Medial Prefrontal Cortex Represents the Object-Based Cognitive Map When Remembering an Egocentric Target Location

Abstract

A cognitive map, representing an environment around oneself, is necessary for spatial navigation. However, compared with its constituent elements such as individual landmarks, neural substrates of coherent spatial information, which consists in a relationship among the individual elements, remain largely unknown. The present study investigated how the brain codes map-like representations in a virtual environment specified by the relative positions of three objects. Representational similarity analysis revealed an object-based spatial representation in the hippocampus (HPC) when participants located themselves within the environment, while the medial prefrontal cortex (mPFC) represented it when they recollected a target object’s location relative to their self-body. During recollection, task-dependent functional connectivity increased between the two areas implying exchange of self-location and target location signals between the HPC and mPFC. Together, the object-based cognitive map, whose coherent spatial information could be formed by objects, may be recruited in the HPC and mPFC for complementary functions during navigation, which may generalize to other aspects of cognition, such as navigating social interactions.

cognitive map, episodic memory, hippocampus, medial prefrontal cortex, representational similarity analysis

Introduction

During navigation, it is necessary to locate our self-position in the current spatial environment as well as to locate the objects relative to the self-body (i.e., egocentric location). To conduct each of the two mental operations, we need map-like representations, called “cognitive map” in our brain (Tolman 1948). After the discovery of “place cells,” the hippocampus (HPC) of the medial temporal lobe (MTL) has been considered responsible for the cognitive map (Buffalo 2015), and crucial contributions of the HPC to spatial memory have also been reported by animal model studies that evaluated behavioral patterns of rodents with an inactivated HPC using maze tasks (Packard and McGaugh 1996; Redish and Touretzky 1998; Nakazawa et al. 2002) as well as human studies that demonstrated the relationship between HPC volume in individual subjects and their amounts of experience exploring spatial environments (Woollett and Maguire 2011; Schinazi et al. 2013, e.g., London taxi drivers). However, it remains largely unknown how neural substrates of the cognitive map are involved in the two mental operations required to locate specific objects within the environment. One possible reason for the difficulty in addressing this question is that despite extensive studies on the spatial elements related to the cognitive map (e.g., self-location, head direction, etc.) (O'Keefe and Dostrovsky 1971; Vass and Epstein 2013; Buffalo 2015; Chadwick et al. 2015; McCormick et al. 2018), there is still a lack of sufficient isolation and characterization of the neural signal of the cognitive map under the previous research paradigms.

In addition to the HPC, the role of the medial prefrontal cortex (mPFC) in goal-directed planning during navigation was demonstrated by a previous human functional magnetic resonance imaging (fMRI) study that showed increased connectivity between the HPC and mPFC (Brown et al. 2016). The mPFC has been long considered as a member of the core-brain system in the retrieval of episodic memory (Konishi et al. 2000; Eichenbaum 2017; McCormick et al. 2018), which is an autobiographical memory consisting of spatial, object, and temporal information (Naya and Suzuki 2011; Squire and Wixted 2011; Suzuki and Naya 2011). Schacter et al. (2007) suggested an involvement of the mPFC in future simulation processing and recollection of past episodes, which depend on mnemonic information stored as declarative memory including both episodic and semantic memory. Recently, they also showed increased connectivity between the HPC and mPFC during future simulation (Campbell et al. 2018). This preceding literature suggests that the HPC and mPFC, which belong to the default-mode network, work together when remembering stored information (e.g., cognitive map) and construct the mental representation of goal-directed information (e.g., target location) from mnemonic information with the current context (e.g., self-location) (Schacter 2012). However, the specific functional role of each of HPC and mPFC during the construction process (Campbell et al. 2018; McCormick et al. 2018) remains elusive, presumably because the construction of goal-directed information (e.g., spatial navigation) includes at least two mental operations described above (locating the self and locating an object target relative to self-location), as previous experimental paradigms did not dissociate these aspects of behavior.

To address these issues, we aimed to devise a novel 3D spatial memory task with spatial environments defined by three different human characters, which would enable us to identify the representation of the cognitive map and to investigate how it is related to the two mental operations (Fig. 1). The spatial configuration pattern of the characters was referred to as a “map” in the present study (Fig. 1b). Using representational similarity analysis (RSA; see Methods for details) (Kriegeskorte et al. 2006, 2008a, 2008b), we investigated the brain regions that code the map when the participants located themselves in the virtual environment (facing period) and when they recalled an egocentric location of a target character (targeting period).

Figure 1

Task design. (a) Spatial memory task. Each trial consisted of three periods. In the walking period, participants walked toward three human characters using the first-person perspective and stopped on a wood plate. In the facing period, one of the human characters was presented, indicating the participant’s current self-orientation. In the targeting period, a photo of another character was presented on the scrambled background. The participants chose the direction of the target character relative to their body upon presentation of a response cue. (b) Maps were defined by the relative position of the three human characters, while the unfilled dot represents the wood plate. (c) The walking directions were defined by the spatial layout of the three human characters from the participant’s first-person perspective.

Materials and Methods

Participants

Nineteen right-handed university students with normal or corrected-to-normal vision were recruited from Peking University (12 females, 7 males). The average age of the participants was 24.9 years (range: 18–30 years). All participants had no history of psychiatric or neurological disorders and gave their written informed consent prior to the start of the experiment, which was approved by the Research Ethics Committee of Peking University.

Experimental Design

Virtual Environment

We programmed a 3D virtual environment using Unity software (Unity Technologies). The environment was designed with a circular fence as a boundary (48 virtual meters in diameter), a flat grassy ground, a uniform blue sky, and with a wood plate surrounded by four vertices of a square placed in the center (Fig. 1b, 4.7 virtual meters for side length). Three human characters (Mixamo, https://www.mixamo.com) were placed on three of the vertices in each trial. A map was defined by the relative relationship of the three human characters (Fig. 1b). From the six possible maps, three of them were pseudorandomly selected for each participant to collect enough number of trials’ data for each condition during the allowable range of scanning duration. The maps were the only environmental cues relevant to the task requirement; no distal cues were used outside the boundary. Participants performed the task using the first-person perspective with a 90° field of view (aspect ratio = 4:3), and they had never seen a top–down view of the virtual environment.

Walking Period

Participants walked from one of four starting locations near the circular boundary (4 virtual meters from the boundary) toward the human characters (Fig. 1c) and stopped on the wood plate. The visual stimuli (spatial environment viewed from first-person perspective) were determined by the combination of the map and walking direction; in other words, each map was presented by four different visual stimuli that were determined by the starting position (Fig. 1c). Importantly, participants were blinded to the map concept throughout the task. The walking period lasted for 6.0 s, during which each character had a 20.6% probability of nodding its head at a random time point between the start and end of walking. There was a 50%, 38.9%, 10.2%, and 0.9% probability for 0, 1, 2, and 3 characters to nod head in each trial; we subjectively selected a 20.6% head-nodding probability for each character to ensure an approximately equal number of trials with head-nodding and no head-nodding. During the walking period, participants were required to pay attention to the heads of the human characters rather than to memorize their spatial arrangement. The height of the participants was 1.8 virtual meters from the ground, which was the same as that of the human characters. No response was required during the walking period.

