Superficial-layer vs. deep-layer lateral entorhinal cortex: coding of allocentric space, egocentric space, speed, boundaries, and corners

Abstract

Entorhinal cortex is the major gateway between the neocortex and the hippocampus and thus plays an essential role in subserving episodic memory and spatial navigation. It can be divided into the medial entorhinal cortex (MEC) and the lateral entorhinal cortex (LEC), which are commonly theorized to be critical for spatial (context) and non-spatial (content) inputs, respectively. Consistent with this theory, LEC neurons are found to carry little information about allocentric self-location, even in cue-rich environments, but they exhibit egocentric spatial information about external items in the environment. The superficial and deep layers of LEC are believed to mediate the input to and output from the hippocampus, respectively. As earlier studies mainly examined the spatial firing properties of superficial-layer LEC neurons, here we characterized the deep-layer LEC neurons and made direct comparisons with their superficial counterparts in single unit recordings from behaving rats. Because deep-layer LEC cells received inputs from hippocampal regions, which have strong selectivity for self-location, we hypothesized that deep-layer LEC neurons would be more informative about allocentric position than superficial-layer LEC neurons. We found that deep-layer LEC cells showed only slightly more allocentric spatial information and higher spatial consistency than superficial-layer LEC cells. Egocentric coding properties were comparable between these two subregions. In addition, LEC neurons demonstrated preferential firing at lower speed, as well as at the boundary or corners of the environment. These results suggest that allocentric spatial outputs from the hippocampus are transformed in deep-layer LEC into the egocentric coding dimensions of LEC, rather than maintaining the allocentric spatial tuning of the CA1 place fields.

1. Introduction

Although hippocampus is essential for episodic memory (Vargha-Khadem et al., 1997; Squire et al., 2004) and significant progress has been made in terms of its structure and function (Andersen et al., 2006), lack of understanding of its neighboring structures has hindered our understanding of the operations within the hippocampus. Entorhinal cortex is situated at a pivotal position within the medial temporal lobe as it mediates most of the interaction between the hippocampus and the neocortex (Insausti et al., 1997; Burwell and Amaral, 1998; Witter and Amaral, 2004). Although rich computational properties related to spatial and nonspatial information processing have been identified in the entorhinal cortex, they have been investigated much less thoroughly than hippocampal processes (Witter and Moser, 2006; Knierim et al., 2014; Moser et al., 2014).

Entorhinal cortex can be divided into the lateral entorhinal cortex (LEC) and the medial entorhinal cortex (MEC). The discoveries of various functional cell types within the MEC have greatly facilitated our understanding of how spatial information propagates within the medial temporal lobe. Grid cells, for example, display multiple firing fields which form hexagonal patterns (Hafting et al., 2005). The relationship of grid patterns between pairs of grid cells remain consistent across different environments, which suggests a neural substrate for path integration (McNaughton et al., 2006; Bush et al., 2015). Boundary/border cells (Savelli et al., 2008; Solstad et al., 2008; Lever et al., 2009), on the other hand, represents environment boundaries, which could anchor the path integrator to external landmarks (Hardcastle et al., 2015; Keinath et al., 2017). Similarly, head direction cells, speed cells, and other cell types perform functions related to spatial representations of the behavioral arena (Sargolini et al., 2006; Kropff et al., 2015; Høydal et al., 2019). Thus, place cells in the hippocampus likely inherit certain spatial properties from the MEC and integrate these functional inputs to form flexible, context-dependent, cognitive maps (Zhang et al., 2013; Moser et al., 2017).

Much less is known about the functions of the LEC. Lesion or genetic techniques have probed the contributions of LEC neurons to episodic memory (Van Cauter et al., 2013; Wilson et al., 2013; Yoo and Lee, 2017; Vandrey et al., 2020) and to the physiological properties of place cells. For example, rate remapping of CA3 neurons across different environments is moderately impaired after LEC lesion (Lu et al., 2013). On the other hand, single neuron recording studies have revealed that LEC cells are much less sensitive to allocentric space compared to MEC neurons (Hargreaves et al., 2005), even in cue-rich environments (Yoganarasimha et al., 2011). Instead, numerous investigations have suggested that LEC neurons carry information about individual items in the environment and/or their association with particular contexts. Olfactory input modulates LEC neuronal activity (Young et al., 1997; Leitner et al., 2016), and LEC neurons associate visual/auditory cues with spatial contexts (Pilkiw et al., 2017). LEC neurons show punctate firing fields related to objects placed in the arena (Deshmukh and Knierim, 2011) and they also fire at the remembered locations of displaced objects (Deshmukh and Knierim, 2011; Tsao et al., 2013). In contrast to the allocentric coding that is predominant in the MEC, LEC cells represent the angular bearing of objects and boundaries (or environmental centers) in an egocentric frame of reference (Wang et al., 2018). LEC neurons may also represent temporal context, likely by processing the temporal flow of sensory inputs related to experience (Tsao et al., 2018).

Structurally, the neurons of entorhinal cortex can be grouped into six layers based on cytoarchitectonic criteria, with different cell types dominating specific layers (except for layers I and IV, which have very few neurons) (Canto and Witter, 2012). Layers II and III of entorhinal cortex (superficial layers) receive convergent cortical inputs and directly innervate different subregions of the hippocampal formations (DG/CA3 and CA1/subiculum, respectively), thus constituting the main input pathways of the hippocampus. In contrast, layers V and VI (deep layers) receive the output from the hippocampus and generate diffuse projections to superficial layers of EC and to other cortical and subcortical areas; they are considered one of the main output pathways of the hippocampus. The aim of the current study was to investigate the presumably distinct behavioral correlates and computational functions of deep- vs. superficial-layer LEC. We predicted that deep layer LEC neurons would inherit spatial information from the hippocampal feedback projections and thus display markedly more allocentric spatial information than superficial LEC cells. Although the results were consistent with this hypothesis, the differences between the layers were modest. These experiments suggest that the feedback connections from CA1 do not endow deep LEC with properties of an allocentric map but instead modulate the egocentric processing that characterizes superficial LEC activity.

