a microdrive containing eight independently driveable stereotrodes (Butler & Taube, 2017). Tetrodes were constructed by twisting together four strands of 17 μm nichrome wire, while stereotrodes were constructed by twisting together two strands of the same wire. Tetrode arrays were constructed by threading tetrodes through a 26-gauge stainless steel cannula and each wire of the tetrode bundle was connected to a pin of a Mill-Max connector. Two center pins of the connector were secured to the cannula to act as an animal ground. Three drive screws were fixed to the connector using dental acrylic creating a microdrive that was moveable in the dorsal/ventral direction. Stereotrode arrays targeted the ADN bilaterally (four stereotrodes per hemisphere). A single stereotrode was inserted in one of eight cannula and affixed with epoxy; the unspun ends of each stereotrode were threaded through the electrode cap and connected to an electrode interface board. Tetrode arrays were implanted above the dorsal-caudal cortex of the right hemisphere with the intention of advancing the tetrodes through the PaS and MEC in one group of rats, the POR in another group of rats, and PoS in another group of rats. The following coordinates were used for electrode implantation in PaS, MEC, and POR: 0.25 – 0.45 mm anterior to the transverse sinus, 4.2 – 4.6 mm lateral to lambda, and 1 – 1.5 mm below the cortical surface. Tetrode arrays targeting PaS, MEC, and POR were slightly angled in the sagittal plane (~10°), such that the tetrodes were pointing in the anterior direction. The following coordinates were used for electrode implantation in the PoS: −7.1 to −7.9 mm relative to bregma, 3.4 to 4.0 mm lateral to bregma, and 1.5 mm below the cortical surface. ADN-targeting electrodes were implanted at the following coordinates: −1.8 mm relative to bregma, 1.2 mm lateral to bregma, and −3.7 mm below the cortical surface. Each electrode array was fixed to the skull using dental acrylic.

Neural Screening and Recording

After 7 days of recovery from surgery, neural activity in the PoS, PaS, MEC, and POR was monitored or “screened” for cellular activity while animals foraged for 20 mg food pellets scattered on the floor of a gray square box (120 × 120 cm; 50 cm in height). The floor of the box was composed of gray photographic backdrop paper and the inside was featureless except for a white cue card attached to the wall (71 × 50 cm). The cue card was maintained at the same position throughout the experiment. Neural activity in the ADN was monitored while animals freely foraged for sucrose pellets in a 76-cm-diameter gray cylinder. A white cue card affixed to the inner wall of the cylinder (~100° of arc) served as a visual landmark. A black floor-to-ceiling curtain (2.5 m in diameter) surrounded each enclosure and eight uniformly arranged lamps were located above the box and cylinder to provide illumination. Olfactory cues were minimized by replacing the floor paper in each environment between sessions, and auditory cues were masked by playing white noise from an overhead speaker. A color video camera was centered above each enclosure.

To detect cell activity, electrical signals were pre-amplified by unity-gain operational amplifiers on the headstage of the animal (HS18-LED). The signals were differentially recorded against a low activity wire from another electrode and then bandpass filtered (600 Hz to 6 KHz) using a Digital Lynx SX Data Acquisition System (Neuralynx, Bozeman, MT). Signals that crossed a pre-set amplitude threshold (30–50 μV) were time-stamped and digitized at 32 KHz for 1 msec. Also attached to the headstage of the animal were red and green light-emitting diodes (LED) secured ~8 cm apart above the head and back of the animal, respectively. An automated video tracking system provided x-y coordinates of each LED position at a sampling rate of 60 Hz. Monitoring of cellular activity continued until each of the electrodes had been examined. If cellular activity exceeded two times the background noise level, the signals were recorded for ~10 to 20 min. If cell activity was not isolatable from background noise, the electrode array was advanced 25–100 μm and the animal was returned to its home cage and monitored the next day.

Offline Spike Sorting

Spike sorting was conducted offline using cluster cutting software (Kilosort, Pachitaru et al., 2016; SpikeSort3D, Neuralynx). Waveform characteristics from each tetrode were plotted as scatter plots from one of the four tetrode wires versus another wire of the same tetrode. Several waveform characteristics were used to isolate cells, including amplitude, peak, valley, width, and principal components. Individual cells formed clusters of points in these plots and the boundaries were identified and manually “cut”. Single cell isolation was further confirmed by the observation of clear refractory periods around 0 μsec in the inter-spike interval histogram. Only cells that were clearly isolated from other cells were included in further analysis.

Classification of HD cells

The HD of the animal was determined by the relative position of the red and green LEDs. The amount of time and number of spikes in each HD was sorted into sixty 6° bins. The firing rate for each 6° bin was determined by dividing the number of spikes by the amount of time at that HD and a firing rate by HD plot was constructed for each cell in the data set. We used a dual criterion comprised of mean vector length and directional stability to classify thalamo-parahippocampal neural activity as an HD cell. First, we computed the mean vector length (Rayleigh r) for each cell. This value represents the concentration of firing across all directions and is equivalent to the circular mean in circular statistics (Batschelet, 1981). The mean vector length ranges between 0 and 1, with higher values indicating that spike occurrence is clustered around a particular direction. Next, we measured the directional stability of each cell by dividing the recording session into four equal time bins and cross-correlating the 60 directional firing bins across each time bin and averaging these values (Stability = (Q1-Q2 + Q1-Q3 + Q1-Q4 + Q2-Q3 + Q2-Q4 + Q3-Q4)/6). We next compared each cell’s mean vector length and directional stability with the same measures calculated from the cell’s shuffled tuning curve. Briefly, each cell had its spike data randomly shuffled relative to time 400 times and mean vector length and directional stability was calculated for each iteration. Cells were classified as HD cells if their mean vector length and directional stability measures exceeded the 95^th percentile level and had mean vector length > 0.3 and directional stability measures > 0.2. In addition, we removed repeat recordings and cells without clear directionality.

Previous studies have shown that inhomogenous sampling of position and HD of an animal’s locomotor behavior can influence measurement of HD cell classification and basic firing characteristics (Burgess et al., 2005; Muller et al., 1994). Thus, we also considered the potential for directional sampling bias in our classification of HD cells. We applied a Maximum Likelihood Model to separate the relative influences of HD and spatial location on cell spiking as described by Burgess and colleagues (2005; also see Caccuci et al., 2004). Cells were removed from further analysis if their “corrected” mean vector length and directional stability measures failed to exceed the criteria cutoff calculated from the “un-corrected” tuning curves (described above). All remaining analyses were conducted on uncorrected tuning curves.

Basic Firing Characteristics of HD cells

Several measures of directional firing characteristics were calculated for each cell (Taube et al., 1990a; Taube, 1995). These included the preferred firing direction, peak firing rate, background firing rate, directional firing range, directional coherence, directional information content, coefficient of variation, and anticipatory time interval. Each measure is described in detail below.

Preferred firing direction.

The preferred firing direction was defined as the directional bin with the highest firing rate.

Peak firing rate.

The peak firing rate was the firing rate corresponding to the preferred firing direction.

Background firing rate.

The background firing rate was defined as the mean firing rate of all bins ≥18° outside the cell’s directional firing range.

Signal-to-noise ratio.

The signal-to-noise ratio was calculated by dividing the peak firing rate by the background firing rate.

Directional firing range.

The directional firing range was defined as the width at the base of a triangle fit to a firing rate vs. HD plot.

Directional coherence.

Directional coherence is a measure of the smoothness in the firing rate vs. HD tuning curve. It was calculated using methods similar to those used for determining the spatial smoothness in the firing rates of hippocampal place and theta cells (Kubie et al., 1990). The firing rate for each directional bin is correlated to the firing rates in the two immediately adjacent neighboring bins (CW and CCW directions) and an overall correlation is calculated across all bins. We would expect this measure to be high for cells that have a strong, continuous, and smooth looking tuning curve and low for cells that have a jagged, irregular, and uneven looking tuning curve. It is possible that conjunctive cells, because they are strongly tuned to a second parameter, might have a lower directional coherence.

Directional information content.

Directional information content is a measure of how many bits of HD information is conveyed by each spike (Skaggs et al., 1993) and was calculated by the following formula: directional information content = p_i (l_i/l) log₂ (l_i/l), where p_i is the probability that the head pointed in the ith directional bin, l_i is the mean firing rate for bin i, and l is the mean firing rate across all directional bins.

Anticipatory time interval.

The anticipatory time interval (ATI) is a measure of the amount of time that cell firing best predicts where the animal will be pointing its head in the future or past. Previous work has estimated that the activity of ADN HD cells anticipate future HDs by ~25 msec, but HD cells in the PoS generally do not anticipate or lag their preferred directions (~0 msec; Blair & Sharp, 1995; Taube & Muller, 1998). The anticipatory firing characteristics of HD cells in the PaS, MEC, and POR is presently unknown. We therefore compared the ATI of cells recorded from the ADN, PoS, PaS, MEC, and POR using the methods of Blair and Sharp (1995). First, firing rate vs. HD plots were constructed for each HD cell by dividing the 360° directional range into sixty 6° bins and then calculating the average firing rate for each bin. Firing rate vs. HD plots were then constructed for CW and CCW directions and the difference between the preferred firing directions for the two functions (i.e., separation angle) was determined. The spike record was then shifted forwards and backwards in time in steps of 16.67 msec (the maximum temporal resolution of the recording hardware) and the separation angle between the CW and CCW functions for head turns ≥ 45°/sec computed for each shift. The spike series was shifted incrementally ±6 times (± 100 msec) relative to the HD series, providing 13 values of CW-CCW separation angles. A scattergram was then constructed from the 13 CW-CCW separation angles and their corresponding time shift. The x-intercept of the best-fit line of this plot is referred to as the anticipatory time interval and is equivalent to the amount of time that the spike series has to be shifted to achieve overlapping CW and CCW functions. Only HD cells with correlations >0.5 were included in the ATI analysis.