Two tasks were completed in two consecutive days. On day 1, the participants performed an HND task that did not include spatial memory trials. On day 2, participants performed a spatial memory task.

Head-Nodding Detection Task

Participants performed 144 randomly ordered head-nodding detection (HND) trials in a behavioral experimental room. In each trial, a photo of one of the characters was presented on a screen after the walking period, and participants were asked to indicate whether the character nodded its head or not (see Supplementary Fig. S1a). For this task, there was a 50% chance that the character in the presented photo nodded its head. Feedback was given after the participants had responded with either green (correct) or red (incorrect) photo border. The stimuli were rendered on a PC and presented on a 27-inch LCD monitor (ViewSonic XG2730) with a screen resolution of 1024 × 768. The HND task was used to examine whether participants paid attention to head nodding rather than memorizing the spatial arrangement of the characters, which would be indicated by high success rates in the head-nodding test.

Spatial Memory Task

During this task, participants performed 144 spatial memory trials (90%) and 16 HND trials (10%) that lasted ~70 min in an MRI scanner. Participants were notified that the remuneration depended only on the performance in the HND trials, although they were also encouraged to perform the spatial memory task as best as they could (videos of trial examples are available online for both tasks). The trial type (i.e., HND or memory task) was distinguishable after the walking period by subsequent stimuli. In the spatial memory task, a scrambled background was presented for 2 s after the walking period, and participants experienced a “facing period” and a “targeting period” sequentially. In the facing period, one of the human characters was presented in the center of the display with the environment background for 2 s as a facing character with the other two characters being invisible. In the targeting period, a photo of another character (targeting character) was presented as a target on a scrabbled background for 2 s. Each of the three experimental periods was followed by a 2-s delay (noise screen). At the end of each trial, participants indicated the direction of the target character relative to their self-body by pressing a button when a cue presented on the screen; no feedback was shown for both trial types (Fig. 1a). The spatial memory task contained four experimental sessions, each containing a spatial information combination of three maps × 4 walking directions × 3 facing character identities in each session, with targeting characters balanced across sessions. After scanning, all participants completed a post-scanning interview and reported the strategy they used to perform the task (see Supplementary Table S3).

fMRI Data Acquisition

Imaging data were collected using a 3T Siemens Prisma scanner equipped with a 20-channel receiver head coil. Functional data were acquired with a multiband echo planar imaging sequence (time repetition [TR] = 2000 ms, time echo [TE] = 30 ms, matrix size: 112 × 112 × 62, flip angle: 90°, resolution: 2 × 2 × 2.3 mm³, number of slices: 62, slice thickness: 2 mm, gap between slices = 0.3 mm, slice orientation: transversal). The signals of the original voxels (i.e., 2 × 2 × 2 mm³) were assigned to the corresponding voxels without gap (2 × 2 × 2.3 mm³) to construct participants’ native space image. Four experimental sessions were collected with, on average, 478, 476, 473, and 475 TRs, respectively. A high-resolution T1-weighted three-dimensional (3D) anatomical data set was collected to aid in registration (MPRAGE, TR = 2530 ms, TE = 2.98 ms, matrix size: 448 × 512 × 192, flip angle: 7°, resolution: 0.5 × 0.5 × 1 mm³, number of slices: 192, slice thickness: 1 mm, slice orientation: sagittal). During scanning, experimental stimuli were presented through a Sinorad LCD projector (Shenzhen Sinorad Medical Electronics) onto a 33-inch rear projection screen located over the subject’s head with a resolution of 1024 × 768 and viewed with an angled mirror positioning on the head coil.

fMRI Preprocessing

Functional data for each session were preprocessed independently using FSL FEAT (FMRIB’s Software Library, version 6.00, https://fsl.fmrib.ox.ac.uk/fsl/fslwiki; Woolrich et al. 2001, 2004). For each session, the first three functional volumes were discarded to allow for T1 equilibration, and the remaining functional volumes were slice-time corrected, realigned to the first image, and high-pass filtered at 100 s. For group-level statistics, each session’s functional data were registered to a T1-weighted standard image (MNI152) using FSL FLIRT (Jenkinson and Smith 2001; Jenkinson et al. 2002), and this procedure also resampled the functional voxels into a 2 × 2 × 2 mm resolution. For RSA, data were left unsmoothed to preserve any fine-grained spatial information (Chadwick et al. 2012). For functional connectivity analysis, data were smoothed using a 5-mm FWHM Gaussian kernel and were high-pass filtered at 0.01 Hz to remove low-frequency signal drifts.

Anatomical Masks

We manually delineated the MTL, including the HPC, parahippocampal cortex (PHC), perirhinal cortex (PRC), and entorhinal cortex (ERC) on each participant’s native space using established protocols (Insausti et al. 1998; Pruessner et al. 2000; Pruessner et al. 2002; Duvernoy 2005), as well as a delineating software ITK-SNAP (www.itksnap.org). The HPC was further segmented into its anterior, middle, and posterior parts because of the anatomical and functional variability along the HPC long axis (Poppenk et al. 2013). The anterior border of posterior HPC (pHPC) and the posterior border of anterior HPC (aHPC) were defined by the appearance of the crus of the fornix and the uncal apex relative to mHPC along the coronal orientation, respectively (Pruessner et al. 2000; Poppenk et al. 2013). For PFC subregions, we used the AAL template (Rolls et al. 2015) and selected four mPFC subregions for region of interest (ROI) analysis, which included the rectus, orbital mPFC (OFCmed), ventral mPFC (Frontal Med Orb), and superior mPFC (Frontal Sup Med). All ROIs were resampled and aligned with the functional volumes, and voxels outside of the brain were excluded.