2. Methods

2.1. Subjects and surgical procedures

Seven male, adult Long-Evans rats (n = 7, 5–6 months old, obtained from Envigo or Harlan) were housed individually on a 12:12 h reversed light-dark cycle. Experiments were carried out in the dark phase of the cycle. Animal care, surgical procedures, and euthanasia were performed in accordance with NIH guidelines and were approved by the Institutional Animal Care and Use Committee at Johns Hopkins University or the University of Texas Health Science Center at Houston. Surgical procedures were performed under general anesthesia. Detailed surgical procedures were reported elsewhere (Yoganarasimha et al., 2011; Wang et al., 2018). Recording drives with 17–18 independently movable tetrodes were implanted and the tetrodes were either left just above the surface of the brain or lowered 0.9 to 1.3 mm into the brain during surgery. A craniotomy was made on the right hemisphere, 7.55–7.6 mm posterior to bregma and 3.0 mm lateral to the midline. The tetrode bundles were angled at 25° medio-laterally. After surgery, the rats were allowed to recover for 4–6 days. Rats were then food restricted until their body weight reached 80–90% of their free feeding weights. Tetrodes were then slowly advanced toward LEC over the course of 1–2 weeks.

2.2. Electrophysiology and recording

Tetrodes were made from 12 μm or 17 μm nichrome or platinum-iridium wire (California Fine Wire, Grover Beach, CA, USA or Kanthal, Palm Coast, FL). The tips of individual wires were electroplated with gold to 120–500 kOhm. During recordings, the electrophysiological signal was first buffered by a unity-gain preamplifier (Neuralynx, Bozeman, MT). For unit recordings, the signal was filtered between 600 Hz and 6 kHz and referenced against a tetrode in a cell-free area (white matter, lamina dissecans, or layer I). Waveforms above the threshold (~70 μV) were sampled for 1 ms at 32 kHz. Local field potentials were recorded against a ground screw anchored on the skull above the cerebellum or frontal cortex, filtered between 1 to 475 Hz, and continuously sampled at 1 kHz or 2 kHz. Red and green LEDs on the head of the subjects were used for tracking the position and head direction of the animals. A color CCD camera was used to capture the position of the rats at 30 Hz sampling rate.

2.3. Histological processing procedures

The rats were perfused transcardially with 4% formalin. The brains were extracted and submerged in 30% sucrose formalin solution. Brain tissue was cryosectioned at 40 μm thickness and mounted onto glass slides. Standard Nissl staining was performed to identify the tetrode tracks. Free-D software (Andrey and Maurin, 2005) was used to register tetrode tracks to the tetrode bundle configuration of the drive. Recording locations of each session were determined based on the amount of tetrode turning each day.

The demarcations of LEC followed standard conventions (Insausti et al., 1997; Burwell and Amaral, 1998; Dolorfo and Amaral, 1998a). Briefly, LEC was distinguished by the large and darkly stained neurons that formed discontinuous islands in layer II and the presence of a cell-sparse lamina dissecans between layer III and layer V. Layer II and III were grouped together as superficial layers and layer V and VI were grouped as deep layers.

2.4. Behavioral tasks

Two tasks were used in the current study.

1. The small-large box task. Details can be found in Savelli et al. (2008) and Yoganarasimha et al. (2011). Briefly, training began after rats recovered from surgery. Training was performed in a large box (135 × 135 × 30 cm³) for 30 min daily for about 10–14 days. The box was located in a room that had many uncontrolled landmarks (e.g., recording electronics racks, computer tables, etc.). During recording, a small box (58 × 58 × 30 cm³) was placed in the center of the large box. Rats first foraged for 6 min in the small box, after which the walls of the small box were removed with the rat still in the box, and the rat continued to forage in the large box for an additional 34 min. Data from two rats performing the small-large box task were included, but only data from the large box part of the task were used for data analysis.

2. The standard random foraging task. Five rats were trained to forage for small food pellets (BioServ, NJ) on a circular platform (122 cm diameter) or a square box (135 × 135 × 30 cm³) for 3–5 days before surgery and for one to two weeks after recovery from surgery while tetrodes were adjusted. For 2 rats, the room had large cues (flattened white cardboard box, poster boards, white circular board with random patterns, etc.) on a black curtain. For the other 3 rats, a rich set of cues was present in the room (stacked brown boxes, ladders, poster board, etc.) without a curtain. Each session lasted for 25–30 min. Rats also performed other tasks such as foraging on a linear track or a circular track in the same day (data not analyzed here). After recording, tetrodes were lowered a minimum of 80–120 μm. We waited for at least 3 hours before the next recording session once tetrodes were adjusted.

2.5. Unit isolation

Single units were isolated manually with a custom-written spike-sorting software (Winclust, J. Knierim). Peak amplitude and energy of the waveforms were used to isolate cells. The quality of each unit was rated with a score ranging from 1 (very good) to 5 (poor) (Supplementary Figure 1). The cluster isolation quality was assigned completely independent of any behavioral correlates of the cells. Units rated as 4 or 5 were excluded from the analysis.

2.6. Data analysis

All analysis were performed with Matlab. All statistical tests were two-sided. Only putative excitatory neurons (mean rate < 10 Hz) with a minimum of 50 spikes were analyzed. The behavioral trajectories of the rats were smoothed with a 5-frame boxcar filter (150 ms) and speed filtered (3 cm/s).

2.7. Spatial firing rate maps

Standard firing rate maps were created by first binning the behavioral arena into 6 cm bins, and the firing rate in each bin was calculated as the number of spikes fired when the rat occupied a spatial bin divided by the amount of time spent in that bin. The maps were smoothed with a Gaussian kernel (standard deviation: 2 bins). Only bins with a minimum occupancy of 0.5 s were analyzed.

2.8. Correlation analysis

To explore if the spatial firing properties are stable within a session, the session was divided into halves. Spatial rate maps of each half session were computed and the Pearson correlation coefficient was obtained between the two spatial rate maps for each cell.

