Dorsoventral gradient of theta sweeps in the medial entorhinal cortex Authors: Zilong Ji, Huiwen Zhang, Callum Marshall, Rokas Stonis, Neil Burgess Published: 2026 ABSTRACT During random foraging, the positional signal decoded from entorhinal grid cells exhibits left-right theta sweeps, alternating from one side of the head direction to the other across successive theta cycles. Here, we report that theta sweeps are topographically organized along the dorsoventral axis of the medial entorhinal cortex, with the angular deviation from head direction increasing gradually from dorsal (smaller scale) to ventral (larger scale) modules. This gradient coexists with a corresponding dorsoventral increase in angular deviation decoded from theta-modulated direction cells, which drive grid cell theta sweeps. These phenomena parallel a broadening of head direction tuning and increasing occurrence of theta cycle skipping in single-cell firing along the dorsoventral axis. Computational modeling demonstrates that these patterns are consistent with continuous attractor dynamics and a dorsoventral gradient in firing rate adaptation. These results highlight how theta sweeps can simultaneously represent multiple potential future locations and reveal a clear neural mechanism underlying this process. INTRODUCTION Theta sweeps are dynamic neural representations observed in the hippocampal formation during the theta rhythm (6 to 12 Hz). In place and grid cells, theta sweeps represent internally generated trajectories that extend outward from the animal’s current location. The trajectory in one theta cycle can appear as forward-directed progression in linear mazes (1), as left-right alternating progressions in T-shape mazes (2, 3) and open-field environments (4). They are driven by theta-modulated direction cells (tmDCs), which represent internal directional sweeps from side to side of the head direction (4). These population-level phenomena correspond to single-cell firing properties identified in earlier studies, including theta phase precession in place (5), grid (6), and tmDCs (7), as well as theta skipping (i.e., firing on alternate theta cycles) in these cell types (3, 7–11). Theta sweeps reflect sequential activation of cell ensembles within individual theta cycles. This temporal organization occurs on a similar time scale as spike timing–dependent plasticity (STDP) (12), suggesting a mechanism by which theta sweeps could support temporal sequence learning and memory encoding (1, 13–15). Theta sweeps during open-field navigation may facilitate efficient sampling of space by transiently activating cells representing locations on the sweep path, thereby reducing the need for exhaustive behavioral exploration (4, 16) and possibly aiding the formation of cognitive maps. Furthermore, recent findings indicate that theta sweeps in open-field navigation not only alternate stereotypically between left and right but can also encode directions toward future goal locations, independently from head or movement direction (17–19). The hippocampal formation exhibits multiple forms of topographical organization along the dorsoventral axis. In the medial entorhinal cortex (mEC), grid cells are arranged into discrete modules, with grid spacing increasing discontinuously from dorsal to ventral regions (20, 21). In the hippocampus, place cells show a corresponding dorsoventral increase in firing field size (22–24). Similarly, in mEC layer III, head direction cells exhibit sharper tuning dorsally and broader tuning ventrally (25). In addition to these functional gradients in spatial tuning, intrinsic physiological properties are also spatially organized. For example, firing rate adaptation (FRA) in mEC layer II principal cells increased progressively toward ventral regions (26). Membrane conductance in layer II stellate neurons likewise exhibit a gradient, potentially reflecting gradients in HCN and leak potassium channels (27), as does the frequency of subthreshold membrane potential oscillations and resonance (28). However, it remains unclear whether population-level dynamics are similarly structured along the dorsoventral axis and which neural mechanisms might underlie such organization. Our recent work identified FRA as a potential mechanism generating theta sweeps at the population level (16), suggesting that circuit-level dynamics may also be topographically organized as a consequence of the graded organization observed at the single-cell level. To address this question directly, we systematically examined whether population-level theta sweeps exhibit a structured gradient along the dorsoventral axis of the mEC. Analyzing additional features of a previously published dataset (4), we reported that, in rats foraging for randomly scattered food, left-right theta sweeps in grid cells exhibit a pronounced gradient along the mEC dorsoventral axis, with sweep angles increasing from dorsal to ventral. A corresponding gradient is also found in tmDCs in the parasubiculum, where sweep angles similarly increase ventrally. These findings indicate that the geometric structure of theta sweeps varies systematically across the mEC and parasubicular dorsoventral axis, paralleling previously known gradients in single-cell features including grid spacing (20) and head direction tuning (25) and suggesting that grid cell firing in different modules does not correspond to a common location at later phases of theta. We further suggest that this gradient in sweep angles may arise from a dorsoventral increase in neuronal FRA (26), which has been proposed to generate theta sweeps in a ring-attractor model of tmDCs and to drive corresponding theta sweeps in downstream grid cells (16). Such an organization may enable the brain to simultaneously represent multiple potential future locations in the environment, compared to when dorsoventral theta sweep directions are aligned, which may be important for efficient navigation, as highlighted in modeling studies (29–31). RESULTS Topography of sweep angles in grid cells We analyzed data from Vollan et al. (4), comprising 29 open-field running sessions recorded from 17 Long-Evans rats. To decode theta sweeps, we followed the method described by Vollan et al. (4), correlating the instantaneous population activity (PV) of all recorded cells in each temporal bin with the session-averaged PV in each spatial bin (see Materials and Methods). The decoded position was then assigned to the spatial bin showing the highest correlation score across all bins. Using this approach, we replicated the decoding of left-right theta sweeps: Across successive theta cycles, the decoded position swept outward from the animal’s location into the surrounding environment, with the sweep direction alternating between left-forward and right-forward (Fig. 1, A and B). The averaged sweep angles range from 5.0° to 47.8° across all recording sessions (median = 24.1°; interquartile range (IQR): [16.1°, 29.0°]), like results reported by Vollan et al. (4). Topography of sweep angles in grid cells along the mEC dorsoventral axis. (A) Example of left-right alternation across four successive theta cycles, generated