Report

Vectorial representation of spatial goals in the hippocampus of bats

See allHide authors and affiliations

Science  13 Jan 2017:
Vol. 355, Issue 6321, pp. 176-180
DOI: 10.1126/science.aak9589

How to get to place B

We constantly navigate around our environment. This means moving from our current location, place A, to a new goal, place B. We have recently learned much about spatial maps in the brain in which place cells indicate current location. However, it is unclear how navigational goals are represented in the brain. Sarel et al. describe a group of neurons in the brains of bats that are tuned to goal direction and distance relative to the bat's current position as it flies toward its goal. The finding elucidates the computations involved in spatial navigation.

Science, this issue p. 176

Abstract

To navigate, animals need to represent not only their own position and orientation, but also the location of their goal. Neural representations of an animal’s own position and orientation have been extensively studied. However, it is unknown how navigational goals are encoded in the brain. We recorded from hippocampal CA1 neurons of bats flying in complex trajectories toward a spatial goal. We discovered a subpopulation of neurons with angular tuning to the goal direction. Many of these neurons were tuned to an occluded goal, suggesting that goal-direction representation is memory-based. We also found cells that encoded the distance to the goal, often in conjunction with goal direction. The goal-direction and goal-distance signals make up a vectorial representation of spatial goals, suggesting a previously unrecognized neuronal mechanism for goal-directed navigation.

Navigation, the ability to reach a desired goal location, is essential for humans and animals. Decades of research have focused on the neural representation of an animal’s own location and orientation, revealing the existence of place cells (13), grid cells (4, 5), and head-direction cells (6, 7). However, a fundamental question that remains unanswered is how an animal’s spatial goals are encoded in the brain. One suggested mechanism posits that the goal is represented by activating a sequence of place cells from the animal’s current location to the goal location (8, 9). An alternative mechanism would be to rely on a vectorial representation of the goal—i.e., encoding the direction and distance to the goal. This mechanism has been suggested both theoretically (1013) and by behavioral studies in a variety of species (1419); however, a representation of such vectors in the brain has not been found to date.

We trained Egyptian fruit bats to fly in highly complex trajectories within a large flight room and land on a single landing site, defined as the goal, where the bat could eat and rest (Fig. 1A and fig. S1). We computed the egocentric azimuthal goal-direction angle, defined as the heading direction of the bat with respect to the goal (Fig. 1B) (we focused all our analyses on the two-dimensional horizontal projection, because the bats’ behavior was mostly confined to a narrow horizontal slab around the z height of the goal; fig. S2). While the bats performed this goal-directed task and sampled all goal-direction angles (Fig. 1C), we recorded the activity of 309 single neurons from hippocampal area CA1 of three bats (Fig. 1D), using a wireless electrophysiology device (20). A subpopulation of hippocampal neurons exhibited angular tuning to the egocentric goal direction [Fig. 1, E (three leftmost examples, top row) and F]. We classified 19% of the CA1 neurons (58 of 309) as significant goal-direction cells on the basis of tuning directionality (95th percentile in a shuffling analysis), tuning stability, and tuning reconstruction analysis (20, 21). The distribution of preferred goal-direction angles across these cells spanned the entire 360° range (Fig. 1, F and G) but exhibited overrepresentation of preferred goal-direction angle 0° (Fig. 1, F and G; see also Fig. 1E, cells 213 and 287)—i.e., a substantial fraction of cells fired maximally when the bat was heading toward the goal.

Fig. 1 Goal-direction tuning in bat hippocampal area CA1.