2.9. Sparseness

Lifetime sparseness refers to the extent to which a cell fires to only a small proportion of possible stimuli, but at a relatively high firing rate (Willmore and Tolhurst, 2001). We used the definition of lifetime sparseness K_L, from a previous study (Willmore and Tolhurst, 2001):

$$K_L=(\frac{1}{M}\sum _{i=1}^M{(\frac{r_i-\overline{r}}{{\sigma }_r})}^4)-3$$

where r_i was the response in the i-th bin, ${\sigma }_r$ was the standard deviation, and M was the number of bins. Two bin sizes were used (1 sec and 0.125 sec).

2.10. Theta modulation

Spike trains were binned with 5 ms resolution and multitaper spectra were computed for the resulting firing rate vectors using mtspectrumc in the Chronux open source library (http://chronux.org/), with 3 tapers and 0.5 Hz bandwidth. We defined the theta modulation index as the ratio of theta band power (6 to 10 Hz) to total power (1 to 100 Hz). For 4 rats, we also defined the theta phase modulation index as in a previous study (Deshmukh et al., 2010), in which theta band LFP (6 to 10 Hz) was extracted, and the theta phase from the Hilbert transform was assigned to each spike. The distribution of spike phases was obtained for each neuron (15 degree per bin) and the peak of the distribution was normalized to 1. The theta modulation index was defined as 1 minus the minimum of the histogram.

2.11. Egocentric bearing tuning

We investigated the egocentric bearing tuning properties of LEC cells. The details of the applied analysis methods were reported in our previous study (Wang et al., 2018). Briefly, we calculated an egocentric bearing tuning curve (bin size 20°) with respect to the nearest environment boundaries; this measurement is approximately equivalent to measuring the egocentric bearing relative to the center of the environment (Wang et al., 2018). The mean vector length (MVL) was used to quantify the strength of the egocentric bearing tuning. Generalized linear models (GLM) were used to quantify the degree of egocentric vs. allocentric bearing preferences. We modelled the overall firing of the neurons with compound models that comprised a spatial component and an angular component.

$$\lambda =c{\lambda }_{\text{space}}{\lambda }_{\text{angle}}$$ (1)

where λ is the firing rate of the cell, c is the best-fit regression coefficient, and λ_space and λ_angle stand for the firing rates of the spatial component and the angular component, respectively. The spatial component was modelled with Zernike basis functions. The angular component could be allocentric head direction or egocentric bearing; both were modelled as linear summations of sine and cosine functions:

$${\lambda }_{\text{space}}=\text{exp}\left({\sum }_i{\beta }_i{Z}_i\right(\rho ,\psi \left)\right)$$ (2)

$${\lambda }_{\text{angle}}=\text{exp}({\sum }_j{\beta }_j\text{cos}(j\varphi )+{\beta }_j^{\prime }\text{sin}(j\varphi \left)\right)$$ (3)

where β_j and β_j′ are the parameters to be fit; $\rho$ and $\psi$ parameterize the Zernike function; and j is a preset parameter for determining the frequencies of sine and cosine functions. The model performance (goodness-of-fit) was accessed with a Bayesian Information Criterion (BIC):

$$BIC=-2\text{ln}\left(L\right)+k\text{ln}\left(n\right)$$ (4)

where L is the maximum likelihood function, k is the total number of parameters, and n is the number of data points. BIC performs model selection by penalizing the complexity of the model with the term k ln(n), thus avoiding overfitting. ΔBIC compares the model-fitting performance of the egocentric and allocentric models:

$$\Delta \text{BIC}=\frac{{\text{BIC}}_{\text{ego}}-{\text{BIC}}_{\text{allo}}}{{\text{BIC}}_{\text{ego}}+{\text{BIC}}_{\text{allo}}}$$ (5)

A value of ΔBIC < 0 indicates that the data were better fit by the egocentric model, and a value of ΔBIC > 0 indicates that the data were better fit by the allocentric model.

2.12. Speed tuning

An empirical speed index was defined as the Pearson’s correlation coefficient between speed and firing rate for the speed tuning curve. A standard spike train shuffling procedure (1000 times) was used to determine if a neuron had a significant tuning for speed with a significance level of 0.05. We performed GLM analyses to fit conjunctive spatial and speed tuning functions to the firing of a neuron:

$$\lambda =c{\lambda }_{\text{space}}{\lambda }_{\text{speed}}$$ (6)

where λ is the firing rate of the cell, c is the best-fit regression coefficient, and λ_space and λ_speed stand for the firing rates of the spatial component and the speed component, respectively. The spatial component was the same set of Zernike basis functions as described above for the egocentric analyses. The speed component was a 9-bin vector of running speeds (7.5cm/s to 47.5 cm with a step size of 5 cm/s) and was thus model-free. The GLM speed index was defined as the correlation coefficient between speed and firing for the fitted speed tuning curve.

2.13. Boundary-related firing

A border score for a cell was defined as the difference between the long edge of the largest firing field (> 20% of the peak) that touched any particular wall and the average distance between the pixels in the firing field to the nearest wall, normalized by their sum (Solstad et al., 2008). As the spatial firing of LEC cells were often not stable, we considered a cell as showing boundary-related firing only if (1) its border score was > 0.5; (2) its border score was > the 95^th percentile of a null distribution of border scores generated from 1000 random shuffles of the spike times relative to position samples of that neuron; and (3) the Pearson’s correlation coefficient of spatial rate maps from the first and second halves of the session was > 0.6.

2.14. Corner-related firing

For each cell, we defined several quantities which described the tendency of the cell to fire near the four corners of the environment. This is similar to a measure used previously (Muessig et al., 2015). We divided the spatial rate map of a neuron into four quadrants and calculated for each quadrant the normalized difference of firing between the corner and center portions of the map (regions a and b in Figure 6c, respectively). This measure was termed the corner score, and each neuron had four corner scores. A positive or negative corner score means the cell has a tendency to fire near the arena corners or centers, respectively. The significance of the corner score was established by shuffling the spike train and recalculating the measure from surrogate rate maps. A neuron was considered to be a corner cell if at least two of its corner scores were significantly greater than both the 95^th percentile of the corner score distribution from its own spike-shuffled rate maps and the 95^th percentile of the corner score distribution from all cells. We also analyzed the complement of the spatial rate maps (peak rate minus the spatial rate map) to search for cells that had a significant decrease in firing at the corners.

