Evidence for allocentric boundary and goal direction information in the human entorhinal cortex and subiculum

In rodents, cells in the medial entorhinal cortex (EC) and subiculum code for the allocentric direction to environment boundaries, which is an important prerequisite for accurate positional coding. Although in humans boundary-related signals have been reported, there is no evidence that they contain allocentric direction information. Furthermore, it has not been possible to separate boundary versus goal direction signals in the EC/subiculum. Here, to address these questions, we had participants learn a virtual environment containing four unique boundaries. Participants then underwent fMRI scanning where they made judgements about the allocentric direction of a cue object. Using multivariate decoding, we found information regarding allocentric boundary direction in posterior EC and subiculum, whereas allocentric goal direction was decodable from anterior EC and subiculum. These data provide the first evidence of allocentric boundary coding in humans, and are consistent with recent conceptualisations of a division of labour within the EC.

3 there were no systematic differences in visual input between, for example, the four allocentric boundary directions.
In addition, to test whether the decoding in EC or subiculum was driven by perceptual features of the stimuli, we have conducted further analyses. Firstly, by selecting different portions of the trial for our decoding analysis, we sub-sampled the data to examine the temporal evolution of the boundary signal. Our hypothesis was that if the passive movement had induced a visual confound, the TRs encompassing the peak of the haemodynamic response function (HRF) should provide maximum information about the visual content. As a consequence, TRs around 4-6 seconds after the onset of passive movement should show higher decoding scores relative to later time windows when the HRF is nearing baseline again. As can be seen in Supplementary Figure 13, in the posterior EC and subiculum, the highest decoding accuracies were achieved later in the trial, and not in the TRs corresponding to the portion of the trial encompassing 4-6 seconds after the viewing of the passive movement where visual content would likely have its strongest influence (i.e., 2-8s).
Secondly, in response to your comment number 4 regarding the decoding of egocentric conditions, we split the trials according to egocentric boundary direction.
This analysis implements a strong visual confound, because the boundary is classified according to its position in the visual field (i.e. right, left or centre). As expected, this visual confound made it possible to decode egocentric boundary information in area V1 (Supplementary Figure 12). In contrast, we did not observe above-chance decoding in either the posterior EC or the subiculum (Supplementary Figure 11).
Together, these findings strongly suggest that it is unlikely that low-level visual features might have driven the decoding of allocentric boundary / goal directions in EC or the subiculum. Importantly, our results are also in line with previous research that has used elegant controls to test for the coding of visual information in medial temporal lobe regions. For example, Chadwick et al. (2015) tried to decode distinct visual information from their virtual environment in their EC/subiculum cluster and found that although this region was sensitive to directional information, it did not code for perceptual qualities of the stimuli. I assume the authors have not done these analyses due to strong visual and button choice confounds, but nonetheless, the results would be useful evidence for future studies that might control for such information.
We thank the Reviewer for highlighting the importance of demonstrating evidence of egocentric coding and for completeness we have now provided in the Supplementary Information the decoding results for the egocentric boundary and egocentric goal direction conditions in the same ROIs. As the Reviewer notes, one should express caution when interpreting these results given the visual confounds due to lower-level visual similarities when averaging the trials in this way. For example, unlike in the allocentric condition, in the egocentric boundary condition the boundary will be located either to the left, right, or straight in front of the participant, which provides very distinct visual information as to the different classes of stimuli. For the egocentric goal direction, although the boundary position will be balanced across the two stimulus classes (goal cue object left versus goal cue object right), the lower-level feature of the cue object position will consistently discriminate the two conditions. In our key ROIs, it was possible to decode egocentric goal direction in anterior subiculum only (Supplementary Figure 11). As can be seen in Supplementary Figure   12, we were able to decode egocentric boundary direction in the CA1 and CA23DG, whereas egocentric goal direction information was contained in CA23DG. Consistent with the Reviewer's intuition regarding visual confounds, we achieved high decoding accuracy for both egocentric boundary and egocentric goal direction in V1. As noted above, however, despite these strong visual confounds in the egocentric boundary condition, it was not possible to decode this information in the posterior EC or subiculum. This is consistent with the analysis presented above with regard to the inability to decode visual information in EC/subiculum.  (left, right, straight) and 50% for egocentric goal direction (left, right). Fig 2 is confusing. In a. it is important to re-state the conditions for making the judgement. E.g. that the correct answer relates to the fact the cue object is to one side of the global landmark. In c. the wall and red dot come across as a bit odd and distracting in the time-line. Seemed to me they make this more confusing than help. Most importantly the trial coding diagram leaves the reader unclear as to the full coding set up. In the right hand allocentric goal panel only two trials are shown (presumably to avoid the over-load of presenting all of them), but in each case why are 2 goals possible rather than 3 since in Fig 2a, 3 images are shown?