Burst index.

Previous work has identified populations of parahippocampal cortical cells that tend to discharge in bursts as opposed to more continuous firing (Coletta et al., 2018; Ebbesen et al., 2016; Latuske et al., 2015; Quinlan et al., 2020). These two modes of firing are usually distinguishable based on their interspike interval (ISI) histograms. Cells that discharge in bursts overall have shorter ISIs than non-bursty cells. To measure the extent of burst spiking by HD cells across the different brain areas, we calculated a burst index score for each cell.

To calculate the burst index, we first produced interspike interval (ISI) histograms (0–100 msec) from the spike timestamped data for each cell. Next, we computed a burst index score by counting the number of ISIs that were <10 msec, but > 2 msec due to the refractory period, and then dividing this value by the total number of ISIs that occurred between 0–100 msec (Quinlan et al., 2020). For this calculation we only used spikes that occurred when the animal’s head was within ± 30° of the cell’s preferred firing direction (PFD) (60° range). Thus, an HD cell that discharges in bursts would have a burst index measure closer to 1. Only cells with at least 50 spikes in their ISI histogram were included in this analysis.

Coefficient of variation.

We also investigated whether there were differences in the regularity of spiking across HD cells from each brain area. There are two fundamental modes that have been postulated for how neurons fire: a rate code and a spike timing code. In a rate code, information is encoded as the average firing rate of the neuron over a specified time interval. In a spike timing code, information is encoded at the level of single spikes and what becomes important is precisely when the neuron fires relative to other neurons. Cells that use rate coding fire in a more regular (tonic) firing pattern, whereas cells that use a spike timing code fire in a more irregular firing pattern. Which of these firing patterns is present in a particular cell can be determined by measuring the variation in the cell’s ISIs. This measurement is referred to as the cell’s coefficient of variation (CV) and is calculated by dividing the standard deviation of all the ISIs by the mean ISI. A regular, uniform rate of firing would have a CV score near 0 and a score approaching or >1 would indicate a cell with a highly irregular rate of firing.

From the spike timestamped data and the histograms we constructed to calculate the burst index above, we determined each cell’s CV, which is defined as the ratio between the standard deviation of the ISIs and the mean ISI. We note that this procedure is different from one we employed in an earlier analysis (Taube, 2010), where the upper acceptable ISI was set at 100 mec as opposed to a variable upper limit that was based on the cell’s peak firing rate. This approach was selected because of the wider variation in peak firing rate ranges across HD cells from each of the brain areas. As with the burst index analyses, only cells with at least 25 spikes in their ISI histogram were included in this analysis.

Linear and Angular Velocity Modulation of HD cells

To investigate whether HD cells were modulated by an animal’s linear or angular head velocity (AHV), we correlated these measures from HD cells across the different brain regions. For linear velocity we first computed the instantaneous speed of the animal for each 1/60^th sec sample, by fitting a best-fit line over a 5-sample window for the x and y dimensions (Bassett et al., 2007; Mehlman et al., 2019). The slopes of the best-fit lines were defined as the change in the x and y dimensions, respectively, and the instantaneous speed for the center time point of the window was defined as the square root of x² + y². We then constructed a linear velocity tuning curve by sorting samples based on their linear velocity (1 cm/s bins) along with their corresponding number of spikes. Next, the number of spikes in each linear velocity bin were summed and divided by the total time in that bin to yield an average firing rate for that linear velocity bin. A firing rate vs. linear velocity plot was then created and a best-fit linear line was computed between 0–30 cm/s. From this linear fit we calculated a Pearson’s r correlation coefficient and its corresponding slope. We classified a cell as modulated by linear velocity if the absolute value of its correlation ≥ 0.7 and the absolute value of its slope ≥ 0.1.

For AHV we constructed a firing rate versus AHV tuning curve for each cell based on the cell’s instantaneous AHVs for each sample (1/60^th sec). Instantaneous AHVs were computed from the HD values by first creating a HD by time function that was smoothed across five time points using the following function: n = (n_t-2 + n_t-1 + n + n_t+1 + n_t+2)/5. From the smoothed HD x time function, we took episodes of five consecutive samples and calculated the best-fit slope (derivative) and this value was defined as the AHV for the center value of the five samples. This procedure was repeated for all samples across the entire session. Then each sample’s AHV, along with the number of corresponding spikes for that sample, were sorted into 6°/s bins. Each AHV bin was summed to determine the total number of samples and spikes for that bin. Dividing the number of spikes by the number of samples yielded an average firing rate for that AHV bin. We then constructed a firing rate vs. AHV tuning curve based on each AHV bin’s values. From this plot we determined two best-fit lines: one for values between −204 and 0°/s (CW values) and one for values between 0 and 204°/s (CCW values). We then calculated the correlation and slope values for each best-fit line. HD cells were classified as AHV modulated if the correlation (Pearson’s r) of either the CW or CCW best-fit lines ≥ 0.5 and the absolute value of their slope ≥ 0.1 °/s (which is equivalent to an increase in firing of 1 spike/10°/s).

Bidirectional cells

A bidirectionality score (Kornienko et al., 2018) was calculated by measuring the firing rate of the second largest peak of the HD tuning curve divided by the firing rate of the largest peak of the tuning curve. HD cells were classified as bidirectional if 1) their bidirectional score was greater than 0.2 and 2) the ratio between the tuning curve with the lowest firing rate and the trough with the highest firing rate exceeded 1.25.

Theta Modulation of HD cells

Theta modulation of HD cells was determined using a theta index (Boccara et al., 2010; Harvey et al., 2018; LaChance et al., 2019; Yartsev et al., 2011). We first created a temporal autocorrelogram for each cell by summing the number of spikes that occurred within each 5-msec bin of a 1-s window centered on each spike. A power spectrum was then computed by performing a fast Fourier transform on the autocorrelogram. The strength of theta modulation (theta index) was computed by first calculating the mean power within 1 Hz on either side of the frequency with the highest power within a 5- to 11-Hz (theta) range and dividing this by the mean power between 1 and 125 Hz. Cells were classified as theta modulated if their theta index exceeded 5.

Grid Modulation of HD cells

The rat’s position was estimated from the red LED and was sorted into a 2 × 2 cm bin matrix of firing rate by time. The matrix was then subjected to smoothing using a Gaussian filter with a standard deviation of two bins. From the smoothed rate maps, spatial autocorrelation maps were computed using similar methods as previous studies (Hafting, et al., 2005; Sargolini, et al., 2006; Winter et al., 2015a; 2015b). First, a Pearson’s correlation was conducted between a cell’s smoothed firing rate map and a copy of the same rate map that was shifted to a particular x-y offset. A Pearson’s correlation was computed for all possible combinations of x-y offsets, and the resulting correlation value at each offset was entered in a new matrix. Grid cells, which fire in a triangular array spanning the entire environment, are defined by the 6-fold symmetry of the hexagon surrounding the central peak within their corresponding spatial autocorrelation maps. We quantified this symmetry by taking a circular sample of the spatial autocorrelation map centered on the central peak, but with the central peak excluded, and computed a Pearson correlation between the circular sample and its rotated copy (in 6° increments ranging from 0 to 180°). The grid score was calculated using a “moving radius” technique, in which the circular correlation was repeated multiple times varying the outer diameter of the circular sample. The outer diameter was increased in steps of 1 bin from a minimum of 7 bins from the central field until it reached the edge of the autocorrelogram. A grid score was calculated from the difference between the highest Pearson correlations at angles of 60° or 120°, and the lowest of 30°, 90° or 150°. Based on qualitative assessment of rate maps, and to provide a conservative estimate of grid cell numbers, cells were classified as displaying grid-like firing only when their grid scores exceeded 0.8.

Border Modulation of HD cells.

A border score was calculated as described in LaChance et al. (2019; see also Kropf et al., 2015; Solstad et al., 2008). Briefly, smoothed rate maps were first thresholded to only include bins higher than 20% of each cell’s maximum firing rate. Firing fields were defined as above-threshold contiguous groups of bins with size > 200 cm. We next determined the firing field with the most bins along one wall of the enclosure, and converted that number of bins into a distance along that wall, d. We then calculated the average distance of each of the bins in that firing field from the associated wall, a. The border score was then computed according to the following equation: B = (d – a) / (d + a). From the smoothed rate maps, we also computed the spatial information content for each cell, which describes how much information is conveyed regarding the spatial location of the animal by each spike discharge. The spatial information content was calculated as described by Skaggs et al. (1993). Units were only classified as border cells if 1) they had border scores that exceeded 0.5, 2) they passed the Generalized Linear Model classification procedure for position modulation (see section below), 3) they had spatial information content scores > 1.0, and 4) they did not pass the grid cell criterion.

Place Modulation of HD cells

HD cells were classified as place modulated if 1) they passed the Generalized Linear Model classification procedure for position modulation (see section below), 2) they had spatial information content scores > 1.0, 3) they did not pass the grid cell criterion, and 4) they did not pass the border cell criterion.

Egocentric Bearing and Distance Modulation of HD cells

Egocentric bearing and distance modulation were calculated as described in LaChance et al. (2019). From the reference point (i.e., environment center), we first computed the allocentric bearing of that location from the animal (defined as the angle between the positive x axis with origin centered on the animal and a line drawn from the animal to the reference point) for each time point in the recording session, using the following equation: Allocentric Bearing = arctan2(yref – yanimal, xref – xanimal). The egocentric bearing of the reference point relative to the animal was then computed by subtracting the animal’s allocentric HD at each time point: Egocentric Bearing = Allocentric Bearing - HD. An egocentric bearing of 0° (“egocentric north”) would indicate that the reference point was in front of the animal (allocentric bearing equal to allocentric heading), whereas an egocentric bearing of 180° indicates that the reference point was behind the animal. Bearings of 90° and 270° indicate bearings to the left and right of the animal, respectively.