Representational Similarity Analysis

Task-relevant information was decoded using RSA. We tried to dissociate the neural effect of walking (8 s), facing (4 s), and targeting (4 s) period from each period onset to the end of following noise period (Zeithamova et al. 2017). First, the trial-based multivoxel activity patterns were obtained separately for each period by creating a univariate general linear model (the first-level GLM) (Libby et al. 2014; Chadwick et al. 2015; Brown et al. 2016; Kim et al. 2017; Tompary and Davachi 2017). In each of walking, facing, and targeting period GLMs, the blood oxygen level–dependent (BOLD) signals across 36 trials (a session) were modeled using boxcar regressors. In addition to the 36 trial-based regressors of interest, two types of nuisance regressors were added. The first type depended on the task periods. For instance, the facing period GLM included 12 nuisance regressors to model the walking period by the visual patterns (3 maps × 4 walking directions) as well as three nuisance regressors to model the targeting period by the three targeting characters. Thus, the first-level GLM for each task period included nuisance regressors, which modeled the other two task periods by their visual stimulus conditions (e.g., modeling facing period by the three facing characters). The first type of nuisance regressors would reduce the influence of interference from other task events on the estimates of BOLD signal to trial-based regressors of interest for each task event. We therefore performed multiple univariate GLMs, which had different sets of the first type of nuisance regressors for the walking, facing, and targeting periods. The second type was not dependent on the task periods and included four HND trials, three directional cues in the response period, and six motion parameters. This procedure generated 36 trial-based multivoxel patterns in participant’s native space (2 × 2 × 2.3 mm voxels) for each period, and those multivoxel patterns were normalized prior to subsequent analysis by subtracting the grand mean pattern of the 36 multivoxel patterns for each session (Vass and Epstein 2013).

Searchlight-Based RSA

Next, we computed the representational similarity for each spatial information based on the 36 multivoxel patterns using a searchlight-based RSA (Libby et al. 2014; Chadwick et al. 2015), which was conducted using custom Matlab (version R2018b, www.mathworks.com/matlab/) scripts. In detail, a sphere with a 6 mm radius was constructed (85 voxels per sphere) for each brain voxel, and the spheres near the edge of the brain with fewer than 10 voxels were excluded from the analysis. The activity parameters within each sphere were extracted from each of the 36 multivoxel patterns, resulting in a 36-column by n-row (number of voxels within the sphere) matrix. The pattern similarity was then calculated between each column-by-column pair using Pearson’s correlation and was normalized using Fisher’s r-to-z transformation. This procedure finally generated a 36-by-36 correlation matrix for each period in each brain voxel. Next, we conducted a GLM as a regression on the correlation matrix, which included each element from one side of diagonal of the matrix but did not include the diagonal elements. To dissociate an effect of each spatial information from potential influences of others, multiple categorical regressors were used in this second-level GLM. In each regressor, either “1 (same)” or “0 (different)” were used to correspond with the correlation coefficient of a given column-to-row element of the correlation matrix. An estimated parameter of each regressor evaluated an increase or decrease in the similarity in the same condition relative to the different condition concerning on the spatial information specified by the regressor. For the facing period, the GLM contained five categorical regressors, which included the (1) “map,” (2) “walking direction,” and (3) “facing-character identity.” Since the participants reported that they had a body rotating-like experience from the walking direction to the self-orientation relative to the environment, we also added the (4) “rotation angle” (turn left/right 45°, turn left/right 135°) and (5) their “self-orientation” into the GLM. For the targeting period, seven regressors were built, which included (1) “map,” (2) “walking direction,” (3, 4) participants’ “rotation angle” and “self-orientation,” (5) “targeting character identity,” and (6, 7) “egocentric and allocentric position of target character.” It is important to note that the “facing character identity” was not included in the targeting period GLM since the effect of each facing character was regressed out in the GLM computing of multivoxel activity patterns. r² was computed and ranged from 0 to 0.03 for the facing period GLM and 0 to 0.04 for the targeting period GLM (see Supplementary Fig. S3). We also calculated variance inflation factor (VIF, 1/(1 − r²) (Craney and Surles 2002) to evaluate multicollinearity effects on the results of searchlight-based RSA (second-level GLM). The VIFs were <1.02 (see Supplementary Fig. S3). Each regressor’s parameter was then assigned to the center voxel of each sphere so that a whole-brain statistical parametric map could be generated for each spatial information for each period, with those spatial representations being finally averaged across the four scanner sessions and normalized to MRI template.

ROI-Based RSA

To validate the neural representation suggested by the searchlight-based RSA, we further conducted an independent RSA on map, self-orientation, and egocentric target direction using anatomical ROIs of MTL and mPFC subregions. We reasoned that if the spatial representations in small portions (i.e., clusters identified by searchlight analysis) of anatomical regions are stable enough, the averaged representational similarity in the corresponding anatomical regions (i.e., ROIs) would increase compared to a chance level. To test this, the chance level was determined by permutation test, in which the trial labels on the correlation matrix were shuffled before calculating the pattern similarity of “same” and “different” conditions; this procedure was performed 5000 times for each ROI, spatial information, session, and period. The baseline-corrected pattern similarity was calculated for each of “same” and “different” conditions from the correlation matrix; with each of the same and different conditions subtracted from the chance level and averaged across the four sessions, we then tested whether or not the baseline-corrected similarity is significantly different from zero across the participants using one-sample t-test (two-tailed). For each spatial information, the initial trial labels are identical across period, and the ROI-based RSA thus provides a direct comparison across periods.

Univariate Analysis

The univariate GLM was performed to model the bold signal of walking (8 s), facing (4 s), and targeting period (4 s); the three periods were included as main regressors. To remove nuisance effects, additional 11 regressors were added, one for modeling the blank period preceding the walking period, three for response key that participants pressed during pointing the target location, one for HND trials, and six for motion parameters. This procedure generated a parameter map corresponding to each period in participant’s native space (2 × 2 × 2.3 mm voxels). The parameter maps were normalized to MNI152 template and averaged across four scanning sessions before submitting to t-test for group-level statistic.

Functional Connectivity Analysis

To investigate the relationship between HPC and mPFC as well as their whole-brain network across task demands, we calculated the time course of functional connectivity between HPC and mPFC, as well as the contrasts of their whole-brain connectivity patterns between the facing and targeting periods, using each subregion of MTL and mPFC as seed (see Supplementary Fig. S8a; 12 for the MTL and 8 for the mPFC). In detail, we first removed the nuisance covariates from the preprocessed functional data by creating a GLM, which specified the signal averaged over the lateral ventricles, white matter, and whole brain; six motion parameters; and their derivatives as regressors. The residual signal was bandpass-filtered, leaving signals within the frequency range 0.01–0.1 Hz and was shifted by two TR intervals (4 s) for subsequent analysis (Tompary and Davachi 2017). To compute the time course of functional connectivity between the HPC and mPFC, for each TR, we extracted regional bold signal across trials, and the two areas’ bold signals were then correlated in each TR. For the whole-brain contrast, regional bold signals were computed for each of facing and targeting periods. To do this, we concatenated the bold signals of the two TRs in one trial with those in the next trial within a session (Ranganath et al. 2005, 72 bold signals in total: 2TRs × 36 trials); each anatomical mask’s bold signal was then correlated with the bold signal of each voxel in the rest of the brain, resulting in a whole-brain connectivity map for each period in each scanning session. The correlation maps were averaged across four scanning sessions, normalized to Montreal Neurological Institute template, and submitted to group-level statistics.