5.
We thank the Reviewer for bringing to our attention that Figure 2 could more clearly explain the task and trial coding. We have now included text in the Figure to describe the conditions used to make the allocentric goal direction judgement. We have now also removed the wall and the red dot from the timeline and have included links to videos of the trials (https://tinyurl.com/y3otgq7w). The reason that only two goals are demonstrated in the goal trial coding panel is because the goal object was located either to the right or the left of the path. There were, therefore, only two goal directions for which this trial could be coded. During the response phase (Figure 2a), however, we provided the option to respond to any of the three landmarks not visible in the virtual environment during the passive movement.

On line 366 the authors describe the method for determining the p-values. It
would be useful to reference another article using this method. It seems sensible, but would be useful to know who has taken a similar approach before with fMRI datasets.
We thank the Reviewer for raising this point. We used a Monte Carlo test procedure to determine the significance of our decoding accuracy (Besag, 1992;Hope, 1968).
We are unaware of this method being used with other fMRI datasets, but the bootstrap methods in which the rank of the observed data is compared to the null distribution make no assumptions as to the distribution of the data. In this respect, they can be more appropriate than comparing accuracies using, for example, a T-test against chance performance. We repeated the same analyses using the non-parametric Wilcoxon Sign-test and the effects were identical for allocentric boundary decoding, and almost perfectly replicated for the allocentric goal direction, with the exceptions of the anterior subiculum (p = 0.08) and CA23DG (p = 0.12).

When considering how the data here relates to previous studies it would worth
considering that the goal locations were never actually visited and walked around. It is unclear what impact this has, but it may have some.
We thank the Reviewer for raising this point and have now added to the Discussion to clarify that the goal direction representations observed in the current study may differ to those studies in which the subject actually visited the goal location. : "It should be noted, however, that our participants never actually visited the goal location which could have prevented the formation of a stable goal representation supported by, for example, hippocampal CA1 54-56 ".