Figure 6.

Boundary- and corner-related firing properties in LEC neurons. (a) Examples of boundary-related cells in LEC. For each cell, 3 rate maps are presented: left to right, spatial firing rate map for the whole session, first half session, and second half session. (b) Examples of corner-related cells in LEC. Format is the same as (a). (c) Schematics showing the procedure for calculating the corner score from the spatial firing rate maps. Four corner scores were defined as the relative differences between center and corner portions of each quadrant of the spatial rate map.

To compare between populations, we calculated a measure termed the ensemble corner score, which was the corner score defined from a population rate map calculated as the sum of the normalized rate maps of individual cells. We limited the ensemble analysis to the cells recorded in the square arena that had a spatial information score greater than a certain threshold (0.05, 0.10, 0.15, 0.2 bits/spike). The conclusions were the same with different thresholds used.

We applied two methods to determine if the observed ensemble corner score is significantly different from null distributions. Both methods are ways of obtaining a spatial rate map from different random processes. (1) The field relocating procedure. For each spatial firing rate map, we thresholded the matrix with a cut-off of 20% of the peak firing rate and randomly relocated the individual blobs that had a minimum size of 5 pixels. The remaining pixels were randomly moved to remaining positions in the matrix. The new map thus had identically shaped blobs and first order statistics (e.g., mean pixel intensity) with the original map. (2) The standard spike shuffling procedure. For each cell, we circularly shifted the spike train relative to the position train, with minimum offset of 50 seconds, and obtained the resulting firing rate map. The two procedures test for the significance of the corner score from different perspectives. The first one preserves the overall spatial selectivity of each rate map while randomizing the locations of the firing fields while the second one preserves the temporal statistics of the spike train while decoupling the spikes from the rat’s location. The two protocols were repeated 1000 times, and the ensemble corner score was recomputed with the surrogate rate maps to form the null distributions.

We applied a permutation method to test if deep and superficial layer LEC neuron showed a significant difference in the ensemble corner score. We first lumped the spatial rate maps from both deep and superficial layer LEC neurons, and then randomly sampled without replacement the number of rate maps in deep layer populations as the surrogate deep layer group; the rest of the spatial maps were considered as the surrogate superficial group. A null distribution of differences of ensemble corner scores was built from 1000 permutations of the original samples and compared with the observed difference from the actual data.

3. Results

3.1. Recording area and behavioral setups

We recorded single units from the LEC while rats foraged for food pellets in a square box or on a circular platform. The recording locations spanned the medial to lateral extent of LEC (Figure 1). A total of 293 LEC units were recorded. Based on the histological analysis, we assigned the cells into the superficial-layer group (n = 153) and the deep-layer group (n = 49) following previous conventions (see Methods) (Insausti et al., 1997; Dolorfo and Amaral, 1998b), excluding those cells with < 50 spikes or with mean firing rates greater than 10 Hz (putative interneurons).

Figure 1.

Example cresyl violet stained coronal sections showing typical recording sites in the LEC. The arrowheads indicate example tetrode tracks. The LEC area was divided into three regions: the superficial layers (layer II and III), lamina dissecans (layer IV), and deep layers (layer V and VI). The dashed and solid lines in the figure denote the superficial- and deep-layers of LEC, respectively. The numbers indicate the rat identification.

3.2. The spatial information scores of deep-layer LEC neurons were slightly greater than those of superficial-layer LEC neurons

We examined classic, allocentric spatial rate maps of superficial-layer (Figure 2a) and deep-layer LEC neurons (Figure 2b). The deep-layer LEC neurons showed statistically significantly higher spatial selectivity than superficial-layer LEC neurons, as quantified by the spatial information score (Skaggs et al., 1993; Figure 2c, Supplementary Figure 2a, Supplementary Table 1) (Wilcoxon rank-sum test, Z = −2.19, p = 0.029; superficial layer: n = 153, deep layer: n = 49), but the difference was small (superficial layer: median = 0.12, interquartile range = 0.05 – 0.21; deep layer: median = 0.15, interquartile range = 0.08 – 0.28). On the other hand, the mean firing rate (superficial: median = 1.03, interquartile range = 0.43 – 3.02; deep: median = 0.79, interquartile range = 0.38 – 3.22) and peak spatial firing rate (superficial: median = 2.37, interquartile range = 0.99 – 4.83; deep: median = 2.19, interquartile range = 1.01 – 5.06) of the subregions of the LEC did not show significant differences (Wilcoxon rank-sum test; mean rate: Z = 0.06, p = 0.64; peak rate: Z = 0.06, p = 0.95). Therefore, the differences in spatial information score are unlikely to be simply explained by different firing rates of the neurons in the two subregions. These results suggest that, like their counterparts in superficial layers, the deep-layer LEC cells typically do not carry strong, allocentric spatial information as seen in the hippocampus and other neighboring areas, such as the MEC.

Figure 2.

Comparison of spatial information scores of superficial- and deep-layer LEC neurons. (a) Fifteen example rate maps of cells in the superficial layers of LEC. The number on top for each cell indicates the peak firing rate. White and black shades in the rate maps indicate 0 Hz (no firing) and maximal firing rates, respectively. (b) Fifteen example rate maps of cells in the deep layers of LEC, as shown in (a). (c) Cumulative distribution of spatial information scores of superficial- (black) and deep-layer (gray) LEC neurons. Deep-layer LEC neurons had significantly larger spatial information scores than superficial-layer neurons, although the difference was small. (d) Cumulative distribution of the Pearson’s correlation coefficients between the rate maps of the first and second halves of the session. Deep-layer LEC neurons had significantly larger correlations than superficial-layer neurons.