from data reported by Vollan et al. (4). Top: Summed spike counts of all recorded cells over a short period, displaying theta rhythmicity. Bottom: Decoded sweeps in four consecutive theta cycles, with dashed lines indicating the animal’s trajectory over the same period as in the top plot. (B) Example sweeps from all recorded cells across the recording period showing in (A), with blue denoting left sweeps and orange denoting right sweeps. (C) Example sweeps from grid cells in each grid module. (D) Left: Schematic outline of the left entorhinal cortex in a rat brain. The dorsoventral gradient of the entorhinal cortex (magenta to blue) spans both the lateral and medial entorhinal subdivisions, with the border indicated by the orange line. Right: Mean left-right sweeps decoded from grid module 1 (magenta), module 2 (purple), and module 3 (blue), in two animals. (E) Mean sweep angle across three grid modules. Each dot represents the sweep angle decoded from one grid module (48 modules in total). Gray lines link the mean sweep angles from grid modules in the same recording session (19 sessions from 11 animals). An LMM yielded a substantial effect of module identity on sweep angle, with χ2(2) = 22.7, P = 1.2 × 10−5. Averaged sweep angles were module 1, 14.6°; module 2, 18.2°; and module 3, 23.8°. Module 2 versus module 1: β = 3.6, P = 0.003. Module 3 versus module 2: β = 5.0, P = 0.001. *P < 0.05, **P < 0.01, ***P < 0.001. Because grid cells are organized into multiple grid modules along the mEC dorsoventral axis (20, 21), we next asked whether sweep features are different across grid modules. To do this, we decoded theta sweeps from grid cells in each module independently (Fig. 1C). Only sessions containing at least two grid modules, each with no fewer than 40 grid cells, were included in the analysis (19 sessions from 11 animals, 48 modules in total). This resulted in 42 to 215 grid cells per module (median = 100; IQR: [69 to 129]). We then calculated the mean sweep angle for each module and found that the sweep angle increased across grid modules along the dorsoventral axis (see Fig. 1D for results from two rats). A linear mixed-effects model (LMM) revealed a substantial effect of module identity on sweep angle [Fig. 1E; χ2(2) = 22.7, P = 1.2 × 10−5]. Averaged sweep angles were module 1, 14.6°; module 2, 18.2°; and module 3, 23.8°. Module 2 versus module 1: β = 3.6, P = 0.003. Module 3 versus module 2: β = 5.0, P = 0.001. To control for potential confounds arising from unequal cell numbers across modules, which could bias the decoded sweep angle (fewer cells leading to noisier decoding), we subsampled each module to match the number of grid cells in decoding and repeated the analysis. The results still showed a substantial module effect and revealed a systematic increase in sweep angles along the mEC dorsoventral axis [LMM with χ2(2) = 14.7, P = 6.4 × 10−4; fig. S1A]. Furthermore, a recent study reported that sweep angle varies with theta frequency during rapid eye movement (REM) sleep (19), with larger sweep angles observed during slow theta oscillations (<6.5 Hz) than during fast theta oscillations (>8.5 Hz). We therefore examined the theta frequency derived from the summed spike activity in each module and did not observe any notable differences between modules [fig. S2; LMM with χ2(2) = 4.77, P = 0.09]. These results ruled out a possibility that a gradient decrease in theta frequency along the dorsoventral axis accounts for the increase in sweep angle observed in the data. Grid cell sweep angle topography reflects that of tmDCs Sweeps in grid cells align with those in tmDCs (20 to 450 per recording session, median = 194; IQR: [85 to 287]; Fig. 2A) (4), which have been identified in the anteroventral thalamus nuclei (AVN) (11, 32), parasubiculum, and mEC (4, 10, 33). First, we selected tmDCs (see Materials and Methods) and performed analyses by combining tmDCs in both the parasubiculum and the mEC. Then, we showed that the decoded direction (hereafter referred to as the internal direction) sweeps continuously back and forth to either side of the head axis (fig. S3), rather than flickering abruptly from side to side of the head axis (4), across successive theta cycles. Furthermore, it has been shown that the locational sweep direction decoded from grid cells matches the internal direction decoded from tmDCs (Fig. 2B) (4). We next asked whether tmDCs exhibit a gradient in sweep angle along the dorsoventral axis. Topography of sweep angles in tmDCs along the parasubicular-mEC dorsoventral axis. (A) Left: tmDC with its directional tuning field relative to head direction. Middle: Segment of the cell’s spike train overlaid on summed spike counts of all corecorded cells with theta cycles marked by alternating white and gray boxes. Right: Autocorrelogram of the cell’s spike train from the full recording session, with the red line a cosine wave fit. (B) Example epochs of decoded sweep direction in tmDCs (arrows) and in grid cells (patches) shown across 13 successive theta cycles, generated from data reported by Vollan et al. (4). (C) Left: Distribution of the grid spacing at different recording sites along the dorsoventral axis from one recording session, with each dot representing a grid cell. Middle: Distribution of the HDMVL calculated for each tmDC from the same session. Right: Distribution of the TMI calculated for each tmDC from the same session. (D) Polar histograms of decoded sweep angles in tmDCs for each assigned module, with each row corresponding to one module and each column to data from one animal. (E) Internal direction sweep angle across the three assigned modules. Each dot represents the sweep angle decoded from one recording session (17 sessions from 10 animals). Gray lines link the mean sweep angles from assigned modules in the same recording session. An LMM yielded χ2(2) = 30.7, P = 2.1 × 10−7. Averaged sweep angles were module 1, 19.4°; module 2, 27.6°; and module 3, 34.5°. Module 2 versus module 1: β = 8.2, P = 1.5 × 10−6. Module 3 versus module 2: β = 6.2, P = 0.006. *P < 0.05, **P < 0.01, ***P < 0.001. First, we examined whether tmDCs are organized into modules, as reported for grid cells (20). We used two criteria to assess this effect: the head direction mean vector length (HDMVL) and theta modulation index (TMI; see Materials and Methods). Neither of these measurements show any discontinuity along the dorsoventral axis, unlike the grid scale (see Fig. 2C for an example from one animal and fig. S4 for more animals). Therefore, to perform population decoding of tmDCs at different locations along the dorsoventral axis, we assigned them into different “modules” according to their nearby grid cells. Specifically, for each recording, we trained a logistic regressor to predict the grid module identity of each grid cell from its dorsoventral recording position (see Materials and Methods). Each tmDC was then assigned to a grid module by predicting the module identity from its recording position. This results in 2 to 251 tmDCs in each assigned module (median = 63; IQR: [36 to 107]). At least two assigned modules per session were required