Each cluster, derived from the contrast analysis in connectivity between the facing and targeting periods based on an initial threshold of P = 0.001, was assigned to each of the three networks based on previous literatures: default-mode network, frontoparietal control network, and dorsal attention network (Vincent et al. 2008; Schacter et al. 2012; Gelström and Graziano 2017). In the present study, the default-mode network contains the clusters of mPFC, MTL, posterior cingulate cortex, and anterior temporal gyrus; the frontoparietal control network contains the clusters of paracingulate gyrus, lateral PFC (lPFC), and inferior parietal lobule; and the dorsal attention network contains the clusters of occipital pole, lateral occipital cortex, cuneal cortex, lingual gyrus, superior parietal lobule (SPL), and postcentral gyrus (see Supplementary Fig. S8b). To examine modulation effects on the connectivity of MTL/mPFC with the large-scale networks by different task demands, we computed connectivity between each subregion of MTL (HPC, PHC, PRC, and ERC) and mPFC (rectus, orbital mPFC, ventral mPFC, and superior mPFC) with the three networks in each period for each participant. The connectivity was averaged across the subregions in each of the MTL and mPFC before submitting it to group-level test. Note that for the default-mode network, we prepared two masks, in which the MTL and mPFC was removed from the network mask for the FC between MTL and default-mode network and between mPFC and default-mode network, respectively.

Statistical Analysis

For searchlight-based RSA, we used an initial threshold of P < 0.001. If no clusters were revealed, a liberal threshold of P < 0.01 was used. For whole-brain FC analysis, an initial threshold of P < 0.001 was used to identify robust network patterns. The reliability of clusters was tested using a nonparametric statistical inference that does not make assumptions about the distribution of the data (Nichols and Holmes 2002; Winkler et al. 2014; Chadwick et al. 2015); the test was conducted with the FSL randomize package (version v2.9, http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/Randomise) and performed 5000 random sign-flips on whole-brain searchlight beta images or FC images; we then reported clusters with the size higher than 95% of the maximal suprathreshold clusters in permutation distribution. For ROI-based analysis, two-tailed t-test was used to examine the RSA, while two-tailed paired t-test was used to examine the FC among MTL and mPFC subregions., Bonferroni corrected for multiple comparison. The univariate analysis used t-test with uncorrected for multiple comparison. Two-way analysis of variance was used to test the connectivity change in the three large-scale functional networks modulated by the period (facing and target period) and brain area (MTL and mPFC). The statistical significance was determined according to whether the corrected P value was smaller than 0.05.

Results

The experiment was conducted over 2 days with 19 participants. On the first day, the participants were familiarized with the 3D virtual environment and the three human characters through a HND task (see Supplementary Fig. S1a). In this task, the participants had the same walking experience as in the spatial memory task but were subsequently asked to indicate whether one of the three characters in a photo had nodded its head during the walking period. On the second day, the participants performed the spatial memory task during fMRI scanning (Fig. 1a). To prevent voluntary memorization of the spatial relationship of the human characters during the walking period, the HND trials were pseudorandomly mixed with the spatial memory trials, and the participants were instructed to focus on head nodding of the human characters during the walking period in all trials. In each trial, its trial type (i.e., HND or memory task) became distinguishable only after the walking period by subsequent stimuli. All participants exhibited ceiling performance with a 93.6% ± 1.5% correct rate (mean ± SE, n = 19) for the spatial memory task and no significant difference was found among each of the task parameters (e.g., maps, walking directions) (see Supplementary Fig. S1c). All participants also showed accuracy that was significantly higher than chance level (50%) in both the head-nodding and no head-nodding trials in the HND task (see Supplementary Fig. S1a). Attempts to memorize the spatial arrangements of the human characters during scanning were examined using post-scanning questionnaires. All participants reported that they did not make any voluntary effort to memorize the spatial relationship of the three human characters nor utilize any special strategy for memorizing it (see Supplementary Table S3). It should be noted that no participant was able to recall the number of map patterns they experienced in the experiment even though only three of the six possible patterns of maps were repeatedly presented to each participant. In addition, no significant changes in performance were found across four experimental sessions (see Supplementary Fig. S1b; F(3,72) = 0.38, P = 0.76), suggesting that the participants performed the spatial memory task with high performance from the beginning and did not learn to use a systematic strategy to improve their performance during the sessions. These behavioral results suggest that the present experimental design allowed us to investigate neural operations for the retrieval after the participants automatically encoded the spatial configuration of three human characters during the walking period when viewing the characters attentively to detect head nodding.

Neural Representation of the Object-Based Cognitive Map

To decode the map information across the whole brain, we conducted searchlight-based RSA, which examined the multivoxel pattern similarity between trial pairs in the “same map” and compared it with that in the “different map” condition across each brain voxel by drawing a 6 mm radius sphere with each voxel in the spherical center. Map information was decoded regardless of other task parameters, such as the walking direction or the identity of the facing character, by balancing the number of trials with other task parameters across the same and different map conditions in the experimental design as well as excluding the effects of other task parameters in the regression analysis (see Methods for details, see Supplementary Figs S2 and S3).

We first assessed the map representation during the facing period (4.0 s including the subsequent delay; Fig. 2a), in which the participants oriented themselves to a presented human character in the 3D environment. We found a cluster located in the left middle HPC (mHPC; Fig. 2b, P <  0.01, voxel-wise threshold; P < 0.05, cluster corrected for multiple comparison), suggesting that the map defined by multiple objects is represented in the HPC. In addition to the mHPC, clusters were revealed in the insula, angular gyrus, superior temporal cortex, and fusiform (see Supplementary Fig. S4b and Supplementary Table S1, P < 0.01, voxel-wise threshold; P < 0.05, cluster corrected for multiple comparison; see discussion). We next assessed the map representation during the targeting period (4.0 s including the subsequent delay), in which the participants remembered the location of a target character relative to their self-body (egocentric target location). In contrast to the facing period, we found clusters representing the map information mainly in the mPFC (Fig. 2c) rather than in the HPC during the targeting period. A peak was revealed in the rectus (x = 4, y = 50, z = −18; t value: 5.62). In addition to the rectus, there was a significant cluster extending from the ACC into the superior mPFC. Outside of the mPFC, we found clusters in the precuneus and middle temporal gyrus and the inferior frontal cortex (Fig. 2c; Supplementary Table S2 and Supplementary Fig. S4c; P < 0.001, voxel-wise threshold; P < 0.05, cluster corrected for multiple comparison). These three brain regions have been consistently reported to be involved in scene construction during recalling of past experience and imagination of new experiences (Hassabis and Maguire 2007; Bird et al. 2010; Gaesser et al. 2013), which is consistent with the post-scanning report that all participants recalled and also imagined the egocentric positions of the three human characters during the targeting period.