We investigated whether there were differences between deep and superficial LEC in the proportion of individual cells that exhibited significant spatial information. For each cell, its spatial information score was compared to a null distribution created from standard spike shuffling procedure (see Methods), and a cell was considered statistically significant if its information score was greater than 99% of the shuffled data points. The number of significant cells in superficial-layer LEC neurons was not different from deep-layer LEC cells (superficial: 65/153, deep: 24/49; χ² = 0.63, p = 0.42), even though (as shown above) deep-layer cells had on average slightly higher spatial information than superficial-layer cells. We next tested whether the weak spatial tuning of the cells was stable across a recording session by calculating the Pearson’s correlation coefficients between spatial rate maps of the first and second halves of the session. Deep-layer LEC cells had significantly larger correlation coefficients than superficial-layer LEC cells (Figure 2d, Supplementary Figure 2b; Wilcoxon rank-sum test, Z = −1.98, p = 0.048; superficial: median = 0.35, interquartile range = 0.16 – 0.57; deep: median = 0.45, interquartile range = 0.19 – 0.80). Thus, the spatial firing patterns of deep-layer LEC neurons were more stable than those of superficial-layer LEC neurons.

Previous reports have suggested that in primary sensory cortical areas, superficial layer neurons fire more sparsely than deep layer neurons (Barth and Poulet, 2012). We compared the lifetime sparseness of superficial- and deep-layer LEC neurons using the kurtosis measure (Willmore and Tolhurst, 2001). Lifetime sparseness refers to the extent to which a particular cell responds to a very small fraction of possible stimuli, but fires at a relatively high rate when it is presented with one of these stimuli. In this case, we defined temporal bins as the different inputs to the cells. There were no significant differences between the lifetime sparseness of deep and superficial LEC neurons, irrespective of the bin size (0.125 s bin size, median of deep-layer LEC neurons = 26.97, median of superficial-layer LEC neurons = 24.78, Z = 0.22, p = 0.82; 1 s bin size, median of deep-layer LEC neurons = 10.91, median of superficial-layer LEC neurons = 7.80, Z = 1.33, p = 0.18, Wilcoxon rank-sum test).

3.3. Theta modulation is weak in both deep and superficial layers

Theta oscillation is an important mechanism to organize information flow within the spatial navigation circuit (Buzsáki, 2005). Although in LEC, both local field potentials and single units show much weaker theta modulation than other parts of the hippocampal system (Deshmukh et al., 2010; Fernández-Ruiz et al., 2017) (Figure 3a and 3b), it is likely that there are functional heterogeneities among LEC neurons; neurons with stronger theta modulation are likely to be functionally more coupled with MEC/hippocampus and thus show greater spatial selectivity. Theta modulation of single units was measured by spectral analyses of spike trains and by creating histograms of the phase of the LFP theta rhythm corresponding with each spike (see Methods). Similar to Deshmukh et al. (2010), we found that LEC neurons showed little theta modulation with the spike-train spectral analysis method, but they showed modest theta modulation when spikes were assigned to phases of LFP theta. Moreover, theta modulation was a statistically significant predictor of the spatial information score with either method, although the correlation was very weak with the spike-train spectral analysis method (Figure 3c, spike train spectrum-based theta modulation index, Spearman’s correlation ρ = 0.18, p < 0.003; Figure 3e, LFP-based theta phase modulation index, Spearman’s correlation ρ = 0.62, p < 0.001). We then tested whether deep-layer LEC neurons had more theta-modulated neurons than superficial-layer neurons. There was no significant difference between the two layers using the spike train spectrum method (Figure 3d; Wilcoxon rank-sum test, Z = −1.95, p = 0.051, median of deep cells: 0.04, median of superficial cells: 0.042), although the results showed a statistical trend in the direction opposite from the a priori prediction that deep-layer cells would show stronger theta modulation. In contrast, there was a statistical trend in the predicted direction when the LFP-based theta modulation index was employed (Figure 3f; Wilcoxon rank-sum test, Z = 1.88, p = 0.06, median of deep cells: 0.48, median of superficial cells: 0.39). Overall, it appears that differences in spatial properties between deep- and superficial-layer LEC neurons may be only weakly, if at all, attributed to differences in the strength of theta modulation.

Figure 3.

Theta modulation of neural firing in deep- and superficial-layers of LEC. (a) Representative spike train autocorrelations of superficial-layer LEC neurons. Most cells did not show evidence of theta-modulated peaks. (b) Example spike train autocorrelations of deep-layer LEC neurons. Numbers on top indicate the spike train spectrum-based theta modulation index. Most cells did not show evidence of theta-modulated peaks. (c) Scatter plot of the spatial information and spike train spectrum-based theta modulation index for all LEC neurons. (d) Cumulative distribution of the theta modulation index for both superficial- and deep-layer LEC neurons. (e) and (f), same as (c) and (d) but for LFP-based, theta phase modulation index.

3.4. The egocentric bearing coding of boundaries is comparable between superficial- and deep-layer LEC neurons

Certain LEC neurons have been shown to represent environment boundaries (or centers) in an egocentric frame of reference (Wang et al., 2018). Sensory information (presumably egocentric) from upstream areas converges onto the superficial layers of LEC (Doan et al., 2019), while the allocentric spatial feedback from CA1 and subiculum may dominate the deep LEC layers. Therefore, we investigated whether there were differences between deep and superficial layer LEC neurons in the context of egocentric representations. First, we constructed egocentric bearing tuning curves for all LEC neurons (see Wang et al., 2018, for details) and found both superficial and deep layers contained neurons that were selective for egocentric bearing of the nearest boundaries (Figure 4a and 4b). (Note that, as described in Wang et al., 2018, egocentric bearing to nearest boundaries is equivalent to egocentric bearing to the center of a radially symmetric environment, and it is difficult to dissociate these two possible reference points even in a square environment). We applied a Generalized Linear Model to model the data and used the normalized difference of the Bayesian Information Criterion (ΔBIC_boundary) to quantify the extent of egocentric/allocentric coding for the nearest boundaries. The ΔBIC_boundary measure is negative or positive if the cell prefers egocentric or allocentric coding, respectively. We found that most neurons in both superficial- and deep-layer LEC preferred egocentric coding (Superficial: n = 152, median = −0.0006, interquartile range: −0.0048 − 0.0007; Deep: n = 49, median = −0.0008, interquartile range = −0.0036 − 0). There was no significant difference between the two groups (Figure 4c; Wilcoxon rank-sum test: Z = 0.55, p = 0.58).