(A) Behavioral setup, showing the flight room (5.8 by 4.6 by 2.7 m; top view) with one elevated landing point (the goal; circle). Five flight trajectories are highlighted in different colors on top of the behavioral coverage for that day (gray). Cameras are shown in the lower corners. (B) Goal-direction angle (blue), defined as the azimuthal angle between the heading direction (top arrow) and the bat-to-goal direction (bottom arrow). (C) Distribution of time spent by all bats in different goal-direction angles. (D) Coronal section through the dorsal hippocampus of one of the bats. Arrowhead, lesion at end of the tetrode track. (E) Four example cells (columns). Top, goal-direction tuning curves. Middle, goal-direction angles along the behavioral session (gray), with spikes overlaid (red). Bottom, spatial firing-rate maps [top view; color scale, zero (blue) to maximal firing rate (red, value indicated)]. Cell 213, goal-direction cell without place tuning; cells 287 and 131, tuned to both goal direction and place; cell 77, pure (classical) place cell. (F) Normalized goal-direction tuning for all significant goal-direction cells (n = 58; rows, sorted by preferred direction). (G) Distribution of preferred goal directions. (H) Total numbers of recorded goal-direction and place cells in CA1. (I) Distribution of the goal/place index for all goal-direction cells (blue; n = 58) and for simulated pure place cells (pink; n = 5625). Goal-direction tuning could not be explained by pure place tuning (P < 2 × 10−4 for all 58 goal-direction cells; vertical lines, 99th and 100th percentile for the simulated cells). (J) Examples of in-field/out-of-field analysis for three cells (columns). Top, goal-direction tuning curves for in-field (dark purple) and out-of-field (light blue) data. Bottom, firing-rate maps (top view; gray line, in-field area). (K) Distribution of differences in preferred goal direction (ΔGD) between in-field and out-of-field tuning curves. (L) Two example cells, showing the normalized goal-direction tuning curves (rows) for different times along the complex flights. (M) Stability of goal-direction tuning (n = 58 goal-direction cells). Shown are correlations of tuning curves between the first and second halves of the flight (short-term stability, left) and between the first and second halves of the behavioral session (long-term stability, right).

Of the 58 goal-direction cells, 26 neurons (45%) exhibited angular tuning to the goal with no significant place tuning (e.g., Fig. 1E, cell 213), whereas 32 neurons (55%) showed both goal-direction tuning and place tuning (e.g., Fig. 1E, cells 287 and 131; summary in Fig. 1H). We also found 101 classical place cells, making up 33% of the recorded CA1 neurons, a similar fraction to that found in previous studies of place cells in rats and bats (2, 3). Most of the place cells (69 cells, 68%) had no goal-direction tuning (e.g., Fig. 1E, cell 77; Fig. 1H), consistent with the classical hippocampal place code.

To verify that the goal-direction signal is genuine and not biased by the place tuning, we did the following. First, one of our inclusion criteria for goal-direction cells required the neuron to have stronger goal-direction tuning than place tuning, based on a reconstruction analysis, yielding a goal/place index > 1 (Fig. 1I and fig. S3) (20). This very strict criterion ensured that the goal-direction tuning of all the neurons that we studied could not be explained through coupling of pure place tuning and the bat’s behavior (P < 2 × 10−4 for all of the 58 goal-direction cells; Fig. 1I and fig. S3). Second, almost half of the goal-direction cells were not significantly place-tuned (e.g., Fig. 1E, cell 213; Fig. 1H; n = 26). Third, for cells that were tuned to goal direction and place, the firing within the place field was highly reduced when the bat was flying in the null direction (180° from the preferred goal direction) (fig. S4). Fourth, we computed the goal-direction tuning separately inside and outside the place field (Fig. 1J) (20). Many of the goal-direction cells exhibited similar preferred goal directions inside and outside the place field (Fig. 1, J and K). Next, we repeated the same types of analyses to dissociate between the goal-direction signal and the head-direction signal reported in the hippocampus (21, 22); we found that the goal-direction signal was largely independent of the head-direction signal (fig. S5). Furthermore, 31% of the goal-direction cells (18 of 58) were significantly tuned only to goal direction but not to head direction or to place.

The neurons stably maintained their goal-direction tuning throughout the flight to the goal [Fig. 1, L and M (left panel)], including well before landing, despite the fact that the bats’ flights were long and complex and spanned highly variable angles (Fig. 1, A and C, and fig. S1). The goal-direction tuning was also stable along the entire behavioral session [Fig. 1, E (three leftmost examples; note the stable raster of spikes along the session) and M (right)]. To verify that the tuning was specific to the goal, we computed the tuning to every location (every pixel) in the room as though it was a goal. The goal-direction cells were sharply tuned to the true goal, but not to other locations in the environment (fig. S6).