Figure 2

Neural representation of cognitive map in MTL and mPFC. (a) Schematic representation of decoding the maps using RSA. (b) In the facing period, RSA revealed a cluster in the left middle HPC (mHPC; t value: 3.84; Montreal Neurological Institute [MNI] coordinates: −28, −25, −16; shown on sagittal and transverse sections) within the MTL (P < 0.01, initial threshold; P < 0.05, cluster corrected for multiple comparison). (c) In the targeting period, clusters were revealed in the mPFC (P < 0.001, initial threshold; P < 0.05, cluster corrected for multiple comparison). The peak was revealed in the rectus within the mPFC (t value: 5.62; MNI coordinates: 4, 50, −18). (d) The left mHPC and bilateral rectus revealed a significantly higher similarity than chance level in the “same map” condition during facing and targeting periods, respectively (the left mHPC: t(18) = 3.26, P = 0.016; the left rectus: t(18) = 3.68, P = 0.007; the right rectus: t(18) = 4.50, P = 0.001). ^***P < 0.001, ^**P < 0.01, ^*P < 0.05. Bonferroni corrected for multiple comparisons, n = 4. Error bars indicate SEM.

To validate the searchlight-based RSA result showing the map representations in the HPC and mPFC, we conducted a ROI-based RSA using the anatomical masks of the mHPC and rectus that derived from the automated anatomical labeling (AAL, Supplementary Fig. S8a, bottom panel) template. In each ROI, we examined the multivoxel pattern similarity in the “same map” condition and in the “different map” condition separately. These similarities were then compared with chance levels that were estimated as mean values among 5000 of trial-based permutation results (see Methods).

In the left mHPC, the similarity was significantly higher than the chance level in the same map condition during the facing period (Fig. 2d; t(18) = 3.26, P = 0.016; Bonferroni corrected for multiple comparisons, n = 4), while it was significantly lower than the chance level in the different map condition (t(18) = −3.26, P = 0.016). On the other hand, the similarities in the left mHPC did not reach a statistical significance for neither of the two conditions during the targeting period (same map: t(18) = 1.99, P = 0.062; different map: t(18) = −1.99, P = 0.062, uncorrected), although the same tendency was observed between the two periods. The map representation (similarity, “same map condition” minus “different map condition”) in the left mHPC did not significantly differ between the two periods (t(18) = 0.39, P = 0.69) nor in the searchlight-based RSA (t(18) = 0.98, P = 0.33). In contrast to the left hemisphere, the similarities did not differ from the control levels in any combination of the conditions and periods in the right mHPC (Fig. 2d), suggesting a striking laterality effect on the map representation in the HPC. On the other hand, both hemispheres of the rectus showed significantly higher and lower similarities than the control levels in the same map (the left rectus: t(18) = 3.68, P = 0.007; the right rectus: t(18) = 4.50, P = 0.001) and different map conditions (the left rectus: t(18) = −3.68, P = 0.007; the right rectus: t(18) = −4.49, P = 0.001), respectively, only during the targeting period. This tendency was observed during the facing period only in the left hemisphere, although the similarities for neither of the two conditions reached a statistical significance (same map: t(18) = 1.60, P = 0.126; different map: t(18) = −1.60, P = 0.126). The ROI-based RSA also confirmed that the superior mPFC and ACC tended to represent the map information only during the targeting period (see Supplementary Fig. S5). To examine a change of the similarity across task periods in a trial, we also calculated the map representation in the walking period. The similarity in the mHPC increased from waking to facing period followed by a decreasing trend in targeting period, while the rectus showed an increased similarity in targeting period relative to walking and facing periods (see Supplementary Fig. S6). The ROI-based RSA thus confirmed the results of the searchlight-based RSA, indicating that the cognitive map was represented in the HPC during the facing period and the mPFC represented it during the targeting period. It should be noted here that the map representation in the different brain areas (i.e., HPC and mPFC) during the facing and targeting periods could not be explained by different background images during the two periods themselves (i.e., environment vs. noise) because the RSA revealed brain regions that discriminated trials with different map information during each period in which the same background was presented. Taken together, the HPC and mPFC may represent the cognitive map information for the different functional operations. Some temporally overlapping map representation in the two brain areas may implicate an interaction to share the map information between them, although it did not reach a statistical significance (i.e., map representation in the mPFC and HPC during the facing and targeting periods, respectively).

Input Signal for the Map Construction

To examine possible signal input from the MTL subregions to the left mHPC for the map construction during the facing period, we examined the neural representation of the facing character identity and walking direction that the participants perceptually and/or mentally re-experienced based on their post-scanning reports (see Supplementary Table S3; Fig. 3a). The results revealed that the bilateral PRC encoded character identity (Fig. 3b; P < 0.001, initial threshold; P < 0.05, cluster corrected for multiple comparison) (Naya et al. 2001; Suzuki and Naya 2014; Chen and Naya 2020), while the PHC and left retrosplenial cortex (RSC) encoded the walking directions reflecting the spatial layout of one empty and three occupied positions perceived by the participants during the walking period (Fig. 1c). In the HPC, the left pHPC selectively represented the spatial layout but not the character identity, while the bilateral aHPC revealed clusters for both character identity and spatial layout (Fig. 3b). These results were consistent with the notion of the “two cortical systems” model, suggesting that object identity and spatiotemporal context are processed in two separate neural systems with the PRC and PHC–RSC as the core brain regions, with the two different information domains interacting in the HPC (Ranganath and Ritchey 2012). Together, the RSA analyses suggest that the MTL is associated with representing the spatial environment in the following ways: elements, such as each object identity and spatial layout, are represented by extrahippocampal areas while the relative relationship between multiobjects is represented in the HPC, suggesting cognitive map representation in the HPC.

Figure 3

Neural representation of the walking direction and character identity in the MTL. (a) Schematic representation of decoding walking direction (left) and the character identity in the facing (middle) and targeting (right) periods. (b) The PRC selectively encodes the character identity across both periods but not the walking direction. The PHC, PPA, and HPC encode both the character identity and walking direction. In particular, there was a clear attenuation in the encoding for the walking direction in the targeting period compared to the facing period (P < 0.001, initial threshold; P < 0.05, cluster corrected for multiple comparison). (c) The clusters located in the lPFC for the walking direction and character identity were revealed across the two periods.