Egocentric bearing tuning curves were then created using 12° bins. For each cell, the amount of time that each bin was sampled, and the number of spikes fired per bin over the course of a session were calculated, and the tuning curve was computed by dividing the number of spikes per bin by the amount of sampling time per bin. The mean vector length and mean angle were then extracted to indicate tuning strength and preferred firing direction, respectively. For center-bearing tuning (egocentric bearing of the geometric center of the environment) a single period cosine curve was also fit to the center-bearing tuning curve, and the explained variance (R2) value of the fit was calculated as the cosine tuning strength. Units were only classified as center-bearing cells if they (1) passed the Generalized Linear Model classification procedure for center-bearing modulation (see section below), and (2) had mean vector lengths > 95th percentile of a shuffled distribution (same procedure as described for HD cell classification).

Egocentric distance was defined as the instantaneous distance between an animal and a given reference location. Egocentric center distance was calculated as the distance between the animal and the geometric center of the environment. For tuning curve construction, 4-cm bins were used. For each cell, the number of spikes fired over the course of the session per bin was divided by the amount of time the animal spent in each bin. A regression line was fit to each curve and the R2 fit of the line was calculated as the linear tuning strength. Cells were only classified as center-distance cells if they (1) passed the Generalized Linear Model classification procedure for center-distance modulation (see section below), and (2) had linear fit values > 95th percentile of a shuffled distribution.

Generalized Linear Model (GLM)

Ten-fold cross-validation with a multiplicative Poisson generalized linear model (Hardcastle et al., 2017) was used to classify single neurons as encoding one of three spatial variables: center bearing, center distance, and spatial location. For a given model, the firing rate vector r for a single cell over all time points was modeled as follows:

$$r=\mathrm{e}\mathrm{x}\mathrm{p}({\sum }_i{X}_i^T{\beta }_i)$$

where X is a matrix containing animal state vectors for a single behavioral variable, $\beta$ represents the parameter vector for that behavioral variable (similar to a tuning curve), and i indexes across behavioral variables included in the model. The parameter vectors for a given model are learned by maximizing the log-likelihood of the real spike train n given the model’s estimated rate vector r:

$$l=\mathrm{ }\sum _t{n}_t\log \left(r_t\right)-\mathrm{ }{r}_t-\log \left(n_t!\right)$$

where t indexes over time points. In order to avoid overfitting for the cross-validation procedure, an additional smoothing penalty P was added to the objective function which penalizes differences between adjacent bins of each parameter vector:

$$P=\mathrm{ }\sum _{i}S\sum _j\frac{1}{2}\mathrm{*}{({\beta }_{i,j+1}-\mathrm{ }{\beta }_{i,j})}^2$$

Here, S is a smoothing hyperparameter (20 for all variables), i indexes over variables, and j indexes over response parameters for a given variable. Response parameters were estimated by minimizing (P – $l$) using SciPy’s optimize.minimize function. To control for potential behavioral confounds, the full model included the following variables: allocentric head direction, egocentric center bearing, egocentric center distance, spatial position, linear speed, and angular head velocity. Thirty bins were used for center bearing and allocentric head direction parameter vectors, and ten bins were used for center distance. Linear speed coding used 10 bins (from 0 m/s to 40 m/s) and angular head velocity used 20 bins (from −250 deg/s to +250 deg/s). Binning for spatial position depended on the size of the arena. For the 120 × 120 cm arena, 20 × 20 bins of side length 6 cm were used. For the 70 cm diameter cylinder, 12 × 12 bins of side length ~5.83 cm were used.

For cross-validation, data for a session was split into training (9/10) and test (1/10) data. Parameter vectors were estimated by minimizing the objective function on the training data using the full model with all variables. Drawing parameter estimates from the full model helps to reduce correlation/artifacts between variables and makes models with different variable combinations more comparable (Burgess et al., 2005). Log-likelihoods for models with all possible variable combinations were computed. This was repeated until all portions of the data had been used as test data (10 folds).

For model selection, the log-likelihood values from the best two-variable model were compared to those from the best one-variable model. If the two-variable model showed significant improvement from the one-variable model (using a one-sided Wilcoxon signed-rank test), then the best three-variable model was compared to the two-variable model, and so on. If the more complex model was not significantly better, the simpler model was chosen. If the chosen model performed significantly better than an intercept-only model, the chosen model was used as the cell’s classification. Otherwise, the cell was marked ‘unclassified’.

Principal Component and Cluster Analyses

A principal component analysis (PCA) was performed based on 12 parameters: peak firing rate; directional firing range; background firing rate, signal-to-noise ratio, directional coherence, directional informational content, Rayleigh r, linear velocity correlation, CCW AHV correlation, CW AHV correlation, and theta and burst indices. A second analysis used k-means clustering to determine the optimal number of distinct clusters the data could be divided into. We used both a k-means silhouette analysis (python: kmeans.intertia_) and a k-means elbow analysis (python: sklearn.metrics.silhouette_score) to determine the optimal number of clusters in the data set.

Histological analysis

At the completion of the experiment, animals were deeply anaesthetized with sodium pentobarbital and a small anodal current (20 μA, 10 sec) was passed through 2 or 3 electrode wires. The rats were then perfused intracardially with saline followed by 10% formalin solution. Each brain was removed from the skull and was post-fixed in a 10% formalin solution for at least 24 hrs. The brains were then cryoprotected in a 20% sucrose solution for at least 24 hr and subsequently frozen. The brain was then blocked such that sagittal sections (30 μm thick) could be taken through the parahippocampal cortices and coronal sections through the ADN using a cryostat. Each section was mounted on glass microscope slides, stained with thionin, and examined under light microscopy to determine the location of recording sites. Locations of recorded cells were determined by measuring backward from the most ventral location of the marking lesions or, if marking lesions were not visible, the electrode tracks (Fig 1B). Delineations of parahippocampal regions were based on Boccara et al. (2010; 2015) and Burwell et al. (2001).

Statistics

Custom software in Matlab (The MathWorks, Natick, MA), Python, and LabVIEW (version 5.0; National Instruments, Austin, TX) was used to acquire, visualize, and analyze neural activity as previously described (LaChance et al., 2019; Taube, 1995; 2010). Between region and cell type comparisons were computed using one-way ANOVAs, independent sample t-tests, and post-hoc comparisons. For ANOVAs and post-hoc tests, violations of homogeneity of variance (Levene’s test) were corrected using Welch’s F test and Games-Howell post-hoc test. All statistical analyses were conducted using Jamovi statistical software (Version 2.3; https://www.jamovi.org).

Results

Robust HD tuning varies across thalamo-parahippocampal regions

Neurons were categorized as HD cells when measures of mean vector length and directional stability exceeded the significance level estimated from their shuffled tuning curves and had a mean vector length > 0.3 and a directional stability measure > 0.2. Overall, a total of 561 cells were classified as unidirectional HD cells in the ADN (n = 107), PoS (n = 111), PaS (n = 133), MEC (n = 140), and POR (n = 70) (Fig. 1C, D). Consistent with previous studies (Giocomo et al., 2014; Taube, 1995; Taube et al., 1990a), we observed a uniform distribution of HD cell preferred firing directions (PFDs) across 360° in each thalamo-parahippocampal region (Fig. 1D). As shown in recent calcium imaging data from the mouse anterior thalamus (Ajabi et al., 2023), we did not find evidence of topographic organization in cell PFDs, i.e., neighboring HD cells have similar PFDs or maintain some fixed value apart from one another. We examined this issue by measuring the difference in PFDs for cell pairs that were recorded on the same tetrode in the same recording session (Fig. 1E). In theory, cells that are recorded on the same tetrode should be in close proximity to one another (≥ 50 μm) (Gray et al., 1995; Buzsaki, 2004; Pedreira et al., 2012). Fig. 1E plots the distribution of angular differences across all cell pairs for the five different regions (note that the angular difference between two cells can range from 0–180°) and shows that there was a relatively uniform distribution of angular differences (χ² Goodness of fit, ps ≥ 0.68), indicating that there was little evidence for neighboring HD cells to contain some fixed value between their PFDs.

Also consistent with previous work (Kornienko et al., 2018), we identified HD cells that exhibited bidirectionality in their tuning (i.e., cells containing two distinct peaks in the tuning curve). As expected, many bidirectional cells were observed in parahippocampal regions (n = 25), although some were also observed in the ADN (n = 16; see Fig. 1F). Each of these cells’ waveforms was well-isolated; it was therefore unlikely that the bidirectionality could be attributed to poor cell isolation. Because these cells were few in number (~6%) and have not been studied in detail before in rats under control conditions, the remainder of this study focuses on the characteristics of unidirectional HD cells.

We first compared unidirectional HD cells from each region across a total of eight measures of directional tuning (peak firing rate, background firing rate, signal-to-noise ratio, mean vector length, stability, directional firing range, directional information content, and directional coherence) (Fig. 2; see inset for results of post-hoc comparisons). Our analysis indicated that each measure varied significantly across thalamo-parahippocampal regions (all Fs ≥ 14.8, ps ≤ 0.001). We observed three general patterns. First, HD cells with a high degree of directional specificity were observed in all five brain regions (see Fig. 1C, D). Nevertheless, there was considerable variability in the sharpness of directional tuning within each brain region. Figure 2 plots the cumulative proportion of mean vector length and directional information content values for HD cells in each brain region. Interestingly, distributions for each region reflect diversity in the sharpness of directional tuning, suggesting a continuum from more directionally modulated HD cells to those reflecting more “classic” or sharp HD cell tuning. Importantly, there was no apparent bimodality or clustering in the distribution.

Figure 2.

Histograms showing the cumulative proportion for basic firing characteristics of HD cells measured across anterodorsal thalamus (ADN), postsubiculum (PoS), parasubiculum (PaS), medial entorhinal cortex (MEC), and postrhinal cortex (POR). *p < 0.05; **p < 0.01; ***p < 0.001.