During real-life navigation, the goal could be invisible to the animal, meaning that goal-directed navigation requires memory (23). We therefore conducted another session at the beginning of every recording day, in which we occluded the goal by an opaque curtain that blocked vision, echolocation, and olfaction (Fig. 2A, top). This hidden-goal session was conducted with two of the three bats (n = 158 recorded cells). Tuning to the hidden goal was computed using only epochs when bats could not see the goal. A substantial fraction of cells (43 of 158, or 27%) exhibited significant directional tuning to the hidden goal [Fig. 2B (examples, top row) and fig. S7]. These cells were tuned significantly more sharply to the hidden goal than to the curtain edges, as quantified by comparing the directionality index [the Rayleigh vector length (20); fig. S8, A to C; t test, P < 0.03]. The goal-direction tuning was invariant to flight trajectory—i.e., the tuning did not change if the bat eventually flew to the hidden goal from the left or right side of the curtain (fig. S8, E to H)—suggesting trajectory-invariant representation of hidden goals “through the wall.”

Fig. 2 Goal-direction tuning is memory-based.

(A) Behavioral setups (top view) for two consecutive sessions. In each session, there was only one goal (black filled circle), located either behind an opaque curtain (top, session 1, hidden goal session) or at the center of the room (bottom, session 2, central goal session). Black horizontal line, curtain position (fixed across sessions and days). Blue and red circles, positions for which we computed the tuning curves in (B) to (D); dotted circles are empty locations (no goal). In this experiment, we recorded 158 cells from two bats. (B to D) Examples of goal-direction cells (columns) for different sessions (rows). Red and blue tuning curves were computed for the red and blue circles in (A). Dotted tuning curves [computed for dotted circles in (A)] indicate tuning to empty locations. All tuning curves were computed on the basis of epochs when the bat could not see the hidden goal. Four examples of cells tuned to the hidden goal in session 1 are shown in (B); the three leftmost cells lost their tuning in session 2. Shown in (C) is an example of a neuron tuned to the central goal only in session 2. Examples of goal-direction cells recorded for three sessions (hidden session → central session → hidden session) are shown in (D). Cell 269 was tuned to the hidden goal in sessions 1 and 3; the other examples (blue) were tuned to the central goal only in session 2. (E) Distributions of changes in tuning modulation depth, comparing the tuning to the same location with versus without a goal. Top, cells tuned to the central goal (n = 31); bottom, cells tuned to the hidden goal (n = 43). The tuning changes were highly significant (t test; top, P < 0.005; bottom, P < 10−5). (F and G) Three example neurons, showing the dynamics of goal-direction tuning over the sessions. Colors are scaled for each neuron, from zero (blue) to the maximal firing rate across all sessions (red, value indicated). Two neurons that changed their tuning abruptly at the transition point between sessions are shown in (F). Shown in (G) is a neuron that was tuned to the hidden goal in the first session and exhibited a gradual decay in its tuning when the goal was moved.

We next examined the effect of changing the goal location. In the second session of the curtain experiment, we moved the goal to the center of the room (Fig. 2A, bottom) (i.e., the original experiment shown in Fig. 1A). Of 43 cells tuned to the hidden goal, 27 (63%) lost this tuning when the goal was not there anymore (Fig. 2B, three leftmost examples, bottom row; cell 207 exemplifies the 37% of neurons that did not lose their tuning). Similarly, 21 (68%) of the 31 cells that were tuned to the centrally located goal in the second session were untuned when the goal was not there (Fig. 2C). These extreme changes in tuning (Fig. 2, B and C) could not be explained by behavioral changes between the sessions, recording instability, or selection bias (fig. S9). For a subset of neurons, we recorded three sessions (hidden goal → central goal → hidden goal) and found that the goal-direction tuning was conserved across the two hidden-goal sessions (Fig. 2D; compare sessions 1 and 3). Among the population of goal-direction neurons recorded in the curtain experiments, most cells were significantly more tuned to the goal than to the same spatial position when the goal was not there—as was quantified by a change in the tuning modulation depth index (20) (Fig. 2E; t test for cells tuned to the hidden goal, P < 10−5; t test for cells tuned to the central goal, P < 0.005)—suggesting that the tuning is indeed goal-related. Further, the majority of goal-direction cells (60 of 74, or 81%) were significantly tuned to only one of the two goals, either the hidden goal in session 1 or the central goal in session 2, indicating that the goal-direction tuning is largely goal-specific.