Figure 4.

Comparison of egocentric properties of superficial- and deep-layer LEC neurons. Some LEC neurons had spatial firing patterns associated with the boundaries or corners of the apparatus. (a) Two example superficial-layer LEC neurons (rows) with egocentric bearing tuning properties with respect to the nearest boundary. Left image, spatial firing rate map, number on top denotes peak firing rate; right curve, egocentric bearing tuning curve, number on top denotes mean vector length. (b) Same as (a) for two example deep LEC neurons. (c) Distributions of ΔBIC_boundary for superficial- and deep-layer LEC neurons.

3.5. Speed tuning is comparable between superficial- and deep-layer LEC neurons

Since the firing rate of a proportion of neurons in both the hippocampus and MEC have been shown to be tuned to running speed (McNaughton et al., 1983; Czurkó et al., 1999; Sargolini et al., 2006; Kropff et al., 2015), we next explored if this property was shared by LEC neurons. We found that 28.3% (83/293) of LEC cells had significant speed tuning (see Methods) (Figure 5a). Most LEC neurons had a negative empirical speed index (the correlation between speed and firing rate); that is, they preferred to fire at lower speeds (Figure 5b; median r = −0.30; Wilcoxon signed-rank test, Z = −4.51, p < 0.001), and this was true for statistically significant, speed selective neurons as well (Figure 5b; median = −0.78; Wilcoxon signed-rank test, Z = −4.49, p < 0.001). Because the rats tended to move more slowly at the corners and boundaries of the environment (Supplementary Figure 3c), it was possible that spatially correlated firing at these locations would provide a confounding influence on the speed tuning. However, the results were the same after removing the potential confound of spatial coding with the GLM framework (Figure 5c, see Methods, Wilcoxon signed-rank test, p < 0.001 for both cases; Figure 5d, Pearson’s r = 0.92, p < 0.0001). Together, these results suggest that speed preference for LEC neurons cannot be simply explained by spatial selectivity for locations in which the rats were biased toward slower speeds. The proportions of speed selective neurons in superficial and deep layers were not significantly different (superficial: 40/153, deep: 15/49; χ² = 0.37, p = 0.54), and the empirical speed index values between the two subregions were also similar (Figure 5e; Wilcoxon rank-sum test, Z = 0.38, p = 0.7).

Figure 5.

Comparison of speed tuning properties of superficial- and deep-layer LEC neurons. Some LEC neurons demonstrated tuning for running speed. (a) Speed tuning curves of 20 example speed selective LEC neurons. (b) Distribution of the empirical speed index for all sampled LEC cells and for the subset of speed selective cells. (c) Same as (b) for the GLM speed index. (d) Scatter plot of the empirical speed index and the GLM speed index for speed selective LEC neurons. (e) Cumulative distribution of the empirical speed index for both superficial- and deep-layer LEC neurons.

3.6. Some LEC neurons showed a tendency to fire near the boundaries and corners of the environment

Boundary (or border) cells are found in a number of brain regions, including the MEC (Savelli et al., 2008; Solstad et al., 2008) and subiculum (Lever et al., 2009). We also observed 8 cells in LEC that passed a statistical test for border selectivity (see Methods) (Figure 6a). Three and four boundary cells were located in superficial- and deep-layer LEC, respectively, and another cell was located too close to the boundary between layers to assign recording location with confidence.

Some LEC neurons showed elevated firing around the arena corners (Figure 6b) and we defined a corner score to quantify this firing bias (Figure 6c; also see Methods). A total of 12 statistically significant corner cells were found, four and three of which were from superficial- and deep-layer LEC, respectively (the other 5 cells could not be localized to deep or superficial layers with confidence). The mean corner score (possible range: −1 to 1) for the significant corner cells was 0.58 ± 0.16 (mean ± S.D.). Of these 12 corner cells, 9 showed preferred egocentric tuning (negative ΔBIC_boundary, median = −0.004, Wilcoxon signed-rank test, p = 0.054). The corner selectivity could not be explained by biased spatial sampling around the corners, speed preference, or selectivity for acceleration or deceleration (Supplementary Figure 3). We also identified 7 “anti-corner/border cells” (or bulls-eye cells) which fired much less near corners or borders (Supplementary Figure 4) (Solstad et al., 2008; Weible et al., 2012). The corner firing bias was also evident in the normalized, population-averaged spatial firing rate map for all the cells that had a spatial information score > 0.2 bit/spike (Figure 7a, leftmost map). To quantify corner-related firing in a population of cells, we defined an ensemble corner score, which is the normalized difference of mean firing in the center versus that in the corners of the ensemble firing rate map (see Methods). The observed ensemble corner indices were significantly different from the null distributions for both the field relocating procedure and the standard spike shuffling procedure (see Methods), and the conclusions were the same when the spatial information threshold was varied from 0.05 bit/spike to 0.2 bit/spike (Figure 7b: the field relocating procedure; Figure 7c: the standard spike shuffling procedure; p < 0.001 for all cases). There were no significant differences between corner representations of the deep- and superficial-layers of LEC (Figure 7d); in all cases, the shuffled data distributions were centered near 0 and the real data (vertical lines) fell within the 95% confidence interval of the shuffled data.

Figure 7.

Comparison of corner-related activity between superficial- and deep-layer LEC neurons. (a) Left, average of the normalized spatial rate maps for cells that had spatial information scores greater than 0.2 bit/spike. Right images, three instances of ensemble average fields after randomly relocating the blobs in the individual maps that had a minimum of 5 pixels that exceeded 0.2 bit/spike (numbers on top, ensemble corner score). (b) The ensemble corner score (vertical line) is significantly different from a null distribution created by randomly relocating the spatial blobs in the rate maps. Left to right, real and shuffled ensemble corner score from spatial rate maps with information score greater than 0.05, 0.10, 0.15, 0.2 bit/spike, respectively. (c) Same as (b) except that the statistics were obtained with standard spike shuffling procedures in which the spike trains were circularly shuffled by a random amount relative to the position train. (d) The ensemble corner score of superficial- and deep-layer LEC neurons shows no significant differences. Left to right, real and shuffled ensemble corner score differences (superficial − deep) from spatial rate maps with information score greater than 0.05, 0.10, 0.15, 0.2 bits/spike, respectively.