Second, we found that HD cell populations in the ADN were skewed toward sharper directional tuning (higher peak firing rates, greater stability and directional coherence) compared to all four parahippocampal regions. ADN HD cells displayed higher measures of mean vector length compared to the four parahippocampal regions, but this difference did not reach significance for comparisons between ADN and PaS (p = 0.18). We also confirmed previous reports of comparatively higher peak firing rates by ADN cells (ps < 0.001), which on average was approximately seven-fold higher than all four parahippocampal regions (Fig. 2). Background firing rates were higher for ADN cells compared to all regions (ps < 0.05). Measures of signal-to-noise ratios were significantly higher for ADN cells when compared to POR (ps < 0.001) – most likely attributed to higher peak firing rates for ADN HD cells. Thus, ADN cells fired more robustly and consistently in the cell’s preferred firing direction.

The data set also revealed that POR HD cells were more coarsely modulated by HD compared to other parahippocampal regions as indicated by significantly lower mean vector length measures (ps < 0.001) (Fig. 2). It is likely that the lower mean vector length measures in this population is related to their wider directional firing range. On average, POR cells expressed directional ranges that were on average >40 degrees wider than cells recorded in other parahippocampal regions (Fig. 2). Post-hoc comparisons between POR cells and the other parahippocampal regions and ADN were statistically significant (ps < 0.001). POR cells also expressed lower signal-to-noise ratio measures when compared to the other parahippocampal regions (ps < 0.01), and higher background rates compared to all regions except PoS (ps < 0.05), which may also be related to the reported lower mean vector length measures among the POR cell population. Finally, PoS cells displayed considerably greater directional stability and directional coherence compared to POR, PaS and MEC (ps < 0.01). Parahippocampal regions were largely similar across other measures (see Fig. 2).

Anticipatory time intervals for HD cells differ across thalamo-parahippocampal regions.

Next, we calculated the time by which HD cells anticipated future HDs, termed the anticipatory time interval (ATI) (Fig. 2, bottom right). As expected, measures of ATI differed significantly between the five regions (F(4, 196) = 43.3, p ≤ 0.001). Consistent with previous studies, ADN cells expressed positive ATI values that were on average significantly greater than all other regions (median: 56.4 msec; ps <0.001). This finding extends previous studies in suggesting that ADN cells anticipate future HDs to a greater extent than PaS, MEC, and POR. For HD cells recorded in the PoS, PaS, and MEC, median ATIs were generally closer to 0 with cells in PoS and MEC slightly lagging and PaS cells slightly anticipating (PoS: −7.42 msec; PaS: 2.88 msec; MEC: −8.47 msec). The ATIs of HD cells recorded in the POR were also closer to 0 but with a lag (median: −32.1 msec). Post-hoc comparisons detected a significant difference between POR and PaS ATIs (p <0.05), but no differences were detected between other parahippocampal regions.

Spike-time discharge characteristics vary across thalamo-parahippocampal regions

We next investigated the spike-time characteristics of HD cells across thalamo-parahippocampal regions. We first determined whether HD cells discharged in bursts of spikes and whether spike timing varied within ±30° of their preferred firing directions. With respect to the burst index, we found that median values for each region were closer to 0 (range: 0.05 – 0.14) indicating that, in general, cells in these regions were not firing in bursts, but rather when cells fired, their spiking was spread out over 100 msec periods. This observation is consistent with previous studies suggesting that burst-like firing is limited among HD cell populations (Latuske et al., 2015; Taube, 2010). While the median burst index measure was generally low across the five thalamo- parahippocampal regions, we found that ADN cells exhibited the highest median burst firing compared to other regions and POR cells displayed the lowest median burst firing (F(4, 258) = 29.5, p ≤ 0.001) (Fig. 3, top; see inset for results of post-hoc comparisons). It is important to point out the possibility that the greater burst index values for ADN HD cells may be related to their greater peak firing rates (cells with higher peak firing rates will have shorter/smaller ISIs than cells with lower firing rates). To control for this possibility, we compared burst index values across regions, but only for cells that had peak firing rates between 10–20 Hz (Fig. 3, middle). We selected this range because 1) there were few cells in the ADN group with peak firing rates < 10 Hz and 2) there were few cells in the parahippocampal groups with peak firing rates > 20 Hz. Using this range of peak firing rates, ADN cells expressed burst index values that were not statistically different than parahippocampal regions. Thus, while ADN cells generally expressed higher burst index values compared to other regions, this finding appears related to their higher peak firing rates.

Figure 3.

Histograms showing the cumulative proportion for burst index measures for all HD cells (top), the burst index from HD cells with peak firing rates between 10 and 20Hz (middle), and coefficient of variation for HD cells (bottom). Anterodorsal thalamus (ADN), postsubiculum (PoS), parasubiculum (PaS), medial entorhinal cortex (MEC), and postrhinal cortex (POR). *p < 0.05 ***p < 0.01 ***p < 0.001.

Next, we compared measures of CV, which assess the regularity in spiking in preferred firing directions of HD cells. We found that the CV varied significantly across regions (F(4, 249) = 39.6, p ≤ 0.001) (Fig. 3, bottom) with ADN, MEC, and PaS cells exhibiting higher values compared to PoS and POR. Post-hoc comparisons between ADN, PaS, and MEC cells did not detect significant differences. However, all other comparisons reached statistical significance (ps < 0.01).

Conjunctive modulation of HD cell firing in thalamo-parahippocampal regions

Consistent with previous research, we found that many thalamo-parahippocampal HD cells were conjunctively modulated by at least one of 8 variables: theta rhythmicity, location-specific firing (grid, border, and place), linear or angular head velocity, egocentric bearing, or center-distance (Fig. 4). However, the proportion of each conjunctive subtype varied considerably across the 5 regions (χ²(4) > 12.8, ps ≤ 0.05). Further, while pure (non-conjunctive) HD cells were observed in all regions, the proportion of these cells varied significantly across the 5 regions (χ²(4) = 25.1, p ≤ 0.001). The greatest number of pure HD cells was observed in the ADN (61% of cells passing HD cell criteria) and PoS (65%), while lower numbers were observed in PaS (46%), MEC (41%), and POR (37%) (see Fig. 4). Note that the proportional values shown for each brain region in Figure 4 do not sum to 100%; this outcome occurs because cells could be sensitive to more than one parameter simultaneously.

Figure 4.

Proportion of each conjunctive subtype found in each region. HD x T: HD x Theta; HD x G: HD x Grid; HD x B: HD x Border; HD x P: HD x Place; HD x LV: HD x Linear velocity; HD x AV: HD x angular head velocity; HD x EB: HD x Egocentric bearing; HD x ED: HD x Egocentric Distance.

With respect to theta rhythmicity and location-specific modulation (grid, border, and place), our findings were largely consistent with previous reports (Boccara et al., 2010; Cacucci et al., 2004; Taube, 1995; Taube et al., 1990a). MEC and PaS exhibited the highest proportions of HD cells that were conjunctively modulated by theta rhythmicity (Fig. 4). Smaller proportions were observed within PoS and POR and were not observed at all in our population of ADN HD cells. Interestingly, when plots of theta index vs. burst index are constructed for HD cells across the five regions (Fig. 5), there was a good correlation for cells in MEC (r = 0.703) and PaS (r = 0.633), but less so for the other three areas (r’s: PoS: 0.429, POR: 0.205, ADN: 0.416). This finding is consistent with the PCA and cluster analyses described below where, in general, HD cells in MEC and PaS appeared to encode slightly different aspects of directional firing than the other three areas, although cells from all these regions overlap on these dimensions.

Figure 5.

Scatter plots showing the relationship between burst and theta indices of HD cells for each region. Trendline (black) and Pearson’s correlations (r) are shown for each plot. Anterodorsal thalamus (ADN), postsubiculum (PoS), parasubiculum (PaS), medial entorhinal cortex (MEC), and postrhinal cortex (POR).

HD cells conjunctive with grid firing were exclusively recorded in the MEC and PaS, but border and place-modulated HD cells were identified in all four regions of the parahippocampal cortex (Fig. 4). While two HD cells (1.9%) in ADN were modulated by place, grid and border modulation was not observed in our population of ADN HD cells.

Although our findings are consistent with previous studies showing an absence of location-specific modulation in ADN cells (Peyrache et al., 2017; Taube, 1995), it is important to note that the ADN cells were recorded while animals locomoted in a small open field cylinder (71 cm in diameter), rather than a larger enclosure (120 × 120 cm), as was done for parahippocampal cells. Thus, it is possible that the small enclosure precluded detection of hexagonal location-specific modulation. To address this possibility, we inspected rate maps of a separate dataset (n = 27) of ADN HD cells that were recorded while animals locomoted a larger 1 m square box. Among this cell population, we failed to detect grid-like modulation (data not shown).

Consistent with a recent report (Spalla et al., 2022), we also found that a small proportion of HD cells in each of the parahippocampal areas was modulated by linear velocity (speed), although the proportions were smaller than those reported by Spalla et al. In contrast, an even larger proportion of ADN HD cells was modulated by linear velocity (Fig. 4). When firing rate vs. linear velocity tuning curves were constructed for samples that fell within ± 60° of the cell’s PFD, the percentage of cells that had linear velocity correlations ′≥ 0.7 (accounting for approximately half of the firing rate variance) and absolute values of slopes ′≥ 0.1 cm/s were PoS: 12.6%, PaS: 4.5%, MEC: 12.1%, POR: 10.0%, and ADN: 33.6%. Interestingly, the percentage of MEC cells that passed this criteria was relatively small, despite previous findings identifying a significant number of MEC cells responsive to speed, including many that were conjunctive with HD sensitivity (Kropff et al., 2015; Hinman et al., 2016). Using a GLM analysis did not change the percentages significantly: PoS: 7.2%, PaS: 9.0%, MEC: 17.1%, POR: 11.4%, ADN: 27.1%. Figure 6 shows representative examples of HD cells from each of the five brain areas that had good correlations to linear velocity. Note that some cells had negative correlations where firing rates decreased as linear velocity increased. An example cell from ADN is shown in the bottom row of Figure 6.