To examine the tuning changes between the sessions at better temporal resolution, we computed the goal-direction tuning over the course of the sessions in short time bins (Fig. 2, F and G). Some of the cells abruptly changed their tuning when the goal moved to the new location (Fig. 2F). A handful of cells that were tuned to the hidden goal in the first session showed a slow decay in the tuning over the ~10 to 15 min after we moved the goal, which might indicate a memory trace for the previous goal location (Fig. 2G).

In addition to knowing the goal direction, it could be useful for the animal to know the distance to the goal. Sixteen percent of the CA1 neurons (49 of 309) were modulated by the path distance to the goal (Fig. 3, A and B, and fig. S10) (20). Most of these goal-distance cells fired maximally at short path distances of between 0 and –2 m, i.e., when the bat was approaching the goal [Fig. 3, A (four leftmost examples) and B]. These neurons could be interpreted as signaling the expectation of imminent reward, which is difficult to disentangle experimentally from coding of distance to goal per se, because for a flying bat, it is always rewarding to land on a goal to rest. However, almost half of the cells fired well before landing on the goal and had preferred path distances of <–2 m, and for some cells even as far as 10 m [Fig. 3, A (two rightmost examples) and B, and fig. S10 (right)]—suggesting that, at least for these neurons, the distance tuning did not reflect the bat’s expectation of reward.

Fig. 3 Encoding of distance to the goal.

(A) Six example neurons exhibiting tuning to path distance, sorted by preferred distance. Top, tuning curves (firing rate versus distance), computed from the rasters below (normalized by time spent at each distance). Bottom, raster plots of the last 12 m of path distance, aligned to landing (distance 0); black dots, spikes. Flights had different lengths and were sorted here by path distance; gray lines indicate flight start. (B) Tuning curves of all significant goal-distance cells (n = 49; sorted by preferred distance). Left, normalized distance-tuning curves for all cells (rows). Right, depiction of tuning widths (at half-height, lines) and preferred distances (dots). (C) Directionality index versus path-distance index for all goal-distance cells (Pearson correlation, r = 0.65, P < 10−6; n = 49). (D) Tuning curves for two example cells conjunctively encoding goal direction and path distance to the goal. Plots depict firing rate as function of goal direction and path distance. (E) Total numbers of different functional cell types recorded in CA1. (F) Tuning curves for two example cells conjunctively encoding goal direction and Euclidean distance to the goal.

We next examined the relation between distance tuning and goal-direction tuning and found a significant correlation between the two: Neurons that were more strongly tuned to path distance were also more strongly tuned to goal direction [Fig. 3C; Pearson correlation between directionality index and path-distance index (20), r = 0.65, P < 10−6]. This suggested a conjunctive representation of goal direction and distance, which was indeed substantiated by plotting the goal-direction tuning for different distances to the goal (Fig. 3D). Some of these cells were also tuned conjunctively to place (Fig. 3E and fig. S11); however, the path-distance tuning could not be explained by the place tuning (fig. S12).