Current Self-orientation on the Map

To compute the egocentric location of a target object (e.g., left, right, or back), information on the current self-position/orientation on the map is necessary (Fig. 4a). Therefore, we examined which brain regions were involved in representing such allocentric “heading-direction” signals (Hargreave et al. 2005; Wang et al. 2018). Interestingly, while no significant cluster was revealed during the facing period (facing character; Fig. 4b), robust clusters were revealed during the targeting period (Fig. 4b, P < 0.001, voxel-wise threshold; P < 0.05, cluster corrected for multiple comparison). These clusters were located in the left ERC, bilateral HPC, and PHC inside the MTL as well as in the lateral occipital cortex, parietal cortex, precuneus, and anterior cingulate cortex outside the MTL (Fig. 4b, P < 0.001, voxel-wise threshold; P < 0.05, cluster corrected for multiple comparison). These results were confirmed by the ROI-based RSA. The MTL subregions showed substantially higher similarity in the same self-orientation condition than chance level (HPC: t(18) = 5.82, P < 0.001; ERC: t(18) = 4.80, P < 0.001; PHC: t(18) = 2.95, P = 0.051; Bonferroni corrected for multiple comparison), while only negligible effects were found in the mPFC (Fig. 4c). Those MTL subregions began to show the self-orientation after the presentation of the targeting character (see Supplementary Fig. S6), presumably because of the necessity to compute the egocentric target location. This interpretation is consistent with the post-scanning report in which participants reported imagining their self-orientation on the map only during the targeting period.

Figure 4

Neural representation of self-orientation on cognitive map. (a) Schematic representation of decoding participants’ self-orientation. (b) In the facing period, no cluster was revealed even with the use of a more liberal threshold (P < 0.01, initial threshold; P < 0.05, cluster corrected for multiple comparison). In the targeting period, clusters were revealed in the MTL (bilateral HPC, PHC, and left ERC) and self-motion areas (inferior parietal cortex, RSC, and lateral occipital cortex). (c) ROI-based RSA revealed, in the “same” condition, a higher similarity than chance level in the MTL (ERC: t(18) = 4.80, P < 0.001; HPC: t(18) = 5.82, P < 0.001; PHC: t(18) = 2.95, P = 0.051) but not mPFC. ^***P < 0.001, ^*P < 0.05, ^~P < 0.1. Bonferroni corrected for multiple comparison, n = 6. Error bars indicate SEM.

Remembering the Egocentric Location of a Target Object

Next, we examined which brain regions signaled the egocentric location (left, right, or back relative to self-body) of a target object (Fig. 5a). The results revealed robust clusters in both the mPFC and MTL (Fig. 5a, P < 0.001, voxel-wise threshold; P < 0.05, cluster corrected for multiple comparison). In the mPFC, we identified the rectus, ventral and superior mPFC, and olfactory cortex. In the MTL, clusters were found in the anterior HPC. These results were confirmed by ROI-based RSA (Fig. 5b; Supplementary Fig. S6, rectus: t(18) = 3.82, P = 0.007; orbital mPFC: t(18) = 5.03, P < 0.001; superior mPFC: t(18) = 3.68, P = 0.01; anterior HPC: t(18) = 5.56, P < 0.001; Bonferroni corrected for multiple comparison, n = 6), which revealed a selective increase of the egocentric target location representation during the targeting period compared with the preceding walking and facing periods (Supplementary Fig. S6). Apart from the mPFC and MTL, clusters were also found in the lateral occipital cortex, inferior parietal cortex, anterior temporal lobe, premotor cortex, and lPFC (middle and superior PFC). We also found clusters in the precuneus and posterior parietal cortex, which were previously reported to represent the egocentric location (Chadwick et al. 2015). The widely distributed clusters (Fig. 5a; Supplementary Table S2) may indicate that the brain regions representing the egocentric target locations can be involved in either generation of the egocentric target location information from multiple pieces of information (cognitive map, self-orientation, and target character identity) or its maintenance while preparing for the following response. These distinct functions might be supported by three different large-scale brain networks: the dorsal attention network, frontoparietal control network, and default-mode network (Spreng and Schacter 2011). In contrast to the robust signal observed across different brain networks for egocentric target location, no cluster was revealed for allocentric target location relative to the spatial layout of the characters (Fig. 5c, P < 0.001, voxel-wise threshold; P < 0.05, cluster corrected for multiple comparison), which implies that the target location may be directly retrieved in the form of egocentric coordinates rather than via its allocentric representation.

Figure 5

Neural representation of retrieved egocentric target location. (a) Left panel: Schematic representation of decoding the egocentric direction of a target character. Right panel: Clusters were revealed across a wide range of brain areas (P < 0.001, initial threshold; P < 0.05, cluster corrected for multiple comparison). Many of the clusters belonged to one of the following three functional networks: the default-mode network, frontoparietal control network, and dorsal attention network. The aHPC is shown on sagittal and transverse sections of volume image for display purpose (P < 0.001, initial threshold; P < 0.05, cluster corrected for multiple comparison). (b) ROI-based RSA revealed, in the “same” condition, a significant higher similarity than chance level in the mPFC (rectus: t(18) = 3.82, P = 0.007; ventral mPFC: t(18) = 5.03, P < 0.001; and superior mPFC: t(18) = 3.68, P = 0.01) and in the aHPC (t(18) = 5.56, P < 0.001). ^***P < 0.001, ^**P < 0.01, ^*P < 0.05. Bonferroni corrected for multiple comparison, n = 6. Error bars indicate SEM. (c) Left panel: Schematic representation of decoding allocentric direction of a target character. Right panel: No clusters were revealed even with the use of a liberal threshold (P < 0.01, initial threshold; P < 0.05, cluster corrected for multiple comparison).

Increased Default-Mode Network Connectivity While Locating a Target Compared with Self-locating

The above results showed that the HPC and mPFC signaled a coherent map coding a spatial relationship of the three human characters during the different time periods in which different task demands were required (i.e., self-locating and target locating). In addition, the MTL, including the HPC, and mPFC signaled the different location information even during the same targeting period; MTL areas tended to represent self-orientation, while the mPFC tended to represent egocentric target location. We next investigated when the HPC exchanged the spatial information with the mPFC as well as how the different functional contributions of the MTL and mPFC were substantiated by whole-brain large-scale networks. For that purpose, we manually segmented the MTL subregions in each participant’s native space (see Supplementary Fig. S8a, top panel) and conducted a task-based functional connectivity analysis using each subregion of MTL and mPFC as seed (see Methods for details).