Figure 6.

Examples of HD cells tuned to linear velocity. Each row displays one cell. Left column: HD cell tuning curve with Rayleigh value. Middle column: linear velocity tuning curve with Pearson’s correlation (r) and slope. Right column: AHV tuning curve. Labels for each plot are as shown for the cell in the top row. The location of each cell is denoted on the left. Two cells are shown for the ADN (rows 5 and 6). One cell increased its firing with increasing linear velocity (row 5), while the other cell decreased its firing with increasing linear velocity (row 6). None of the cells shown here were classified as sensitive to AHV.

Further, a few HD cells (3.4%) in ADN were significantly modulated by angular head velocity (see also Bassett et al., 2007; Taube, 1995). In contrast, none of the PoS, PaS, POR, and MEC HD cells were significantly influenced by the animal’s AHV using the criteria we selected (r ′≥ 0.5, absolute value of slope ′≥ 0.1); Figure 7 shows some examples of this cell type. Using GLM analyses, the percentages for AHV sensitivity were a little higher for all parahippocampal areas (PoS: 4.5%, PaS: 6.0%, MEC: 2.9%, POR: 7.1%), as well as for ADN cells (9.4%). Finally, while some co-modulation with egocentric variables was apparent in each thalamo-parahippocampal region, the largest numbers appeared in POR (Fig. 4).

Figure 7.

Examples of HD cells tuned to AHV. Each row displays one cell. Left column: HD cell tuning curve with Rayleigh value. Middle column: AHV tuning curve with Pearson’s correlation (r) and slope values for CW and CCW portions. Right column: Linear velocity tuning curve with Pearson’s correlation (r) and slope values. Labels for plots are shown for the cell in the bottom row. The cells in the top two rows were from ADN; the cell in the bottom row was from POR. Note that the ADN cell in the second row has an inverted AHV function. The POR cell in the bottom row was sensitive to both AHV and linear velocity.

HD cells displayed direction-specific firing equally at all AHVs

In a preliminary study in the presubiculum of bats, Finkelstein et al. (2019) reported that direction selective activity only emerged at high AHVs in a significant number of cells and that there was little direction-specific firing at lower AHVs. We examined this issue for HD cells in our data set. Although cells in our data set were selected based on their overall HD tuning curve that was independent of AHV, any cell that was only directionally modulated at high AHVs would still be expected to display directional tuning in its overall tuning curve, although its peak firing rate would be reduced because its tuning curve would have included many samples from low AHVs, which presumably did not contain any directional sensitivity. For this analysis, we constructed a firing rate vs. HD tuning curve for each cell for three AHV ranges (0–30, 30–90, 90–300 °/s). We then computed the ratio of the peak firing rate between 1) the middle AHV range and the low AHV range, and 2) the high AHV range and the low AHV range. If cells were more directionally modulated at high AHV ranges, then their ratios should > 1. However, we found that across all five brain areas ratios were generally close to 1 for both range comparisons and were seldom > 2. Indeed, across all five brain areas, there were only 3 cells (1 in MEC and 2 in PaS) that had a peak firing rate ratio > 2 for the 90–300°/s : 0–30°/s category and only 3 cells (in MEC) that had a peak firing rate ratio > 2 for the 30–90°/s : 0–30°/s category. Further, for each of these cases the peak firing rate for the 0–30°/s range was very low (< 2 spikes/s), which would make it easier to obtain high ratios with relatively small increases in the cell’s firing rate at high AHVs. Table 1 depicts the mean peak firing rate ratios for each brain area.

TABLE 1.

Peak Firing Rate Ratios for different AHV ranges

Brain Area	n	Peak Firing Rate Ratio (30–90 °/s : 0–30 °/s)	Peak Firing Rate Ratio (90–300 °/s : 0–30 °/s)
ADN	107	0.983 ± 0.012	1.045 ± 0.016
PoS	111	0.968 ± 0.020	1.000 ± 0.025
PaS	110	1.002 ± 0.021	1.044 ± 0.027
MEC	140	1.019 ± 0.025	1.034 ± 0.027
POR	70	1.003 ± 0.026	0.976 ± 0.032

Figure 8A shows a ranked ordering of the peak firing rate ratios for the high AHV range to the low AHV range for each brain area. For each plot there is a smooth continuous distribution from low to high ratios. Were cells to display HD sensitivity at only high AHVs, as reported by Finkelstein et al. (2019), then peak firing rate ratios for some cells would be expected to be much higher than 2. Moreover, we found that all cells that displayed a high peak firing rate ratio also contained direction-specific firing at lower AHVs. Indeed, for our entire population of cells across all five brain areas, we never found a cell that was only directionally sensitive at high AHVs. Plots of the peak firing rate ratios for the medium AHV range to the low AHV range looked similar to the high AHV range ratio plots shown in Figure 8A (data not shown). Figure 8B depicts the tuning curves for the three different AHV ranges for six representative cells in PoS that displayed a range of different peak firing rate ratios. The cells selected are shown by the red dots in Figure 8A in the PoS plot. Figure 8C shows a similar set of 6 representative cells for the ADN; the six cells selected are shown by the red dots in Figure 8A in the ADN plot. In each case the tuning curve for the high AHV range looked relatively similar to the tuning curve for the low AHV range. Even when the tuning curve for the high AHV range appeared higher than for the low AHV range (numbers 5, 6, 11, 12 in Fig. 8B, C), the difference between the tuning curves was relatively minor and appeared to be more accounted by natural variability in firing rather than by a property of the cell. Similar results were also found for PaS, MEC, and POR (data not shown). In sum, in contrast to the Finkelstein et al. (2019) study, there was little, if any, modulation of firing rate based on AHV across all five brain areas.

Figure 8.

HD cell tuning curves based on different AHV ranges. A) Ranked ordered distribution of peak firing rate ratios based on the high AHV range (90 – 300°/s) to low AHV range (0 – 30°/s) for all HD cells in each of the five brain areas. Ratios generally ranged between 0.5 to 2.0 with smooth continuous distributions for each brain area. The dashed line in each plot represents a ratio of 1.0. B) HD tuning curves for three different AHV ranges (see legend) for the 6 cells shown in the PoS ranked-ordered plot in A by the red dots (numbered 1–6). C) Same as B except for ADN and cells numbered 7–12 in the ranked ordered plot for ADN in A. Note that even in cells that had high ratios (e.g., cells 5, 6, 11, 12), there was still significant direction specific firing at the low AHV range. Axis labels for each plot are shown in the top row for each column.

Robust HD cell firing varies across pure and conjunctive HD cells in MEC, PaS, and PoS

We next investigated the possibility that weaker directional tuning observed in thalamo-parahippocampal regions may be related to the high proportion of co-modulation by theta, location-specificity, velocity, and/or egocentric firing characteristics within parahippocampal regions. For instance, cells co-modulated by these variables might be more likely to spike outside their preferred firing direction resulting in diminished or broadening of directional tuning. To test this possibility, we first directly compared directional tuning measures between pure vs. conjunctive HD cells (theta, velocity, and location-specific/egocentrically) in each brain region (Fig. 9; see inset for results of post-hoc comparisons). Because of the small numbers of theta, velocity, location-specific, and egocentric subtypes in some of the regions, conjunctive cells were pooled across subtypes for statistical comparisons. Our analysis revealed that conjunctive HD cells recorded in PaS and MEC showed lower peak rates, lower signal-to-noise ratios, lower directional coherence, lower mean vector length and directional information content measures, exhibited less stability, and had wider tuning curves (ps < 0.05). Conjunctive cells in PoS showed a slightly similar pattern for all measures of directional tuning but only directionbal tuning was significantly wider (p < 0.05). POR conjunctive cells were similar to pure HD cells across all other measures. ADN conjunctive and pure cells were similar across all measures of directional tuning with the exception of peak firing rates which were higher in ADN conjunctive cells (p < 0.01). Thus, weaker directional modulation is likely a general feature of conjunctive HD cells in parahippocampal cortex, especially within PaS and MEC regions.

Figure 9.

Scatter plots showing raw data and the median (black horizontal line) for basic firing characteristics of pure (red circles) and conjunctive (blue circles) HD cells in each region. *p < 0.05; **p < 0.01; ***p < 0.001.

We also investigated whether measures of ATI and spike time discharge characteristics were affected by conjunctive cell firing (Fig. 9). In POR, conjunctive HD cells were observed to slightly lag in their firing, but this difference failed to reach statistical significance (p = 0.11). All other regions did not show a difference in ATI across pure and conjunctive cells. Median values for CV and burst index were generally higher for conjunctive cells in MEC, PaS, PoS, and ADN (ps < 0.01; not significant for ADN), and in particular when compared to POR.

PCA and Cluster Analyses

We next performed two types of analyses to determine whether certain parameters clustered together and whether any clusters were associated with particular brain regions. The first analysis involved a principal component analysis (PCA) and was based on 12 parameters: peak firing rate; directional firing range; background firing rate, signal-to-noise ratio, directional coherence, directional informational content, Rayleigh r, linear velocity correlation, CCW AHV correlation, CW AHV correlation, and theta index. We first determined how many dimensions should be selected to optimally account for the variance in the data. To reduce the number of dimensions from 12, we first ran the PCA on all 12 parameters. 70% of the variance could be explained by three components, with additional components accounting for < 10% of the explained variance (Fig. 10A). We therefore used three components in subsequent analyses. The first component (PC-1) accounts for a large portion of the variance (~45%). The factors with the largest weights were (in descending order) cells that contained high Rayleigh r values, high directional information content scores, and low directional firing ranges (Fig. 10B, left). In general, these parameters correspond to HD cells that can be characterized as ‘classic’ HD cells, with sharp tuning curves and little out-of-range firing. The second component (PC-2) accounted for 16% of the explained variance and corresponded to cells with high peak firing rates, high linear velocity correlations, and relatively low directional information content scores (Fig. 10B, middle). The third component (PC-3) accounted for 9% of the explained variance and corresponded mostly to cells with a large directional firing range, a high theta and burst indices, and a moderately high Rayleigh value (Fig. 10B, right).