for within-animal comparison of directional sweep angles. Next, we decoded the internal direction from the PV of tmDCs within each assigned module. Modules were included only if they contained at least 20 tmDCs, yielding 42 modules across 17 sessions from 10 animals (see Materials and Methods). To compute a single directional sweep angle for each module, we identified the theta phase at which tmDC PV peaked, extracted the corresponding internal directions across theta cycles, and computed their circular mean. We found a systematic increase in the directional sweep angle across assigned modules along the dorsoventral axis [Fig. 2, D and E; an LMM with χ2(2) = 30.7, P = 2.1 × 10−7; internal direction sweep angles were module 1, 19.4°; module 2, 27.6°; and module 3, 34.5°. Module 2 versus module 1: β = 8.2, P = 1.5 × 10−6. Module 3 versus module 2: β = 6.2, P = 0.006]. To control for potential differences in cell numbers, we subsampled each module to match the number of tmDCs used for decoding, and the results still showed a substantial modular effect and revealed a systematic increase in sweep angles along the parasubicular/mEC dorsoventral axis [LMM with χ2(2) = 27.5, P = 1.1 × 10−6; fig. S1B]. For the analysis above, the same theta phase window was used to extract directional sweep angle across all assigned modules, which corresponds to the center (180° phase) of each LFP theta cycle. However, we also observed that tmDCs within each assigned module fired at slightly different phases of the global reference theta cycle––dorsal cells at earlier phases and ventral cells at later phases (fig. S5)––in line with the gradual theta phase shift along the mEC dorsoventral axis (34). To account for this variation, we extracted the directional sweep angle for each assigned module using the phase window corresponding to its own peak activity within the theta cycle. The dorsoventral gradient in directional sweep angles remained [LMM with χ2(2) = 25.3, P = 3.2 × 10−6; fig. S5]. Topography of directional sweeps reflects topography of single-cell firing features in tmDCs The topography of directional sweeps in tmDCs mirrors corresponding gradients in single-cell firing properties along the dorsoventral axis. First, because of the left-right alternation of directional theta sweeps (4), a tmDC can be activated even when the animal’s head is not pointing toward its preferred firing direction, which broadens its directional tuning field and reduces the MVL (Fig. 3A). As sweep angles are larger ventrally, this reduction should be stronger in ventral tmDCs. Consistent with this prediction, we observed a progressive decrease in HDMVL across assigned modules along the dorsoventral axis [Fig. 3B; see also Fig. 2C; an LMM yielded χ2(2) = 415.0, P = 7.6 × 10−91; HDMVL values were module 1, 0.574; module 2, 0.482; and module 3, 0.461]. This provides a mechanistic explanation for the reported decrease in HDMVL along the dorsoventral axis (25). Apart from assessing head direction tuning (plotting firing rate against the head direction), we also examined internal direction tuning (plotting firing rate against the internal direction from the assigned module the tmDC belongs to) (Fig. 3A). The internal direction mean vector length (IDMVL) did not show a gradient [Fig. 3C; LMM with χ2(2) = 2.60, P = 0.27]. This supports the idea that tmDC activity is driven more strongly by internal direction than by head direction and that the broadening in head direction tuning reflects the increasing difference between the two. Topography of single-cell firing features in tmDCs along the parasubicular-mEC dorsoventral axis. (A) Three example tmDCs from each of the three assigned modules in the same recording session. Column 1: directional tuning fields relative to the head direction. Column 2: directional tuning fields relative to the decoded internal direction. Column 3: segments of cells’ spike trains overlaid on the local field potential, with theta cycles marked by alternating white and gray boxes. Column 4: autocorrelograms of the cells’ spike trains from the full recording session, with each red line a cosine wave fit. TSI stands for the theta skipping index (see Materials and Methods). (B) HDMVL across assigned modules. Each dot represents the mean HDMVL from all tmDCs in one assigned module (42 modules from 17 sessions). Gray lines link the mean HDMVL values from assigned modules in the same recording session. An LMM yielded χ2(2) = 432.6, P = 1.2 × 10−94. Mean HDMVL values were module 1, 0.574; module 2, 0.482; and module 3, 0.461. Module 2 versus module 1: β = 0.086, P = 3.35 × 10−63. Module 3 versus module 2: β = 0.018, P = 8.87.6 × 10−4. (C) IDMVL across assigned modules [LMM with χ2(2) = 2.60, P = 0.27]. (D) TSI across assigned modules. An LMM yielded χ2(2) = 629.3, P = 2.2 × 10−137. Mean TSI values were module 1, 0.048; module 2, 0.172; and module 3, 0.232. Module 2 versus module 1: β = 0.097, P = 4.89 × 10−54. Module 3 versus module 2: β = 0.071, P = 3.52 × 10−27. In (B to D), n.s. P > 0.05, *P < 0.05, **P < 0.01, ***P < 0.001. Second, because theta sweeps in tmDCs alternate from side to side of the head direction across successive theta cycles, especially during straight runs (4, 16), this predicts that a tmDC with its preferred firing direction on one side of the current head direction should be activated on every other theta cycle—a phenomenon known as theta skipping (Fig. 3A) (8–11). As sweep angles are larger ventrally, tmDCs in the ventral part should display a stronger theta skipping effect. Consistent with this prediction, we observed an increase in the theta skipping index (TSI; see Materials and Methods) across assigned modules along the dorsoventral axis [Fig. 3D; LMM with χ2(2) = 629.3, P = 2.2 × 10−137; mean TSI values were module 1, 0.048; module 2, 0.172; and module 3, 0.232]. Together, these single-cell properties converge to demonstrate a dorsoventral organization of directional theta sweeps in tmDCs. Dorsoventral increase in FRA can explain the gradient of theta sweeps in a coupled continuous attractor model Left-right sweeps in grid cells can be modeled using a two-layer continuous attractor network (CAN; Fig. 4A) (16), following the circuit described by Vollan et al. (4). The model comprises a ring attractor for tmDCs and a two-dimensional (2D) CAN for grid cells (see Materials and Methods). Introducing FRA into the model causes the activity bump in the tmDC ring attractor—representing the animal’s internal estimate of heading (note that, in the model, we do not dissociate head direction and movement direction and instead assume that they are aligned)—to oscillate around the true head direction rather than remaining stationary. This behavior arises from a dynamic competition between two opposing forces: sensory input, which anchors and stabilizes the bump at the current head direction, and FRA, which acts as a form of activity-dependent negative feedback. As neurons within the bump remain active, adaptation progressively reduces their firing, weakening the stability of the bump and pushing it away from its current position. The resulting interplay between