First, we examined the time course of the functional connectivity between the HPC and mPFC (Fig. 6a). The connectivity was the highest during the inter-trial interval, which may be the characteristics of the HPC and mPFC as members of the default-mode network. The connectivity decreased after the trial start, but it showed an increase from facing to targeting period. The rectus, ventral mPFC, and superior mPFC showed significantly larger connectivity with the HPC in the targeting period relative to facing period. (rectus: t(18) = 3.29, P = 0.01; ventral mPFC: t(18) = 4.76, P < 0.001; superior mPFC: t(18) = 3.38, P = 0.01; orbital mPFC: t(18) = 1.34, P = 0.78; Bonferroni corrected for multiple comparison (n = 4)). Second, we examined the connectivity between each seed and the whole brain (Fig. 6b). Both the MTL and mPFC showed significantly larger connectivity to brain areas that belong to the default-mode network and those to the dorsal attention network during the targeting period compared to the facing period (Fig. 6c, P < 0.001, voxel-wise threshold; P < 0.05, cluster corrected for multiple comparison). By contrast, both the MTL and mPFC showed significantly larger connectivity to the frontoparietal control network during the facing period relative to the targeting period (Fig. 6b, P < 0.001, voxel-wise threshold; P < 0.05, cluster corrected for multiple comparison). These results suggest that both the MTL and mPFC changed their connectivity to the three functional networks depending on the task demands. We next evaluated how strongly the MTL and mPFC connected with the three networks during each task period using a functionally defined mask for each network (see Supplementary Fig. S8b). This ROI analysis revealed that the default-mode network was positively correlated with the MTL (t(18) = 7.98 for average across the subregions within MTL, P < 0.001) and mPFC (t(18) = 9.63 for average across the subregions within mPFC, P < 0.001) for both time periods with a significant increase during the targeting period (Fig. 6d left panel; Supplementary Fig. S9; F(1,72) = 4.51, P = 0.03), regardless of the seeds (MTL or mPFC; F(1,72) = 0.00, P = 0.98). In contrast, the frontoparietal control network showed significantly negative connectivity with the MTL (t(18) = −10.50, P < 0.001) and mPFC (t(18) = −6.55, P < 0.001) during both task periods, and this negative connectivity was stronger during the targeting period (Fig. 6d middle panel, F(1,72) = 5.58, P = 0.02; also see Supplementary Fig. S9). These results suggest that the default-mode network contributes more to the retrieval of the target location than the self-location to an external reference. Interestingly, despite both the MTL and mPFC being part of the default-mode network, they showed opposite connectivity patterns to the dorsal attention network during both periods (Fig. 6d, right panel; Supplementary Fig. S9; F(1,72) = 55.07, P < 0.001); the MTL positively with the network while the mPFC negatively correlated with it. The connectivity between the MTL and the dorsal attention network increased from the facing to targeting period (F(1,72) = 8.43, P = 0.005). These results suggested that the dorsal attention network, which contains the SPL that represented egocentric target location (Fig. 5a), showed increased coupling with the MTL during the targeting period.

Figure 6

Increased default-mode network connectivity while locating a target compared with locating oneself. (a) The time course of connectivity between mPFC and HPC demonstrated an increased trend from facing to targeting period (shaded area indicates 95% confidence interval). (b) The frontoparietal control network showed enhanced connectivity strength with the MTL and mPFC in the facing period compared to the targeting period (P < 0.001, initial threshold; P < 0.05, cluster corrected for multiple comparison). (c) The default-mode network and dorsal attention network showed enhanced connectivity strength with the MTL and mPFC in the targeting period compared to the facing period (P < 0.001, initial threshold; P < 0.05, cluster corrected for multiple comparison). (d) The mean connectivity strength of the subregions in each of the MTL and mPFC with each of three networks was shown in each of the facing and targeting periods. Two-way analysis of variance with the brain regions (MTL vs. mPFC) and the task periods (facing vs. targeting) as main factors revealed a significant main effect of the task periods for all the three networks (default mode: F(1, 72) = 4.51, P = 0.03; frontoparietal control: F(1, 72) = 5.58, P = 0.02; dorsal attention: F(1, 72) = 55.07, P < 0.001). Note that the connectivity between default-mode network and MTL/mPFC was examined using default-mode network mask without the MTL/mPFC, respectively. Error bars indicate SEM.

Discussion

In this study, we examined neural representations of space defined by three objects and found that both the HPC and mPFC represented the object-based space around the participants. Interestingly, the HPC represented the object-based map when the participants locate their self-body in the environment constructed by the three objects, while the mPFC represented the map when the participants remembered the location of a target object relative to the self-body. These results suggest that the cognitive maps in different brain regions play different functional roles. In addition, during the targeting period, we found differential spatial representations across the MTL and mPFC: the MTL generally reinstated self-orientation, while the mPFC represented egocentric target location relative to self-location. Increased functional connectivity was observed between the MTL and mPFC under the necessity of the retrieval of the target location from the stored memory (targeting period) compared to when they actually faced the reference object to locate their self-body (facing period). These results suggest that mental representations of the external world formed by the coherent space and its constituent elements may be shared in the default-mode network including the MTL and mPFC. The special role of the mPFC in this scheme might be to select the object location based on the mnemonic information including the cognitive map and current self-location on it, which might be propagated from the MTL.

To examine the representation of spatial “maps” (Fig. 1b), the present task was designed to cancel out the effects of a particular encoding experience related with the walking direction as well as a particular object identity that the participants viewed during the facing and targeting periods in each trial by balancing number of trials with each of those confounding factors in each map (see Methods). Therefore, the neural representation of the map information revealed by the RSA could not be explained by perceptual information in the present study. Moreover, the participants always stood on the center of the virtual environment during the facing and targeting periods, during which the map effect was examined. Because of this task design, the map information does not directly indicate self-location information like place fields of place cells in the HPC (O'Keefe and Dostrovsky 1971). On the other hand, the representations of place fields are reportedly influenced by the animal’s cognitive map, and the existence of cognitive maps could be most clearly demonstrated by a phenomenon known as “remapping,” which reportedly occurs in populations of place cells in the rodent HPC (Moser et al. 2017). Therefore, it might be reasonable to interpret the map representations in the left mHPC during the facing period as experimental evidence of “remapping” of place cells in the human HPC even though the participants stood in the same position. However, holding this interpretation predicts that human place cells are localized in the left mHPC. This prediction is against consistent evidence from previous human studies reporting that the right HPC was more involved in encoding and retrieving spatial information than the left HPC (Abrahams et al. 1997; Maguire et al. 1997; Ekstrom et al. 2003; Doeller et al. 2008; Schinazi et al. 2013). The other possible interpretation for the map representation in the left mHPC is that it may encode an allocentric spatial relationship of the three objects itself. This interpretation is consistent with previous human imaging studies reporting contributions of the left HPC to the imagination of visual scenes, which could be constructed from multiple spatial elements (Addis et al. 2007; Bird et al. 2010). The specific role of the left HPC in relational memory was also reported in nonspatial information domains, including associative learning (Kumaran et al. 2009; Suarez-Jimenez et al. 2018) and social interactions (Tavares et al. 2015). Taken together, it might be more reasonable to interpret that the clusters in the left mHPC was related to a coherent space constructed by the multiple objects rather than its influence on representations of individual spatial elements such as self-location or head direction. RSA also suggested the involvement of the PRC and PHC in MTL signaling the object identity and egocentric view of their spatial layout, respectively, which might be used for constructing the coherent map from its constituents in the left mHPC. Future studies should address how the coherent map can be constructed by multiple objects in the MTL.