Discussion

In this study, we examined the spatial firing properties of the input and output layers of LEC, inspired mainly by their respective anatomical relationships with components of the medial temporal lobe in terms of connectivity and cell type differences (Insausti et al., 1997; Canto et al., 2008; Agster and Burwell, 2009, 2013). We found that: (1) the deep-layer LEC neurons were only slightly more informative about allocentric space than the superficial-layer LEC cells; (2) the extent of egocentric bearing tuning of environment boundaries was comparable between superficial- and deep-layer LEC neurons; 3) some LEC cells demonstrated speed tuning, with the majority preferring slower speed; 4) spatially selective LEC cells tended to fire near the boundaries and corners of the environment.

Our initial prediction was that deep-layer LEC neurons would contain much more allocentric spatial information than superficial-layer LEC neurons. Superficial-layer LEC neurons receive projections from inputs mainly from perirhinal and postrhinal cortex (Doan et al., 2019), both of which show weak or large-scale spatial selectivity (Deshmukh et al., 2012; Furtak et al., 2012; Bos et al., 2017; Connor and Knierim, 2017). In contrast, deep-layer LEC receives inputs from subiculum and distal CA1, both of which are considerably selective for space and would presumably endow deep-layer LEC cells with much greater spatial biases than superficial-layer cells (Sharp and Green, 1994; Henriksen et al., 2010; Deshmukh et al., 2012; Furtak et al., 2012; Gofman et al., 2019; Deshmukh, 2021). In addition, deep vs. superficial functional differences have been reported in MEC (Sargolini et al., 2006). Because the microcircuitry of LEC is similar to that of MEC, it was predicted that functional dissociations between deep and superficial layers would likewise be found in LEC (Nilssen et al., 2018). Indeed, fan cells in layer 2 of LEC have been suggested to be specifically required for sophisticated binding of object, place, and context information, but not simpler object-context association, in contrast to full LEC lesions (Vandrey et al., 2020). Although we found that deep-layer LEC cells tended to have higher spatial selectivity and more stable spatial firing patterns than the superficial-layer LEC cells, it is perhaps surprising that there were only minor, quantitative differences in the spatial firing patterns between these two subregions. The lack of a strong difference between superficial and deep layers may be explained by the fact that deep-layer LEC is also innervated by brain regions with much less spatial information, e.g. prefrontal cortex (Fujisawa et al., 2008; Zielinski et al., 2019). Furthermore, there is local connectivity between different layers of the cortex arising from complex dendritic and axonal organizations, thus potentially blurring the differences between the two subregions of LEC (Canto et al., 2008; Ohara et al., 2018). The lack of difference in firing sparseness between the two subregions also suggests that both systems employ a highly sparse code for the representation of memory- and navigation-related information. The present results reinforce the view that LEC is not merely a relay station between association cortex and the hippocampus, and that integration and refinement of information exist across different layers of the entorhinal cortex (Canto et al., 2008; Knierim et al., 2014).

Theta oscillation dominates the local field potential in the hippocampus and provides temporal organization for hippocampal communication with other brain regions (Sirota et al., 2008; Mizuseki et al., 2009; Schomburg et al., 2014; Tamura et al., 2017; Quirk et al., 2021). In this regard, theta-modulated activity might serve as an indicator for the coupling strength between LEC neurons and the rest of the spatial navigation circuit, and the extent of theta modulation might correlate with their spatial coding properties. Indeed, we observed a correlation between the spatial information score and the theta modulation index. It is likely that the sensory processing which imparts LEC neurons with the spatial selectivity was accompanied by sporadic bouts of theta oscillation (Xu and Wilson, 2012), in contrast to the sustained theta oscillations that are present in MEC and hippocampus. Nevertheless, deep and superficial neurons cannot be distinguished by the extent of theta modulation, suggesting that both subregions operate similarly in this aspect. This adds to the growing body of evidence that, functionally, LEC operates in a different way than MEC.

Our results demonstrate that both superficial- and deep-layer LEC participate comparably in egocentric processing, with no significant differences with respect to the egocentric tuning of environment boundaries (or the apparatus center, which were equivalent in the current experiment). The field has seen a renewed interest in egocentric coding properties in the hippocampus and related regions in recent years (Wang et al., 2018, 2020; Gofman et al., 2019; LaChance et al., 2019; Alexander et al., 2020; Kunz et al., 2021; Mao et al., 2021; Ormond and O’Keefe, 2022). LEC receives convergent multisensory input from a variety of brain regions and is part of larger circuits involving parietal cortex and postrhinal cortex that show coding of spatial information in an egocentric frame of reference. Does this egocentric representation of external items only pertain to the inputs to the hippocampus, or is it a shared property across the layers of LEC? Our results show that the LEC output stream that goes back to neocortical areas also carries egocentric information. It thus appears that allocentric spatial information from distal CA1 and proximal subiculum may be recoded into egocentric coordinates in LEC. Future study on the substrate for this particular coordinate transformation is needed.