stabilization and destabilization leads to a continual sweeping motion around the head direction (Fig. 4, C and G). This, in turn, drives the activity bump in the 2D CAN to move outward from the current position along the same internal direction, alternating between left and right (Fig. 4, C and H). Dorsoventral increase in FRA strength explains the gradient of theta sweeps and single-cell firing features in a coupled continuous attractor model of tmDCs and grid cells. (A) Schematic of the computational model, comprising a ring attractor of tmDCs and a 2D attractor of grid cells. (B) FRA strength increases along the dorsoventral axis for 35 mEC cells recorded using patch-clamp electrophysiology (Pearson correlation: r = 0.48, P = 0.003), generated from data reported by Yoshida et al. (26). (C) Three modeling examples where only FRA strength was varied. Column 1: directional sweeps in the tmDC attractor, with the gray line marking head direction. Column 2: locational sweeps in the grid cell attractor, with simulated trajectory marked in dark. Column 3: head direction tuning map of a tmDC. Column 4: autocorrelogram of the cell’s firing. (D) Directional and locational sweep angles as a function of normalized FRA strength in the model (left) and in the data (right). (E) Mean HDMVL as a function of normalized FRA strength in the model and data. (F) Mean TSI as a function of normalized FRA strength in the mode and data. (G) Model prediction of continuous directional sweeps across theta cycles with a straight run. (H) Model prediction of palm tree–like sweep patterns in grid cells, overlaid on the straight trajectory (left) and reoriented with the head direction pointing upward (right). Backward sweeps were excluded. (I) Averaged directional sweeps in each assigned module in empirical data (merged across animals, with the number of sessions indicated in brackets). The dark dashed line marks the head direction (reoriented to zero), and the white dashed line marks the mean directional sweep angle at each theta phase. (J) Examples of palm tree–like locational sweeps decoded from grid cells in empirical data. In vitro whole-cell patch-clamp recordings (26) showed that FRA strength in principal neurons in the mEC increases along the dorsoventral axis, measured as the ratio of the instantaneous firing rate between the first two induced spikes and the last two induced spikes (Fig. 4B). In the model, increasing FRA strength caused the activity bump in the ring attractor to sweep further away from the head direction (Fig. 4C), thereby increasing the sweep angle in both the tmDC ring attractor and the grid cell 2D CAN (Fig. 4D). Furthermore, because sweeps in grid cells are driven by upstream tmDCs, the model predicts larger sweep angles in tmDCs than in grid cells, which we further verified with experimental data (Fig. 4D). The network model also explains single-cell firing features in tmDCs we observed empirically. Specifically, increasing FRA strength increased the directional sweep angle and therefore broadened the head direction tuning curves of tmDCs and reduced their HDMVL values (Fig. 4, C and E). It likewise increased the TSI, reflecting the stronger influence of left-right directional sweeps under higher FRA strength (Fig. 4, C and F). Last, because the tmDC ring attractor is a 1D CAN, we predicted that the activity bump should sweep continuously, rather than flickering between left and right (Fig. 4G). This prediction was verified empirically (Fig. 4I and fig. S3). Because the internal direction gradually deviates from the head direction from the early to the middle phase in each theta cycle, we further predicted a gradual increase in the angle of locational sweeps across the early-to-middle phases of the theta cycle, resulting in a palm tree–like pattern rather than straight locational sweeps (Fig. 4H). This prediction was likewise confirmed empirically (Fig. 4J and fig. S3). In summary, this computational model provides a coherent mechanistic explanation for the empirical findings reported above and generates experimentally testable predictions. Efficient learning of hippocampal spatial maps via theta sweeps with varied angles The varied sweep angles observed here, together with the varied sweep lengths reported by Vollan et al. (4), suggest an efficient mechanism for sampling the surrounding space without requiring extensive behavioral exploration [see also Marshall et al. (31)]. Theta sweeps can activate cells encoding locations that do not lie on the animal’s actual trajectory and can modify their synaptic connections accordingly. On the basis of this property, we propose that theta sweeps facilitate the formation of hippocampal spatial maps, particularly when sweeps with different angles are generated by multiple grid modules. Specifically, we constructed three CAN modules to mimic three grid modules along the mEC dorsoventral axis (Fig. 5A). Theta sweeps generated in these CANs were directed toward the same side of the agent but differed in both length and angle. Neuronal outputs from the CANs projected to a downstream network representing hippocampal CA3, mimicking the direct input from the mEC to CA3 via the perforant path. We assumed that, before learning, the firing fields of CA3 place cells were established and uniformly yet randomly tiled the environment, and synaptic connections have not been formed yet. Synaptic connections were then modified according to STDP (12). In the absence of theta sweeps, place cells whose firing fields did not overlap with the behavioral trajectory remained inactive. In contrast, theta sweeps enabled the sequential activation of place cells whose firing fields were located away from the behavioral trajectory (Fig. 5B). Efficient learning of hippocampal spatial maps via theta sweeps. (A) Schematic of the extended model. Top: From magenta to cyan, three CANs corresponding to three grid modules. Middle: Examples of theta sweeps with varying lengths and angles generated by the three CANs, with moving trajectories shown in dark. Bottom: Place cells predefined in the hippocampal CA3 network without recurrent connections before learning, with each dot representing the field center of a place cell. (B) Schematic of synaptic strengthening through STDP, driven by the sequential activation of two place cells along a sweep path. (C) Moving trajectory (dark) overlaid with the trajectory of the bump center in the CAN (blue). Left: Without theta sweeps. Right: With theta sweeps. (D) Correlation between the learned synaptic connectivity and the ideal synaptic connectivity (characterized by a Gaussian profile that decays with the distance between the place field centers of two cells) as a function of learning time. (E) UMAP of the learned synaptic connectivity of all place cells. Each dot represents the projected connectivity profile of a place cell, with nearby dots indicating similar connectivity patterns. Colors represent the distance between a cell’s firing field center and the center of the environment. Notably, only under theta sweeps with varied angles do the learned connectivity profiles reflect the geometry of the square