Figure 10. PCA and Cluster Analyses.

A) PCA using 12 different parameters and displaying the three best components (PC-1, PC-2, PC-3). Each plot displays how much each parameter contributes to that component. The plot on the left shows the amount of variance accounted for by the first 5 components. B) Cluster analyses using three different PCA components. C) Same as B, but analyzed by brain area. D) Analyses using k-means clustering to determine the optimal number of distinct clusters the data could be divided into; top: k-means silhouette analysis, bottom: k-means elbow analysis. E) Three-dimensional scatter plots color-coded by brain area using left: Rayleigh r, peak firing rate, and theta index; middle: peak firing rate, theta index, and burst index; right: Rayleigh r, peak firing rate, and absolute value of linear velocity correlation. See text for more details.

Figure 10C (top row) plots PC-1 vs. PC-2, PC-1 vs PC-3, and PC-2 vs. PC-3, and while the distribution of points in all three plots is continuous, there is a clear separation of the two components around the 0 value for PC-1 (left and middle plots). The plot displaying PC-2 vs. PC-3 did not yield a clear separation between the two components.

The bottom row of Figure 10C shows the same three plots as in the top row, but points are now color-coded based on their brain location. The clearest differentiation between brain regions appears in the PC-1 vs. PC-2 plot, although aspects of the differentiation can be seen in the other two plots. Cells from MEC, PaS, and PoS (blue, red, and green, respectively) are intermixed, but generally separated from ADN cells (orange), which had high PC-2 values. Cells from POR (black) generally contain low values for both PC-1 and PC-2, although they are intermixed with cells from the other parahippocampal brain regions, but not from ADN. Finally, there is a group of cells that appears mostly confined to MEC and PaS and are not present in the other three brain regions (although there are a few cells of this type in PoS); this result is most evident in the two plots on the right and can be best-characterized as cells that contain high PC-3 values. Thus, four characteristics emerge from the PCA: 1) HD cells from the ADN represent one category of HD cells, 2) HD cells from POR represent a second distinct category of HD cells, 3) HD cells from the other three parahippocampal regions (MEC, PaS, PoS) contain a sub-population of HD cells similar to POR, but also include a third category of HD cells that are generally distinct from and not observed in either ADN or POR, and 4) a fourth HD cell category appears mostly confined to MEC and PaS.

The second analysis used k-means clustering to determine the optimal number of distinct clusters the data could be divided into. We used both a k-means silhouette analysis and a k-means elbow analysis to determine the optimal number of clusters in this data set. Figure 10D shows that both methods yielded similar results with an optimal value of 2 clusters – the elbow occurs at 2 clusters for the elbow method and the maximum silhouette score occurs at 2 clusters for the silhouette method.

Figure 10E displays 3D scattergrams with different combinations of parameters and color-coded based on the cell’s brain region. The parameters selected for display were based on those parameters that captured a large portion of how each principal component was defined in the PCA. The first plot (left scattergram) contains the parameters: peak firing rate, theta index, and the Rayleigh mean vector length. The second plot (middle scattergram) displays the parameters: theta index, burst index, and peak firing rate. The third plot (right scattergram) depicts the peak firing rate, linear velocity correlation, and Rayleigh mean vector length. In general, there was considerable overlap of cells from all five brain regions across each of these parameters, but there were some distinct clusters. One cluster that was evident was cells with high peak firing rates (left and right plots), which only occurred in ADN cells, although ADN cells could also have low peak firing rates. A second cluster that emerges are for cells with high theta and burst index scores (middle plot), which was localized to the PaS and MEC. This result is consistent with the results reported above for theta rhythmicity (see Fig. 5) and with previous studies applying PCA to MEC and PaS, which identified two types of HD cells in these regions (Latuske et al., 2015; Ebbesen et al., 2016). It is also important to note, that although cells could be classified along two primary components (see PCA above), the scattergrams show that the various parameters are best characterized as continuous distributions and there is no evidence for a bimodal distribution. This observation is also true for the parameters not displayed in these scattergram plots (data not shown).

Overall, the results of the PCA and cluster analyses largely recapitulate the conclusions from our individual measurement analyses, such that, while HD cells from individual brain regions can occupy distinct regions in cluster space (particularly ADN in terms of peak firing rate and MEC and PaS in terms of theta/burst indices), the overarching distribution of HD cells varies continuously along even the most informative dimensions. However, it is worth noting that, unlike the individual measurement analyses, k-means clustering indicates a separation of the HD cell population into two groups along PC-1, corresponding roughly to strongly-tuned or ‘classic’ HD cells with positive PC-1 values, and broadly-tuned or ‘HD-modulated’ cells with negative PC-1 values.

Robust HD cell firing varies across superficial and deep layers of MEC, PaS, and PoS

Previous studies have identified distinct populations of HD cells in the superficial layers (II/III) of MEC that exhibit weak directional tuning (i.e., lower measures of mean vector length) (Giocomo et al., 2014; Zutshi et al., 2018). Consistent with these observations, we found that HD cells within the superficial layer were less directionally specific than deep layer cells (Fig. 11). Specifically, within our population of MEC HD cells, superficial layer cells exhibited significantly lower mean vector lengths and directional information content, less directional stability, and higher background rates when compared to deep layer (V/VI) cells (ps < 0.05) (Fig. 12A). A similar pattern was observed in PaS with cells in the superficial layers showing significantly weaker and less specific directional tuning across most measures (ps < 0.001). For PoS, we found that superficial layer cells were less stable, exhibited less directional coherence, lower signal-to-noise ratio, lower mean vector lengths and directional information content, and had wider tuning curves (p < 0.05) compared to deeper layers. For POR cells, measures of directional tuning did not significantly differ across layers.

Figure 11.

Polar HD cell tuning curves (firing rate x HD) are shown for each laminar region (Sup = layers II/III; Deep = layers IV/V). The first 5 cells that exceeded the median mean vector length for each laminar region are shown. Above each tuning curve is the peak firing rate (top) and mean vector length (bottom) for that cell.

Figure 12.

A. Scatter plots showing raw data and the median (red horizontal line) for basic firing characteristics of superficial (gray circles) and deep layer (black circles) parahippocampal HD cells. B. Bar graph showing the proportion of pure HD cells (black bars) per superficial and deep layers of each parahippocampal region. *p < 0.05; **p < 0.01; ***p < 0.001.

We further investigated whether measures of ATI and spike time discharge characteristics varied across superficial and deep layers (Fig. 12A). In general, there were no significant differences in measures of ATI across laminar location. Measures of burst index were generally higher in the superficial layers for parahippocampal regions but reached significance only in PaS (p < 0.001). Measures of CV were similarly higher in the superficial layers, reaching significance for PoS, PaS, and POR (ps < 0.05).

Figure 12B displays a bar graph summarizing the proportion of pure vs. conjunctive cells across superficial and deep layers in each of the parahippocampal areas. The plot shows that there was a trend toward a slightly higher proportion of pure HD cells in the deep layers than in the superficial layers for MEC, PaS, and PoS; there was no difference for POR, which is consistent with the weaker directional tuning in superficial layers compared to deep layers across all parahippocampal areas.

Giocomo et al. (2014) also reported that layer III MEC cells exhibit a topographical organization, with HD cells in the dorsal regions expressing greater mean vector length values compared to those recorded in ventral regions. Among our HD cell population that could be clearly distinguished as layer III cells (n = 40), we did not observe a similar topographical organization in directional tuning (Fig. 13). Specifically, a linear correlation between the dorsal-ventral recording location (distance (μm) from MEC border) and mean vector length failed to reach significance (Pearson’s r = −0.03, p = 0.84). A similar observation was made when correlating recording location (distance (μm) from MEC border) with directional firing range (Pearson’s r = 0.02, p = 0.91). Although these results fail to replicate the findings by Giocomo et al. (2014), it is important to note that the dorsal-ventral range of recorded layer III cells is smaller in the present study (0–952 μm vs. 0–2000 μm in Giocomo et al., 2014), which may preclude sufficient sampling of broadly tuned cells in ventral MEC.

Figure 13.

Scatter plots showing the mean vector length (left) and directional firing range (right) of each MEC layer III HD cells plotted as a function of recording location relative to the dorsal border of MEC. Trendline (black) is shown for each plot.