Cells were more tuned to path distance than to time to landing (fig. S13), although both distance to goal and time to goal could be useful signals for navigation [with timing information possibly represented by time cells (24)]. The goal-distance cells were significantly more tuned to path distance than to Euclidean distance to the goal (fig. S14, A to D). During complex nonstraight flights, as in our experiment—or when flying in the wild through interconnected cave systems—the encoding of path distance may be particularly useful for guiding navigation. In contrast, during long-distance natural navigation, bats are known to fly in very straight lines (25), in which case the path distance would be similar to the Euclidean distance to the goal. Under the latter circumstances, neurons tuned to path distance and direction (Fig. 3D) might carry information about the length and angle of the vector to the goal. Further, we found that a subset of cells explicitly encoded the Euclidean distance to the goal (fig. S14E), often in conjunction with goal direction (Fig. 3F)—forming a full vectorial representation of the goal.

In summary, we report here on neurons that encode the egocentric direction (Fig. 1) and distance (Fig. 3) to navigational goals. We found these neurons in hippocampal area CA1, indicating that the hippocampus contains a variety of navigation-related signals, including a place signal, a head-direction signal, and a vectorial representation of goals. The goal-direction and goal-distance tuning could not be explained by the classical CA1 place tuning (Fig. 1 and figs. S3, S4, and S12) or by the head-direction tuning (fig. S5). Notably, goal-direction neurons were able to represent occluded goals, suggesting that goal-direction tuning is memory-based rather than sensory-based (Fig. 2). Together, our data demonstrate genuine goal-direction and goal-distance signals in the hippocampus, constituting a vectorial representation of spatial goals.

The goal-direction cells are very different from hippocampal neurons reported previously in rats that fire on one side of a landmark but that are not affected by the animal’s orientation relative to the landmark (26). In addition, the goal-direction signal was strongly affected when we moved the goal (Fig. 2); it is therefore goal-specific and conceptually different from spatial-view cells reported previously in monkeys, which are tuned to an abstract point on the wall (27). The goal-direction signal is memory-based (Fig. 2) and is thus different from cells found in the rat parietal cortex, which respond to the direction of visible cues (28).

We argue that the mammalian navigation system, which contains the major spatially tuned classes of cells in the brain (1, 4, 6), is incomplete without a representation of the goal location to which the animal is navigating. Our finding of a vectorial representation of goals in the hippocampus might fill this gap. The preferred goal directions spanned the entire range of 360° (Fig. 1, F and G), and therefore the goal-direction signal can allow the bat to decode its deviation from the goal (e.g., during navigational detours) and thus to reach its goal; the additional presence of a distance signal may enable even more sophisticated computations. Our results are consistent with a recent functional magnetic resonance imaging study in humans, which reported hippocampal activation related to a combination of goal direction and path distance to the goal (29); importantly, here we demonstrate single neurons encoding goals in a vectorial manner in egocentric coordinates. We speculate that the hippocampal representation of hidden goals reported here might provide an unexpected explanation of why hippocampal lesions impair navigation to hidden goals (e.g., in a water maze) (30). Last, we propose that this neural representation could underlie the vector-based navigation strategies described for many species, from insects to humans (1419)—suggesting a previously unrecognized mechanism for goal-directed navigation across species.

Supplementary Materials

www.sciencemag.org/content/355/6321/176/suppl/DC1

Materials and Methods

Figs. S1 to S14

References (3234)

References and Notes

  1. Materials and methods are available as supplementary materials.
  2. Acknowledgments: We thank S. Romani, A. Treves, A. Rubin, T. Stolero, D. Omer, T. Eliav, G. Ginosar, D. Blum, and S. Maimon for comments on the manuscript; O. Gobi and S. Kaufman for bat training; A. Tuval for veterinary support; C. Ra’anan and R. Eilam for histology; B. Pasmantirer and G. Ankaoua for mechanical designs; and G. Brodsky and H. Avital for graphics. This study was supported by research grants to N.U. from the European Research Council (ERC-StG NEUROBAT and ERC-CoG NATURAL_BAT_NAV), the Israel Science Foundation (ISF 1319/13), and the Minerva Foundation. Equipment support was provided by the Krenter Institute at the Weizmann Institute of Science. The data are archived on the Weizmann Institute of Science servers and will be made available on request. The results reported in this study were presented earlier in abstract form (31).
View Abstract

Stay Connected to Science

Navigate This Article