environment without holes. Because sweep trajectories extended into locations not directly experienced by the agent (Fig. 5C), they activated a larger population of place cells and increased opportunities for synaptic modification. Ideally, place cells with nearby firing fields should develop stronger synaptic connections, whereas place cells with distant firing fields should develop weaker connections, resulting in Gaussian-like connectivity profiles. To quantify the emergence of such connectivity, we measured the correlation between the learned synaptic connectivity and an ideal Gaussian-like connectivity throughout learning. In the absence of theta sweeps, place cells relied solely on behavioral experience to update their synaptic connections, and the correlation with the ideal connectivity increased only gradually. When theta sweeps generated by different CANs shared a fixed sweep angle, the correlation increased more rapidly. However, when theta sweeps from different CANs exhibited varied sweep angles, the correlation increased most rapidly (Fig. 5D). After learning, we applied UMAP to visualize the connectivity profile of each neuron by reducing the dimensionality of the connectivity vectors from N dimensions to two dimensions, where N denotes the number of CA3 place cells. We found that, only under the multiangle theta sweep condition, the learned synaptic connectivity accurately reflect the spatial geometry of the environment (Fig. 5E). In summary, our modeling results suggest that theta sweeps generated by multiple grid modules can simultaneously extend toward multiple locations in the environment without requiring extensive behavioral sampling. Consequently, place cells whose firing fields overlap with sweep trajectories can be activated and their synaptic connections efficiently modified. Through this process, a spatial map of the environment can be rapidly learned and encoded in the synaptic connectivity of the hippocampal network. DISCUSSION This study provides empirical evidence for a topographical organization of theta sweep angles and a potential mechanism underlying this organization. Although topographical gradients in single-cell tuning in the mEC have been reported, e.g., grid spacing (20, 21) and directional tuning breadth (25), topographical gradients in population-level spatial representations, such as theta sequences, have not yet been described. Using a large-scale dataset (4) together with computational modeling, we showed that (i) theta sweep angle increases dorsoventrally across mEC grid modules, thereby allowing different modules to represent multiple potential future locations simultaneously at late phases of theta; (ii) this increase coexists with a corresponding dorsoventral increase in directional sweep angles decoded from tmDCs; (iii) the gradient of theta sweeps accounts for the decrease in directional tuning and the increase in theta cycle skipping effect in tmDCs; (iv) continuous directional sweeps drive locational sweeps, which exhibits a palm tree–like pattern; and (v) these features can be unified within an attractor model in which FRA increases along the mEC dorsoventral axis. The computational model provides mechanistic explanations for empirical results we observed, as well as several experimentally testable predictions. First, it suggested that alternating theta sweeps in tmDCs and theta skipping are two manifestations of the same underlying mechanism, as previously discussed by Brandon et al. (10). Because theta skipping was also observed in tmDCs in the AVN (11, 32), it raises the possibility that alternating directional theta sweeps originate in a region upstream of parasubiculum. It further predicted that, when animals transition from straight running to turning, alternating directional sweeps should transform into forward-directed sweeps: The decoded direction should begin behind the current head direction and propagate toward the future turning direction in each theta cycle (7), analogous to forward-directed theta sweeps observed in place cells during linear track navigation (1). This mechanism produces theta phase precession of turning angle during rotation in single-cell firing of tmDCs, observed recently in the AVN (7). Given the stronger alternating sweep effect in more ventral modules, we would therefore expect the length of forward-directed theta sweeps during turning to also increase dorsoventrally. This, in turn, predicts a topographical organization of theta phase precession in tmDCs during turning. Testing these predictions in tmDCs will be important once the full dataset from Vollan et al. (4) becomes available as the current dataset excludes stationary periods during which turns typically occur. We propose that the topographical organization of sweep angles reflects a gradient of FRA strength along the dorsoventral axis in the mEC (26). This gradient likely reflects changes in biophysical properties of cell firing, specifically a decrease in the amplitude and an increase in the duration of the medium after-hyperpolarization potential (mAHP) along the dorsoventral axis (26). Several questions remain open. First, it is not yet clear whether the principal cells recorded in this region were grid cells. Second, little is known about FRA organization in mEC layer III or the parasubiculum, which contain tmDCs (4, 10, 33). Third, if the FRA hypothesis is correct, the underlying neurophysiological mechanisms that support its gradient along the mEC or parasubicular dorsoventral axis remain to be identified (27, 35). Addressing these issues will be essential for developing a detailed mechanistic understanding of the topographical organization of theta sweep angles and that of the related single-cell firing features reported here. Although theta sweeps can reduce the need for extensive behavioral sampling (e.g., navigating between locations) by virtually traversing space, sweeps with a fixed angle and distance would still cover only a limited portion of the environment. The present observation of simultaneous multiangle theta sweeps—together with recent evidence that sweep length increases along the dorsoventral axis of the mEC (4)—suggests a more efficient sampling strategy (Fig. 5), as well as an efficient gold-directed navigation strategy, as noticed in a previous modeling work (29). Dorsal sweeps are short and predominantly forward directed, providing fine-grained sampling ahead of the animal, whereas ventral sweeps are longer and extend laterally, capturing information to the sides. Operating in parallel, these sweeps could link environmental features across multiple spatial scales and directions, enabling a cognitive map to form more rapidly than would be possible through physical exploration alone. This constitutes a working model for how multiscale theta sweep dynamics may facilitate rapid map formation (Fig. 5). Last, in addition to efficient sampling via stereotypic left-right alternating theta sweeps, theta sweeps in open-field environments can be goal directed; that is, they can propagate from the animal’s position toward future goal locations, thereby providing a neural substrate for online