HD cell properties in other brain areas

Given the similarities and differences reported here across different HD cell populations in the parahippocampal regions and ADN, we next compared our results to HD cell properties from four other brain areas: retrosplenial cortex (RSC; M.L. Mehlman, unpublished data), dorsal striatum (DS; Mehlman et al., 2019), medial precentral cortex (PrCM; Mehlman et al., 2019), and the lateral mammillary nuclei (LMN; Stackman & Taube, 1998; Yoder et al., 2015). In each case, however, we note that there were some differences in how the HD cells in these other brain areas were recorded (i.e., cell isolation using single wires and window-discriminators vs. tetrode recording and cluster-cutting analyses), as well as the type of apparatus used (cylinder vs. square-shaped environments). In addition, the number of cells available for analysis was smaller for each of these other brain areas. Despite these differences, we thought it was worthwhile to conduct an initial comparison. HD cell properties for each brain area are shown in Table 2. Results showed that HD cells from the LMN had similar characteristics to ADN HD cells in terms of high peak firing rates, larger ATIs, and a propensity to be correlated to linear velocity. Similar to POR, LMN cells had higher directional firing ranges and lower information content values than all the other brain areas. While coefficient of variation values were comparable across these other brain areas, burst index scores were also higher for LMN, RSC, striatum, and PrCM than for the ADN and parahippocampal areas, although it was comparable to the ADN. Much of this difference is likely attributed to the higher mean peak firing rates for these cells, since previous studies have shown that cells with higher peak firing rates generally have higher burst index scores (Coletta et al., 2018). In terms of linear velocity, there was a much higher percentage of cells from the LMN, striatum, and PrCM that were sensitive to AHV than the other brain areas. In addition, cells from each of these other brain areas, as well as ADN, were also generally more sensitive to linear velocity than cells from the parahippocampal areas. Importantly, none of the RSC cells were sensitive to AHV, but about half of them (46.7%) were sensitive to linear velocity. With regards to AHV, there was a much higher percentage of cells from the LMN, striatum, and PrCM that were sensitive to AHV than the other brain areas. Finally, it is noteworthy that we did not find any AHV sensitivity in RSC HD cells, which contrasts with a recent study by Keshavarzi et al. (2022) who reported a substantial number of conjunctive HD x AHV cells (~60%) in RSC. Possibilities contributing to this difference could lie in the different species used (mice vs. rats), different areas of recording within RSC, or different criteria used for classifying cells as directionally tuned.

TABLE 2.

HD Cell Properties Compared Across Brain Regions

Parameter	ADN	PoS	PaS	MEC	POR	LMN	RSC	PrCM	Striatum
Number of cells (n)	107	111	133	140	70	33	15	16	24
Rayleigh r	0.710 ± 0.018	0.628 ± 0.020	0.655 ± 0.018	0.623 ±0.017	0.464 ± 0.014	0.608 ± 0.027	0.606 ± 0.065	0.710 ± 0.036	0.710 ± 0.034
Peak Firing Rate (spikes/s)	36.95 ± 1.72	8.54 ± 1.07	5.14 ± 0.39	5.96 ± 0.36	6.91 ± 0.64	52.91 ± 9.64	34.31 ± 5.60	65.18 ± 7.14	36.88 ± 4.42
Directional Firing Range (°)	117.8 ± 3.0	104.8 ± 4.3	120.2 ± 4.6	117.8 ± 5.1	165.7 ± 6.7	170.2 ± 6.5	108.6 ± 6.2	108.6 ± 6.38	100.4 ± 3.7
Background FR (spikes/s)	1.89 ± 0.22	0.70 ± 0.09	0.30 ± 0.04	0.45 ± 0.06	1.11 ± 0.14	3.38 ± 0.82	4.06 ± 1.61	3.31 ± 0.85	1.75 ± 0.33
Signal-to-Noise Ratio	27.1	19.0	23.7	16.1	6.9	15.2	13.8	36.0	40.9
Directional Coherence	0.971 ± 0.003	0.938 ± 0.005	0.909 ± 0.006	0.910 ± 0.006	0.895 ± 0.009	0.962 ± 0.006	0.957 ± 0.017	0.958 ± 0.008	0.938 ± 0.015
Information Content (bits)	1.197 ± 0.066	1.101 ± 0.074	1.174 ± 0.065	1.121 ± 0.065	0.477 ± 0.033	0.746 ± 0.087	0.984 ± 0.211	0.939 ± 0.128	1.168 ± 0.127
Directional Stability Score	0.850 ± 0.012	0.767 ± 0.017	0.663 ± 0.020	0.658 ± 0.017	0.647 ± 0.022	0.831 ± 0.020	0.833 ± 0.043	0.662 ± 0.047	0.652 ± 0.039
Coefficient of Variation	0.714 ± 0.008	0.619 ± 0.011	0.697 ± 0.011	0.714 ± 0.013	0.549 ± 0.013	0.825 ± 0.045	0.751 ± 0.037	0.767 ± 0.025	0.720 ± 0.019
Burst Index	0.167 ± 0.010	0.097 ± 0.007	0.118 ± 0.007	0.145 ± 0.008	0.062 ± 0.006	0.289 ± 0.048	0.204 ± 0.036	0.251 ± 0.035	0.171 ± 0.020
ATI (msec)	56.3 ± 3.7	−6.63 ± 4.81	6.01 ± 5.61	−20.3 ± 8.8	−40.5 ± 13.8	81.9 ± 9.5	48.43 ± 12.90	50.54 ± 9.94	−11.40 ± 22.84
Theta Index	3.36 ± 0.07	2.76 ± 0.14	3.72 ± 0.21	3.75 ± 0.20	3.19 ± 0.16	3.70 ± 0.13	3.56 ± 0.16	3.72 ± 0.12	3.47 ± 0.253
AHV (percent classified)	3.7	0.0	0.0	0.0	0.0	28.6	0.0	31.3	33.3
LV (percent classified)	33.6	12.6	4.5	12.1	11.4	34.3	46.7	18.8	29.2

All values are mean ± SEM, except for Signal-to-Noise ratio, where the median is reported, and for AHV and LV measures, where the percentage of cells classified as sensitive to AHV or LV are reported. See text for significant differences across parameters x brain area.

Discussion

The present study reports several major differences in firing characteristics across thalamo-parahippocampal HD cells. First, we found that ADN HD cells had higher mean vector lengths, greater directional stability, and peak firing rates that were roughly 7x higher than other regions. This observation is consistent with other studies reporting higher firing rates by ADN cells when compared to PoS (Peyrache et al., 2015; Sharp, 1996; Taube, 1995; Taube & Muller, 1998). The variability of peak firing rates across different HD cells, varying from as little as a few spikes/s to as high as 150 spikes/s (rates averaged over 8–16 min sessions), has been evident since the time of their initial discovery (Taube et al., 1990a). However, the significance of this aspect in terms of identifying a functional role for this variability and understanding its impact on downstream neurons remains elusive. In addition, ADN HD cells discharged in anticipation of future HDs to a greater extent than the four parahippocampal regions and had a larger number of cells modulated by linear and angular head velocity. Second, we found that while parahippocampal HD cells were similar across most directional firing characteristics, POR cells were more coarsely modulated by HD, where they exhibited substantially lower mean vector length measures and wider tuning curves. In addition, POR cells were more likely to be co-modulated by egocentric variables. Third, the discharge characteristics of HD cells recorded in PoS, PaS, and MEC could be differentiated across laminar layers. Compared to superficial layers, sharper directional tuning (e.g., mean vector length, directional information content, stability, directional range) was observed in deep MEC, PaS, and PoS layers and among pure HD cell populations.

Both recording and lesion studies have indicated that the HD signal is generated and maintained through an ascending vestibular-motor circuit and is modulated by descending parahippocampal input from PoS and projections from the retrosplenial cortex (Fig. 1; also see Clark & Taube, 2012; Mehlman et al., 2020; Taube, 2007). Evidence for this proposal stems from studies reporting that damage to subcortical regions that contain HD cells leads to a complete loss in “downstream” HD cell signaling in ADN and parahippocampal regions (Bassett et al., 2007; Blair et al., 1998; Butler & Taube, 2015; Clark et al., 2012; Sharp et al., 2008), and damage to the ADN leads to a loss of HD tuning in PoS and PaS/MEC neurons (Goodridge & Taube, 1997; Winter et al., 2015). These previous findings, coupled with the results of this study, suggests that the ADN relays a robust HD signal that drives directional tuning in parahippocampal targets (Petrof & Sherman, 2009; Sherman & Guillery, 2006). Descending projections stemming from the retrosplenial cortex and PoS are more modulatory in that they are important for stabilizing the ADN HD cell signal with respect to environmental cues (Clark et al., 2010; Goodridge & Taube, 1997). These latter projections are thought to gain their influence over ADN HD signaling through PoS projections to the lateral mammillary nuclei (Yoder et al., 2015).

Dudchenko et al. (2019) have recently argued that different regions of the HD cell circuit may contain more than one directional representation, possibly driven by distinct sensory-motor information, where one population of cells is driven more by landmark cues, while other populations are driven more by self-motion cues such as vestibular information. Our findings here, along with previous work (e.g., Giocomo et al., 2014; Jacob et al., 2017), support this basic hypothesis. For instance, HD signal generation in thalamo-cortical regions is thought to depend in part on angular head velocity signaling within the dorsal tegmental and lateral mammillary nuclei (Blair et al., 1998; Bassett et al., 2001; Stackman & Taube, 1997). ADN HD cells reflect these early processing stages with greater angular head velocity modulation and greater anticipation of future HDs compared to parahippocampal HD cell populations. It is likely that the unique ADN firing characteristics are driven by the large excitatory inputs stemming from the lateral mammillary nuclei (Fig. 1; also see Petrof & Sherman, 2009; Sherman & Guillery, 2006). Previous studies have shown that the lateral mammillary nuclei contain populations of high firing rate HD cells that anticipate future HDs and can be modulated by angular head velocity (Blair et al., 1998; Stackman & Taube, 1997). In addition, lesions to the lateral mammillary nuclei results in a loss of HD specific tuning in the ADN (Bassett et al., 2007; Blair et al., 1998; 1999). Previous studies have indicated that ADN HD cells have slightly narrower directional ranges when compared to HD cells recorded in the lateral mammillary nuclei (Stackman & Taube, 1998). Feedforward inhibition via a retrosplenial/PoS → reticular thalamus → ADN circuit may play an important role in narrowing the ADN HD signal as disruption of this pathway results in significantly wider directional tuning in ADN (Vantomme et al., 2020; see also Peyrache et al., 2019).