In contrast to the facing period, the representation of the map information was found largely in the mPFC during the targeting period, while the left mHPC had only a tendency to exhibit the map information that might be sustained from the facing period. In addition to the map information, the mPFC signaled the egocentric location of a targeting object, while the HPC concurrently signaled both the egocentric location of a targeting object and the self-orientation. These results indicate the selective recruitment of the mPFC during the targeting period compared with the involvement of the HPC across the two periods. The differential contribution patterns between the mPFC and HPC during the two task periods were also supported by the univariate analysis (see Supplementary Fig. S7). Taken together, the mPFC may be selectively involved in constructing goal-directed information in the current context, which is consistent with accumulating evidence showing that the mPFC contributes to decision-making or action selection (Saxena et al. 1998; Gallagher et al. 1999; Feierstein et al. 2006; Spiers and Maguire 2007; Kable and Glimcher 2009; Young and Shapiro 2011; Balaguer et al. 2016; Yamada et al. 2018). These previous studies consistently supported the notion that the mPFC function becomes obvious when an appropriate selection requires mnemonic information in addition to incoming perceptual information (Bradfield et al. 2015). In this study, together with perceptual information responsible for target object identity, mnemonic information, such as the map information and self-orientation, was required to solve the task. Considering that the MTL could provide all the necessary mnemonic information, a reasonable interpretation is that the mPFC was involved in the selection of a target location among alternatives rather than the recollection or generation of it.

In addition to the HPC and mPFC, the map information has been observed in other brain areas, such as the angular gyrus (Seghier 2013; Price et al. 2016), lateral temporal gyrus (Karnath 2001; Himmelbach et al. 2006), and precuneus (Cavanna and Trimble 2006), that also belong to the default-mode network. The brain areas in the default-mode network, particularly the MTL subregions except for the PRC represented self-orientation during the targeting period. On the other hand, RSA analysis showed that the representation of the egocentric target object location recruited widely distributed brain regions, which belong not only to the default-mode network but also to the dorsal attention network and frontoparietal control network. The increased functional connectivity between the MTL and the brain regions of dorsal attention network as well as the mPFC suggests that the egocentric target location signal is transmitted from the mPFC to the dorsal attention network, such as the SPL (Evans et al. 2016), via the MTL, which implies a pivotal functional role of MTL as a hub of mental representation of object-related (e.g., identity and location) signals. This signal transmit across different brain networks may be related to the fact that the egocentric target location was the main behaviorally relevant feature in the present task. These results contrast with the present result that no brain regions represented the allocentric target location relative to the spatial layout of the characters, which was not required to answer in the task.

Interestingly, the frontoparietal control network showed a strong negative correlation with both the MTL and mPFC during both the facing and targeting periods, although the lPFC in the frontoparietal control network represented both the map information and egocentric direction during the targeting period. In addition, the lPFC represented walking direction as well as character identity during both periods. These results suggest that the lPFC computes the target location independently of the default-mode network. The parallel contributions of the lPFC and MTL-mPFC in choosing the target location may reflect their different cognitive functions (Jimura et al. 2004). lPFC has long been considered as a center of executive functions (Funahashi 2017; Miller et al. 2018) equipped with working memory (Andrews et al. 2011; Barbey et al. 2013; Brunoni and Vanderhasselt 2014; Funahashi 2017). In human fMRI studies, the lPFC has been shown to contribute to the retrieval of task-relevant information when more systematic thinking is required (Epstein et al. 2017; Javadi et al. 2017). In the present study, the behavioral task was designed to ensure participants neither actively maintained a spatial configuration of the human characters during the walking period nor any systematic strategy to solve the task, which was confirmed by the post-scanning test results. The greater signal for the cognitive map and the egocentric target location in the mPFC than that in the lPFC may reflect that the current spatial memory task was enough easy to allow participants to depend only on the involuntary encoding and subsequent memory retrieval for their top-ceiling performance (Epstein et al. 2017; Javadi et al. 2017).

In contrast to previous memory/navigation studies, which examined brain functions using spatial environments consisting of immobile landmarks (e.g., stores) and/or landscapes (e.g., mountains) (Bird et al. 2010; Woollett and Maguire 2011; Schinazi et al. 2013; Chadwick et al. 2015; Brown et al. 2016), the present study used a spatial environment constructed by only mobile objects that could become targets and references of self-location as well as determine the space (i.e., map) around oneself. This task design allowed us to extract a mental representation of the spatial environment consisting of the minimum essential constituents. This reductionist method could be useful for future studies investigating the construction and functional mechanisms of a cognitive map because of its simplicity. One critical concern might be whether the findings discovered by this reductionist method can be applied to a more complicated cognitive map consisting of large numbers of immobile spatial elements, which could be learned through extensive explorations over a long time period (e.g., the city of London) (Woollett and Maguire 2011). Another related concern might be whether our brain system holds only one cognitive map or multiple ones at a time (Meister and Buffalo 2018). For example, we may hold an object-based cognitive map consisting of relevant mobile objects such as same species, predators, and foods, while we may also hold the other cognitive map consisting of landmarks, landscapes, and other immobile objects such as trees. Future studies should address the relationships of different types of cognitive maps (e.g., mobile vs. immobile, short term vs. long term) and their underlying neural mechanisms.

The present study found neural representations of the space specified by objects around us. This object-based cognitive map seems to interact with representation of self-location in HPC and to mediate a selection of egocentric target location in mPFC, which would serve for leading us to the goal position. In addition to the spatial navigation, an existence of the object-based cognitive map may equip us with a space representation for persons separately from the background, which may serve for our social interactions (Damasio et al. 1994; Stolk et al. 2015) as well as the encoding and retrieval of episodic memory (Tulving 2002; Squire and Wixted 2011).

Notes

We thank Li Sheng and Sun Pei for helpful discussions and also thank Arielle Tambini, Lusha Zhu, Koji Jimura, Rei Akaishi, and Cen Yang for comments on an early version of the manuscript. Computational work was supported by resources provided by the High-performance Computing Platform of Peking University. Conflict of Interest: None declared.

Author’s Contributions

B.Z. and Y.N made the experimental design. B.Z. conducted all experiments and data analysis under supervision of Y.N. B.Z. and Y.N. wrote the paper.

Funding

National Natural Science Foundation of China Grant 31421003 (to Y.N.).