goal-directed spatial planning (17–19). MATERIALS AND METHODS Identifying tmDCs For calculating the TMI for each cell, the theta phase was extracted from the population spiking activity of all units. Specifically, spike times were binned into 10-ms bins and the resulting spike counts were band-pass filtered between 5 and 10 Hz. Principal components analysis (PCA) was applied to the matrix of band-pass-filtered spike counts with the first two principal components (PCs) extracted. The theta phase at time t was then defined as the direction of the projection of the instantaneous PV at time t onto the plane defined by PC1 and PC2. The phase of peak activity is roughly 180° relative to a theta cycle. For each neuron, a theta phase tuning curve was constructed by calculating the spike rate in each theta phase bin and the TMI was then quantified as the MVL of the theta phase tuning curve asMVL=1∑kfk∑k[fkcos(ϕk)]2+∑k[fksin(ϕk)]2(1)where fk is the firing rate in the kth theta phase bin ϕk. For calculating head direction mean Rayleigh vector length, the directional tuning curve relative to head direction was first computed by calculating the spike rate in 60 angular bins of the head direction. Then, HDMVL was calculated as the mean Rayleigh vector length using the firing rate in each head direction angular bin. In our analysis, tmDCs are identified with two conditions: (i) TMI > 0.5; (ii) HDMVL > 0.3. These two values are taken from the processed data in Vollan et al. (4). With these two criteria, we found 27 to 450 tmDCs across 24 sessions in 15 rats (median = 190; IQR: [96 to 235]). Determining modular identity for individual tmDCs Although tmDCs do not show modular organization comparable to grid cells (Fig. 2C), we still wanted to decode directional representation from tmDCs at different recording locations along the dorsoventral axis. Therefore, we assigned module identities to each of the tmDCs by using their nearby grid cells which have been classified into grid modules in the Vollan data. Specifically, for each recording session, we trained a multinomial logistic regression model using the recording shank positions of grid cells as the predictor and their known module identities as the categorical outcome. We then applied the trained model to shank positions of tmDCs to assign their module identities. This resulted in 2 to 251 tmDCs in assigned modules (median = 63; IQR: [36 to 107]). Note that the fourth module, when present, was excluded from all analyses as most recording sessions contained fewer than four grid modules. Decoding directional theta sweeps in tmDCs To decode the sweep angle from tmDCs, we correlated the PV vector from selected tmDCs with the session-averaged PV in each head direction bin (60 bins in total), following the method used by Vollan et al. (4). Because tmDCs were mostly activated around the 180° phase in each theta cycle during straight runs (4) (fig. S5), the time bin corresponding to this phase was used to extract the sweep angle in each theta cycle. Specifically, two inputs were fed into to the PV correlation decoder: first, an N × T matrix (MPV) of temporally smoothed firing rates (Gaussian kernel with σ = 10 ms), with N representing cell numbers and T representing the number of theta cycles; and second, an N × P matrix (reference matrix; MrPV) of session averaged angular tuning curves for head direction, with P representing the directional bins (P = 60). For the tth theta cycle, the Pearson correlation coefficient was calculated between the tth column of MPV and each column of MrPV, resulting in a 1 × P vector representing the correlation coefficients with all angular bins (T × P correlation coefficients in total). The sweep angle in the tth theta cycle was defined as the circular mean of P angular bin values, weighted by the P correlation coefficients matrix (dot product), hence producing a single scalar value representing the decoded sweep angle in each theta cycle. We applied this process to each theta cycle resulting in a 1 × T vector representing all the decoded sweep angles in one recording session. These sweep angles are allocentric, being defined by room-referenced directional tuning curves. To get the egocentric directional sweep angle, we then subtracted the decoded sweep angle from the animal’s head direction at the decoded time bin. Theta cycles for which the preceding egocentric internal direction pointed to the right of the head axis were grouped into the left sweep class, whereas those whose preceding internal direction pointed to the left of the head axis were grouped into the right sweep class. Kernel density estimation was then applied separately to the left sweep and right sweep groups to obtain their respective distributions of egocentric sweep angles (Fig. 2D). The peak of each distribution was extracted to represent the averaged sweep angles for left directional sweeps and right directional sweeps. Then, we took their absolute values and calculated their mean as the averaged directional sweep angle. To calculate an overall averaged sweep angle in one recording session (irrespective of which module tmDCs were assigned to), we used the PV from all tmDCs in the PV correlation decoder (median = 190 tmDCs in each session, IQR: [96 to 235]). To calculate an averaged sweep angle in each assigned module, we used the PV from each assigned module to decode their internal direction sweeps. Modules containing fewer than 20 tmDCs were excluded from analysis to prevent unreliable decoding. This resulted in 42 modules from 17 sessions in 10 animals with at least 2 modules involved in decoding (median = 82 tmDCs per assigned module, IQR: [55 to 110]). Given that tmDCs within each assigned module fired at different phases of the theta cycle (fig. S5), we first summed the spikes from all tmDCs in each module within each phase bin and identified the time window in which spike count was maximal. We then selected the decoded angles occurring within the same time window across all theta cycles and computed their mean. This yielded a sweep angle for each tmDC module, based on its preferred firing phase. Decoding locational theta sweeps in grid cells The locational theta sweeps in grid cells were decoded using the same PV correlation method as described above, with the PV from grid cell activities and the session-averaged spatial tuning maps as inputs to the decoder. The spatial tuning maps of grid cells were computed by binning positions into spatial bins of 2.5 cm by 2.5, calculating firing rates per bin, and smoothing the resulting map with a Gaussian kernel (4). Note that, like most decoding algorithms, this method can only decode locations that the animal has previously visited. However, because the animal’s trajectory provides sufficient spatial coverage of the environment, the spatial tuning maps often densely cover nearly all spatial bins. To calculate overall grid cell sweeps, we calculated the correlation between the PV of all recorded cells and the session-averaged spatial tuning maps (median = 741 cells per session, IQR: [606 to 992]). To calculate grid cell sweep in individual grid module, we conducted the same