Egocentric processing in LEC is consistent with the “local view” idea proposed for spatial processing in theoretical and computational studies (McNaughton et al., 1989; Sharp, 1991; Redish, 1999). Egocentric representations are dependent on the relationship between the subject and external landmarks, and this first-person perspective may create view-specific responses that work together with the path integration system to form the cognitive map (O’Keefe and Nadel, 1978). View-specific responses have been unveiled under body-fixed virtual reality conditions, in which the normal, allocentric firing of place cells was replaced by egocentric responses to specific visual items in the environment (Purandare et al., 2022). Note that egocentric representations could be based on senses other than vision; for example, egocentric border firing in retrosplenial cortex persists in darkness (van Wijngaarden et al., 2020). LEC is a multisensory gateway to the hippocampal memory system and receives rich, presumably egocentric, information from perirhinal and postrhinal cortices (Doan et al., 2019; LaChance et al., 2019). Given that egocentric frames of reference have been found in a variety of association and sensory areas associated with the medial temporal lobe system (Wang et al., 2018, 2020; Gofman et al., 2019; LaChance et al., 2019; Alexander et al., 2020; Kunz et al., 2021; Long et al., 2021; Mao et al., 2021; Ormond and O’Keefe, 2022), the presence of egocentric processing in the output layers of LEC might facilitate information transfer between the hippocampus and parahippocampal areas during memory consolidation and recall (Luo et al., 2022). In particular, Wang et al. (2018) have argued that the LEC provides information to the hippocampus about the content of an experience, experienced from the subject’s egocentric (first-person) viewpoint and remembered from that same viewpoint as an episodic memory. These responses would be similar to view cells reported in the primate hippocampus (Rolls and O’Mara, 1995; Georges-Françoiss et al., 1999), although perhaps not as sharply tuned, given the non-foveal, wide field of view of the rodent compared to the primate. These considerations predict that the egocentric coding of LEC cells may provide the inputs that generate view-coding in the primate hippocampus, as well as view-coding that is unmasked under certain virtual reality conditions (Jercog et al., 2019; Purandare et al., 2022).

To our knowledge, we have demonstrated for the first time that certain LEC neurons are sensitive to running speed with most of them preferring lower speed. Speed representation is essential for navigation (McNaughton et al., 2006), and the entorhinal-hippocampal network has been shown to represent speed with either a firing rate code (McNaughton et al., 1983; Czurkó et al., 1999; Sargolini et al., 2006; Kropff et al., 2015; Hinman et al., 2016) or frequency of oscillation code (Vanderwolf, 1969; Burgess et al., 2007; Geisler et al., 2007; Hinman et al., 2016). In MEC these two signals could be dissociated (Hinman et al., 2016). However, few studies have examined the speed representation in LEC. Compared with MEC, theta oscillations in LEC are much weaker (Deshmukh et al., 2010) for both local field potential and spike firing, making it unlikely to represent running speed with an oscillation frequency code. However, we found there was a significant correlation between speed and firing rate in many LEC neurons. The finding that most speed-selective cells had negative speed modulation is particularly striking given that the majority of MEC and hippocampal neurons are positively modulated by speed (McNaughton et al., 1983; Czurkó et al., 1999; Sargolini et al., 2006; Kropff et al., 2015; Hinman et al., 2016). As LEC is thought to process multimodal sensory inputs about external items (Knierim et al., 2014; Doan et al., 2019), it is possible that these negatively speed-modulated cells in LEC are sensitive to sensory inputs that are available to the animals when they are moving at a slow speed, which is presumably when they are paying more attention to their surroundings.

We identified boundary- and corner-related firing in both superficial- and deep-layer LEC neurons. Geometry is a powerful cue in orienting the subject during navigation (Cheng, 1986; Keinath et al., 2017; Zeng et al., 2022); arena boundaries and corners are among the most salient features of the environment and constitute important nodes that determine the geometry of the arena. There are rich local sensory features afforded by the boundaries or corners, e.g., tactile inputs and visual inputs (shadows). LEC neurons could integrate these sensory inputs and form a spatial representation of nearby boundaries or corners. Together with boundary-related firing in MEC and subiculum (Savelli et al., 2008; Solstad et al., 2008; Lever et al., 2009), boundary and corner representation in LEC could stabilize the spatial representation and bind the path integration system to the external environment. Because of the rats’ bias for slower movement in corners, one must be cautious in interpreting the corner-selective firing and speed-selective firing reported here, as either could be an artifactual result of the other. Our GLM and other analyses provide evidence that both of these variables make contributions to LEC firing, and it is unlikely that either one can be explained as a simple confounding influence of the other. Nonetheless, it is likely that they may be intertwined in some cells, and further experiments will be necessary to tease apart the relative influence of speed and corner selectivity on LEC activity.

The laminar structure of the cortex provides essential organizing principles with respect to the functions of cortical circuits. Entorhinal cortex is situated as an important nodal point between the neocortex and the hippocampus, with superficial layers relaying neocortical input to the hippocampus and deep layers broadcasting the hippocampal output back to neocortical regions (Agster and Burwell, 2013). Functionally, superficial- and deep-layer entorhinal cortex have been reported to be associated with the encoding and retrieval of memories, respectively (Maass et al., 2014). On the level of single cells, considerable progress has been made in understanding the functional heterogeneity of MEC neurons between different layers. For example, the proportion of grid cells is largest in layer II of MEC, and conjunctive grid by head direction cells are only present in layer III and deep-layer MEC (Sargolini et al., 2006); furthermore, phase precession can be observed in layer II but not layer III MEC neurons (Hafting et al., 2008). Although the laminar structure of the cortex clearly indicates a certain degree of segregation of information processing, the projection characteristics of neurons make it possible to transmit and process information across different layers or even sublayers of entorhinal cortex (Sürmeli et al., 2015; Ohara et al., 2018; Rozov et al., 2020; Gerlei et al., 2021), which could provide potential mechanisms for building functional cell types, e.g., grid cells (Gerlei et al., 2020). However, given that the intrinsic anatomical organization of LEC is grossly similar to MEC, similar cross-layer functional heterogeneities may still exist in LEC but along different coding dimensions from those examined here, e.g., binding objects to spatial contexts (Vandrey et al., 2020) or feedforward and feedback processing of sensory stimuli (Leitner et al., 2016), and investigating these functions across layers of LEC could conceive a fruitful way to gain mechanistic insights into its involvement in episodic memory. Together, we have demonstrated that the superficial- and deep-layer LEC neurons have minor differences in allocentric spatial information processing, and behave similarly in egocentric properties. Both subregions demonstrate boundary or corner-related firing. The results suggest that there is a considerable functional interaction between input and output regions of LEC, and that the hippocampal outputs were transformed to different coding dimensions than the pure, allocentric spatial domain.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.