The significantly broader tuning observed in POR cells also remains unexplained but supports the hypothesis of a functional distinction within parahippocampal circuitry. The POR is reciprocally connected with the MEC (Burwell & Amaral, 1998), but recent studies have shown that neurons in this region are distinct from PoS, PaS, and MEC in that they respond to egocentric variables to a greater extent than location-specific firing (Gofman et al., 2019; LaChance et al., 2019). Our results are supportive of this previous finding in terms of HD cells, as a larger proportion of POR HD cells were coactive with egocentric variables and a smaller proportion were classified as pure HD cells. POR neurons have also been linked to modulation by a wide range of visual-spatial cues suggesting a broader role in processing spatial context (Burke et al., 2018; Furtak et al., 2012; Knierim et al., 2013). Whether the broad directional tuning by POR cells observed here is related to its role in contextual modulation is unclear as sensory stimuli were not manipulated in our study. However, a recent study by LaChance et al. (2021) showed that the firing of POR HD cells can be modulated by changes in the stability and number of visual landmark cues in the environment. Finally, although the ADN appears to endow directional signaling to PoS, PaS, and MEC either directly or indirectly via the retrosplenial cortex (Goodridge & Taube, 1997; Winter et al., 2015), it is unknown whether the ADN exerts a similar influence on POR cells.

An additional conclusion is that the sharpness of HD cell tuning varies based on the laminar layer within the parahippocampal area. Specifically, in MEC, PaS, and PoS, we found that cells in layers II/III were more coarsely tuned to the animals HD than HD cells localized to deeper layers. The finding that the sharpness of HD signaling decreases with a topographical gradient from layers V/VI → II/III is consistent with previously reported observations. First, Boccara et al. (2010) reported that only a few cells in layer II of MEC could be categorized as HD cells (see also Sargolini et al., 2006), with most cells expressing low mean vector length values. Consistent with these observations, in a recent paper Wang et al. (2023) reported that cells tuned to place in the lateral entorhinal cortex displayed slightly greater allocentric spatial selectivity (increased information content and stability measure) in the deep layers compared to the superficial layers, although there was no difference found for egocentric coding properties. Overall, these findings suggest that increased allocentric spatial selectivity may be a general feature of deeper layer cells across parahippocampal areas. Second, Giocomo et al. (2014) reported that HD cells in layer III of the MEC were topographically organized such that HD cells in the dorsal regions of layer III expressed sharp tuning while those recorded more ventrally expressed broad HD tuning. Although this observation parallels previous demonstrations of dorsal-ventral expansion of MEC grid cell firing fields (Brun et al., 2008), a similar expansion in HD cell tuning has not been observed in layers V/VI of MEC or along the dorsal-ventral axis of presubiculum (Giocomo et al., 2014). Among our layer III MEC cell population, we did not observe a similar topographical organization in directional tuning. However, it is important to note that the dorsal-ventral range of recorded layer III cells was smaller in our study (0–952 μm vs. 0–2000 μm in Giocomo et al., 2014) which may have precluded sufficient sampling from broadly tuned cells in more ventral regions of the MEC layer III. Zutshi et al. (2018) suggested that HD cells in layers II/III of the MEC were comprised of two distinct populations: one that was broadly tuned to HD, and a second that was sharply tuned to HD. Finally, Preston-Ferrer et al. (2016) reported that layer II cells of the PoS were weakly modulated by the animal’s HD.

Differences in directional tuning across parahippocampal lamina may reflect the organization of intrinsic parahippocampal circuitry (Figs. 11, 12). Using optogenetic-mediated inactivation, Zutshi et al. (2018) recently provided evidence that the broadly tuned HD cells within layer II/III of the MEC were largely dependent on local recurrent collaterals. In contrast, sharply tuned HD cells were unaffected by inactivation of terminals from recurrent layer II connections. Thus, coarse coding by HD cells in superficial MEC might be partially generated within local circuitry. However, it is important to note that HD cells in the superficial MEC might also be influenced by coarse coding HD cells located “upstream” to the MEC – possibly in the PoS, which sends extensive projections stemming from layer III (Simonnet & Fricker, 2018; van Groen & Wyss, 1990) or in PaS, which also forms a major input into the MEC (Caballero-Bleda & Witter, 1994; Tang et al., 2016). POR also sends extensive projections to MEC (Burwell & Amaral, 1998; cf., Doan et al. 2019). While projections from PoS terminate in layers II/III and V, projections from PaS are largely confined to layer II, perhaps exerting a greater influence on coarse coding in the MEC. In addition, the ADN and retrosplenial cortex send excitatory input to projection cells in layer III of PoS (Kononenko & Witter, 2012; Nassar et al., 2018), which may have an indirect role in modulating superficial MEC cells. While cells in the superficial layers of MEC/PaS/PoS may gain their course directional modulation by local circuitry, it is likely that the relatively sharper HD signaling within the deep layers may be driven in part by “upstream” ADN input (Fig. 1A). ADN projections also target PoS/PaS deep layers (Simonnet & Fricker, 2018; van Groen & Wyss, 1995), where HD firing was generally sharper compared to those recorded in the superficial layers.

A final conclusion from our study is that conjunctive HD cells express less robust HD cell tuning particularly within PoS, PaS, and MEC where location-specific modulation was most prominent. While less robust directional tuning in conjunctive cells may not be surprising, our findings show that directional information is more integrated with other spatial parameters (place, grid, and border) once directional information reaches parahippocampal regions. However, there was no evidence for a bimodal or clustered distribution within each brain region. Our analyses indicate that HD cells in parahippocampal circuitry are likely made up of a continuum, or gradient, ranging from sharp to more coarsely modulated HD cells.

Other studies have pointed to the possibility that parahippocampal HD cell populations are made up of multiple sub-classes of HD cells with distinct firing characteristics. For instance, recording from mice, Kornienko et al. (2018) reported two HD cell sub-types in MEC – one that was modulated by theta and one that was theta independent. Kornienko et al. (2018) suggested that theta modulation of HD cells increased with recording depth in parahippocampal cortex, likely reflecting the transition from deep layer recordings to the superficial layers. Kornienko et al. (2018) also observed that simultaneously recorded theta-modulated HD cells were more likely to display coherency in their directional firing, i.e., cells’ PFDs maintained the same angular relationship with one another (remained in-register) during environmental cue manipulations. In contrast, non-theta modulated HD cells did not maintain remain in register during cue manipulations, also suggesting that distinct HD cell populations might be organized with respect to theta rhythmicity. We note that these findings were in mice and whether HD cells remain in register, could differ between rats and mice, since spatial properties in place cells (mice: Kentros et al., 2004 vs. rats: Rotenberg et al., 2000) and grid cells (mice: Chen et al. 2016 and Perez-Escobar et al. 2016 vs. rats: Hafting et al., 2005) sometimes differ between these two species. Nonetheless, provided that the deep layers of MEC are the recipient of projections from cortical regions involved in visual-spatial processing, such as the retrosplenial cortex (Witter et al., 2017), it is possible that sharply tuned HD cells are more sensitive to changes in external sensory stimuli.

A final issue is how conjunctive HD cell firing might be generated within parahippocampal circuits. Again, some recent studies point to the possibility that both local circuits and upstream modulation is necessary. For instance, Miao et al. (2017) found that chemogenetic silencing of inhibitory interneurons selectively disrupted grid cell activity in the superficial layers of MEC, while silencing the same neurons failed to disrupt HD cell signaling within the same region. A large proportion of stellate cells in MEC are likely grid cells (Sun et al., 2015), suggesting that GABAergic neurons, which mediate signaling between stellate cells within the MEC superficial layers, play a critical role in grid cell generation. This observation, coupled with previous findings that lesions/inactivation of the ADN abolished conjunctive grid x HD cell coding (Winter et al., 2015a), strongly suggests that conjunctive cells are dependent on both local circuitry and external inputs. A similar circuitry may underlie conjunctive theta modulation by HD coding, as theta modulation is critically dependent on local inhibitory interneuron firing (Stark et al., 2013). Alternatively, it is also possible that theta modulation and HD are encoded afferent to parahippocampal cortices in the anteroventral thalamus, which contains theta-modulated HD cells (Tsanov et al., 2011; Welday et al., 2011), which contrasts with HD cells in the ADN, which are not thought to be extensively modulated by theta (Taube, 2010).

Previous studies have identified and characterized HD cells in the retrosplenial cortex (Alexander et al., 2020; Chen et al., 1994; Cho & Sharp, 2001; Jacob et al., 2017), LMN (Blair et al., 1998; Stackman & Taube, 1998), dorsal striatum (Wiener, 1993; Mehlman et al. 2019), and medial precentral cortex (Mehlman et al. 2019; Mizumori et al., 2000). While we were able to compare the current data sets with that of smaller populations of HD cells we have recorded in these regions, it remains to be determined the extent to which HD cells in other brain areas that were not included in the present analyses, such as HD cells in anterior ventral thalamus (Tsanov et al., 2011) nucleus reuniens (Jankowski et al., 2014), and parietal cortex (Wilber et al., 2014), share similar firing characteristics with the brain regions discussed in the present manuscript.

In summary, our findings suggest that the sharpness of directional firing by HD cells is topographically organized such that robust HD coding is observed in ADN and deep layers of MEC/PaS/PoS while more coarse HD coding is found in POR. Our findings also show that the strength of HD tuning is degraded when cell tuning is conjunctive with other variables within PoS, PaS, and MEC. The characterization of the HD cell signal across architecturally diverse parahippocampal and thalamic circuits will further constrain our understanding of the basic mechanisms underlying HD signal generation, as well as the functional relationship between HD signaling and parahippocampal spatial and oscillatory activity. It remains to be determined how HD cells in other brain areas not included in the current analyses compare to the pattern of results we found here across these parahippocampal areas, and how distinct firing characteristics across thalamo-parahippocampal circuits map onto functional roles in spatial behavior (Clark & Harvey, 2016; Dudchenko et al., 2019; Harvey et al., 2017; Peckford et al., 2014; Yoder & Taube, 2014).

Data Availability:

The data that support the findings of this study are available on request from the corresponding authors.