correlation but with the PV of grid cells in each grid module. Modules with less than 40 grid cells were excluded from the analysis to prevent unreliable decoding. This resulted in 48 modules from 19 sessions in 11 animals with more than 2 modules in decoding (median = 100 grid cells per module, IQR: [69 to 129]). Within each theta cycle, the theta sweep trajectory was defined as the longest decoded trajectory in consecutive time bins, which was truncated to remove folds and maximize the Euclidian distance from the start point to the end point (4). After calculating locational sweeps in every theta cycle, the session-averaged locational sweeps were computed by first interpolating each head-centered sweep trajectory at 50 time points linearly spaced from the beginning to the end of the trajectory. Interpolated sweeps were grouped into those that followed a right sweep or left sweep, before an average sweep was computed for each group by taking the median position at corresponding time points within the sweep (4). The locational theta sweep angle was then calculated as the direction of the vector from the start point to the end point of the session-averaged locational sweep, relative to the head direction, and lastly averaged across the left and right angles. It is noteworthy that, although we observed increasing sweep angles across grid modules, the first grid module detected in each recording session was always designated as module 1, although it may not correspond to the anatomically most dorsal module. This does not violate the assumptions of the LMM because our inference concerns within-animal module effects, and all statistical analyses were performed at the animal level. This is also the case in the analysis of increased internal direction sweep angle from tmDCs. Analyzing directional sweeps and locational sweeps across theta phase bins To show that decoded internal direction sweeps continuously from side to side of the head direction across theta cycles, instead of flickering abruptly across the head direction, as reported by Vollan et al. (4), we checked the decoded probability map as a function of theta phase. To do this, we binned each theta cycle into 20 phase bins (each bin is 18°) and binned the corresponding PV correlation matrix of directional sweep (described above) into 60 angular bins (each bin is 6°). Then, we raised the correlation matrix to the fourth power for better visualization and normalized each column into a probability distribution. This resulted in a 20 × 60 probability matrices (PMs) of decoded direction in each theta cycle. We then averaged the PMs across all theta cycles in which the internal direction on the preceding theta cycle lay on the left side of the head direction. We further averaged these mean PMs across all sessions, separately for modules 1, 2, and 3. This shows a right sweep starting from the head direction at early theta phase (0°), propagating to the right side until the middle theta phase (180°) and then sweeping backward to the head direction until the late theta phase (360°) (fig. S3). This is also true by averaging PMs across all theta cycles in which the internal direction on the preceding theta cycle lay on the right side of the head direction, which gives left sweeps. To demonstrate that the locational sweep angle increases continuously within a theta cycle, we first selected locational sweep angles along the outward sweep trajectory. Only sweeps with 3 to 10 valid points and a path straightness score (computed for each time step of the recording by dividing net travel distance by cumulative travelled distance within a 2-s moving window; high values correspond to straight paths) greater than 0.5, obtained during running speeds above 20 cm/s, were included in further analysis. For the remaining sweeps, we interpolated each sweep trajectory into 10 points and calculated nine angular values of sweep direction, defined by the vectors from the starting sweep point to the 2nd to 10th points, respectively. We then subtracted the head direction to obtain nine egocentric sweep angles. These values were averaged across all selected sweeps within each recording session, separately for grid modules 1, 2, and 3. This yielded session-averaged locational sweep angles as a function of interpolated theta phase bins 1 to 9 (not exact phase bins but close approximations). Last, we averaged across all sessions from all animals, revealing a continuous increase in locational sweep angles across the theta cycle (fig. S3C). Last, to demonstrate that the directional sweep angle drives the continuous increase in locational sweep angle across a theta cycle, we first calculated the histogram of theta phases corresponding to the start points of the selected locational sweeps and identified the first phase bin as the peak of this histogram. Starting from the subsequent phase bin to the first phase bin, we then calculated the directional sweep angles in the subsequent eight consecutive phase bins (18° per bin). These values were averaged across all theta cycles, separately for grid modules 1, 2, and 3. This procedure yielded session-averaged directional sweep angles as a function of theta phase bins that were approximately aligned with those used for the locational sweeps described above. Last, we averaged across all sessions from all animals, revealing a continuous increase in directional sweep angles across the theta cycle (fig. S3D). Calculating the TSI in tmDCs To quantify theta skipping effect in tmDCs, we applied the method used in Brandon et al. (10), fitting the autocorrelogram with a cosine wave of frequency ω and an interfering cosine wave of frequency ω/2y(t)=[a1(cos(2πωt)+1)+a2(cos(πωt)+1)+b]·exp(−∣t∣τ1)+c·exp(−t2τ22)(2)where t is the time lag of the autocorrelogram (in seconds), and a1,a2,b,c,τ1, and τ2 are parameters determined by fitting. These parameters were constrained within the following ranges: a1∈[0,100],a2∈[0,100],b∈[0,100 ],c∈[0,0.8],ω∈[6,12],τ1∈[0,8], and τ2∈[0,0.05]. The interfering cosine wave was used to capture the theta skipping effect on the autocorrelogram. The TSI was then calculated as the difference between the first and second peaks on the fitted curve, normalized by the larger of the twoTS=p2−p1max(p1,p2)(3)where p1 is the model value at one cycle with t = 2π/ω and p2 is the model value at two cycles with t = 4π/ω. This index is bounded between −1 and 1, with higher values indicating a greater degree of theta skipping. Note that the TSI was calculated only on the basis of spike trains during fast running periods with animals’ speed > 15 cm/s because stationary periods were not included in the published dataset. Computational modeling The computational model consists of a two-layer CAN in which tmDCs are modeled as a ring attractor and grid cells are modeled as a 2D attractor on a neuronal sheet [GC attractor; see Ji et al. (16) for more details]. The dynamics of the tmDC ring attractor are defined asτd∂di∂t=−hi+∑j=1NdJijdfd(dj)−aid+Iid(θ)(4)where di is the presynaptic input to ith tmDC (100 simulated cells in total), fd(dj) is the firing rate of the jth tmDC, Jijd denotes Gaussian-like recurrent connections, aid is the FRA term (defined below), and Iid(θ) is the sensory input encoding the current head/movement directionIid(θ)=Adαdexp[−(xi−θ)24bd2](5)where αd represents the theta modulation from the medial septum (see below) and θ is the current head/movement direction. With FRA, the activity bump sweeps from side to side around the current head direction, with the sweep frequency controlled by the induced theta modulation. The GC attractor is defined asτg∂gi∂t=−gi+∑jNgJijgfg(gj)−aig+Iig(β)+Iiloc(β)(6)where gi is the presynaptic input to the ith grid cell (100 × 100 cells in total tiling the 2D neuronal sheet), fg(gj) is the firing rate, Jijg denotes the 2D Gaussian recurrent connectivity, aig is the FRA term (defined below), and β is the location expressed in a phase space (not a physical space). Note that the GC attractor runs in a phase space instead of the physical space, so all the inputs has been converted from physical space to the phase space. Iig(β) is the input from conjunctive grid cells, and Iiloc(β) is the sensory input representing location information. Iig(β) is expressed asIig(β)=αg∑jIconj−g(βj)=αg∑jAgfh(hj)exp[−‖ϕi−ψ−w(βj)‖24bg2](7)where Iconj−g(βj) is the input from the jth conjunctive grid cell group, and αg represents theta modulation (see below). Note that the jth group of conjunctive grid cells will be activated when the jth tmDC is activated in the ring attractor. The total input is a weighted sum of many 2D Gaussian bumps from all the conjunctive grid cells with the weights as the firing rate of upstream tmDCs fh(hj). ψ is the animal’s location in the phase space. w(βj) is a vector with w(βj)=w0(cosβj,sinβj), where w0 represents the offset length, with the offset representing the grid phase offset between putatively connected conjunctive grid to pure grid cells, as shown by Vollan et al. (4). Equation 7 means that activation of conjunctive grid cells corresponding to direction βj generates a shifted Gaussian input centered around ψ+w(βj). The location-dependent input Iiloc(β) is expressed asIig(β)=Agαginvexp[−‖ϕi−ψ‖24bg2](8)where αginv is the theta modulation with an opposite phase compared to αg, mimicking encoding and retrieval occurring at opposite phases of the theta cycle (36). Thus, in early theta phases, the GC attractor is pulled toward the current location ψ, whereas in late phases, it is driven toward the offset input from conjunctive grid cells, generating directional sweeps along the internal direction. Theta modulation of tmDCs isαd=1+α¯dvsin(ωθt)(9)where the oscillation magnitude scales linearly with the angular speed v, with α¯d as a scaling factor. The theta modulation on the GC attractor isαg=1+α¯gvsin(ωθt)(10)where the oscillation magnitude also scales linearly with the running speed v, with α¯g as a scaling factor. FRA has the following dynamicsτa∂as(ϕi)∂t=−as(ϕi)+msfs(si)(11)where as(ϕi) represents FRA in the ith cell with s∈{d,g}. d represents tmDCs, and g represents grid cells. fs(si) represents the firing rate of the ith cell. τa is the time constant of FRA, which is much larger than τd and τg, indicating slow dynamics. ms is the FRA strength, which can differ between tmDCs and grid cells. We systematically varied ms to model the dorsoventral increase in FRA observed empirically (26). For the parameter settings, the simulated model consists of 100 tmDCs (Nd) and 100 × 100 grid cells (Ng) arranged on a 2D neuron sheet. Time constants for both tmDCs (τd) and grid cells (τg) are set to 10 ms, as is typical in rate-based models. For tmDCs, recurrent connections are modeled as Gaussian profiles with an SD (bh) of 0.7 rad. For grid cells, recurrent connections are modeled as 2D Gaussian profiles with an SD (bg) of 0.8, corresponding to ∼13% of the width of the 2D neuronal sheet. The offset value projected from conjunctive grid cells to pure grid cells, w0, is set to 0.1, corresponding to roughly 1.6% of the width of the sheet, which is 2π. The time constant for FRA, τa, is set to a much larger value of 100 ms compared with the time constant τg and τd. This means that it takes ∼100 ms for a neuron’s firing rate to adapt (i.e., decrease) in response to a sustained stimulus. To model the increase in sweep angles along the dorsoventral axis [where adaptation strength increases (26)], the adaptation strength of tmDCs (md) varied between 2 and 20, with 2 corresponding to the most dorsal part and 20 to the most ventral part of the mEC/parasubiculum. For Fig. 4C, the FRA strength was set to 5, 10, and 15 for the three example simulations, respectively. For Fig. 4D, the FRA strength ranged from 10 to 11, whereas for Fig. 4 (E and F), it ranged from 5 to 10, chosen to best match the empirical data. Note that the adaptation strength is dimensionless, and its absolute value is not important. Rather, the key result is that increasing FRA strength explains the topographical gradient in theta sweeps. Although it reflects a slow negative feedback inhibition proportional to the instantaneous firing rate, its numerical range is not directly comparable to that reported by Yoshida et al. (26). Specifically, in Yoshida et al., FRA strength was quantified from the instantaneous firing frequency by fitting the function A+Be−2t/τ to the spike train, where B represents the amplitude of the exponential component and is plotted on the y axis, yielding values in the range of ∼10 to 60 (see Fig. 4B). Statistical analysis All statistical analyses were performed in Python (version 3.12). Statistical analyses in Figs. 1 to 3 and figs. S1 and S4 were performed using the statsmodels package (MixedLM implementation). Statistical significance was defined as α = 0.05, and all tests were two sided. Exact P values, test statistics, and sample sizes are reported in the main text or corresponding figure captions. To account for repeated measurements across animals, LMMs were fitted with sweep angle as the dependent variable and module identity as a categorical fixed effect. Animal identity was included as a random intercept to account for clustering of observations within animals. Models were fitted using maximum likelihood estimation when comparing nested models and restricted maximum likelihood otherwise. The overall effect of the module was assessed by comparing the full model to a null model containing only the intercept and random effect, using a likelihood ratio test. The test statistic was computed as twice the difference in log-likelihoods and evaluated against a χ2 distribution with degrees of freedom equal to the difference in the number of fixed-effect parameters. Pairwise comparisons between modules were obtained by refitting the model with alternative reference levels, and regression coefficients (β), χ2 statistics, and exact P values are reported. Pearson correlation in Fig. 4B coefficients were computed using the scipy.stats module in Python. Correlation strength (r) and corresponding two-tailed P values were calculated to assess linear associations between variables. Statistical significance was defined as α = 0.05. Sample size, correlation coefficients, and exact P values are reported in the main text or figure legends.