Evidence for allocentric boundary and goal direction information in the human entorhinal cortex and subiculum
This study used a VR navigation training task in combination with a fMRI pattern analysis during a relative direction judgment task using passive observation of on-screen movement (through various paths in the same virtual environment).
The main regions of interest were EC and Subiculum (Sub), with control areas CA1, CA23DG, and PHG. The four "directions" analyzed were with respect to four distal landmarks (e.g., mountains, buildings) and four freestanding boundaries within the environment (two walls oriented along the axis defined by two opposing distal landmark, and the other two walls oriented 90-degrees from the first two, along the axis defined by the other two opposing landmarks). The main claim is that there is information about the allocentric direction of the boundaries with respect to the subjects and that there is information about the allocentric direction of the "goals" (cued by an object on the screen), but that these two signals originate from the posterior and anterior regions, respectively.
It is true that there are some conceptual replications of previous findings presented here, such as the general boundary-related activity in the human hippocampal subregions and the LEC/MEC (or anterior/posterior) distinction.
However, the paper also has some notable merits as well, such as the commendable efforts to decode allocentric boundary direction and to distinguish between directions of distal landmarks and local boundaries all in one task. The task and methods are generally very well thought out and the reported effects quite strong. Nevertheless, there are several serious concerns about the paper that I have summarized below. I hope that they help the authors revise and potentially even reframe the central claim of the paper to create a stronger paper in the end: 1. Issue of boundary direction: Border/boundary cells are defined by their response to environmental boundaries such as walls. Often, these cells are directionally tuned; however, the indication of the direction of a boundary is less essential to their characterization than their representation of the proximity of boundaries. Hence, the border score used to classify EC border cells is simply a comparison between firing distance from the wall with the maximum coverage of a field of any of the walls. This means that many border/boundary cells will respond not only to one wall of a quadrilateral-shaped environment. On the other hand, the BVC model does afford directional tuning of boundary cells, but even with a distance component to this model, most subicular boundary cells activate proximally to a boundary. Given this, the allocentric directional coding discussed in this paper is slightly unclear, with respect to the analogy the authors are making with the rodent brain. Although these issues are mentioned in the paper, it could be more directly tied to or contrasted with the study itself.
We thank the Reviewer for clarifying this point regarding the properties of boundary direction coding in the EC. In addition to the boundary vector cell model, the rationale for our study was also based on findings such as those reported by Solstad et al. (2008): in that study, 52/69 cells in MEC responded to a single wall in the environment; the remaining 17 had multiple fields. In a subsequent analysis, 12/22 border cells continued to show a boundary response to a newly inserted wall in the environment, and the direction of firing for these cells was maintained between the peripheral boundary wall, and the inserted boundary. These data support our assertion that it should be possible to observe allocentric boundary representations in EC. We have now added a sentence to the manuscript that acknowledges that a subset of border cells show directional modulation. We thank the Reviewer for raising this valid point as to whether our reported effects can be considered boundary-specific, or if they could be considered evidence instead for object-vector cells in humans. Using fMRI, we are unfortunately not able to test for boundary/object effects in real time as can be carried out in rodent electrophysiology.
It is true that the "boundariness" of a stimulus is a continuous and abstract property (Lever, Burton, Jeewajee, O'Keefe, & Burgess, 2009), and it can be difficult to determine the point at which one might expect an object to be classed, both 13 physically and psychologically, as a boundary. In our virtual environment, the length of the boundary was 40 virtual meters and its height 4 virtual meters. This is over 20 times as long as the height of the participant rendered in the environment, and twice as tall. These features are consistent with the definition of a boundary being an extended 3D surface (Lee, 2017), and even though our walls were isolated (i.e., not joined-up) it is similar to the stimuli used in Solstad et al. (2008) where it was demonstrated that border cells are active for walls even when discontinuous with respect to other environment boundaries. Furthermore, our walls impeded movement -a primary feature of boundaries as described previously in the literature (Buckley, Smith, & Haselgrove, 2015;Høydal, Skytøen, Andersson, Moser, & Moser, 2019;Lee, 2017;Lever et al., 2009;Solstad et al., 2008) -and participants spent the majority of their exploration time near the boundaries, which would have made it abundantly clear that they were impassable during learning in the fully-immersive virtual reality setup.
That these walls would still be considered as such even in a virtual world is supported by work showing that navigational behaviour during obstacle-avoidance is highly correlated between virtual and real-world setups (Fink, Foo, & Warren, 2007).
Distinguishing further our walls from objects, our boundaries did not change position throughout the entire experiment, meaning that that they are experientially different to a transient object. In human behavioural experiments, boundaries influence objectlocation memory relative to single objects (Negen, Sandri, Lee, & Nardini, 2018), and in fMRI extended boundaries, but not single objects, are associated with hippocampal activity (e.g., Doeller, King, & Burgess, 2008), with this boundary-related medial temporal lobe activity evident when having to imagine even a single boundary (Bird, Capponi, King, Doeller, & Burgess, 2010). Finally, in rodents there is no evidence that subicular BVCs respond to individual objects, and individual objects fail to control the position of place cell firing unless they are arranged to form a boundary (Cressant, Muller, & Poucet, 1997).
Given that there is no published evidence of object responses in BVCs, we are left in a situation where we would have to argue for different neural effects underpinning the same pattern of data in two different ROIs (i.e., above-chance decoding in posterior EC and posterior subiculum). In future studies it will be important to put these two different properties (object versus boundary) in more controlled opposition, but in the current manuscript we have now acknowledged the possible contribution of object vector cells to our decoding accuracy observed in the posterior EC.

Discussion (Lines: 615-631): "A recent study discovered object-vector cells in the rodent MEC that show an allocentric directional response to objects within the environment 51 . Given that both boundary and object-related responses are evident in MEC, and that object-vector cells respond also to boundary-like structures (i.e., elongated objects), it is not possible to distinguish whether our effects are driven by border and/or object-vector cells. Consistent with previous definitions of boundaries,
in the current study participants had experience that the walls impeded movement 10,18,51 , and they comprised an extended 3D surface 52 . Moreover, to our knowledge, there are no reports of object-related firing in the rodent dorsal subiculum.
While it is possible that object-vector cells contribute to the posterior EC effect observed in the current study, a more parsimonious explanation is that the walls were considered boundaries and that the decoding performance reflects border and boundary vector cell responses in the EC and subiculum, respectively. The distinction between object-vector and boundary responses in the EC awaits further clarification in humans."

3.Given the lack of significant differences across regions, the main finding is that anterior vs. posterior EC distinction. However, the theoretical interpretation here is not very clear. The introduction (Lines 41-47) reviews the what/where division of LEC/MEC, and then mentions that it has been modified to be about
the spatial vs. external sensory inputs (e.g., landmarks and prominent objects).

However, the paper that is cited (Knierim et al) does not make that claim,
actually; the contrast that is made in that paper is the difference between global and local. If that is the case, it is not very much in line with the current findings, in which the distal landmarks serve as orientation cues and the boundaries could, by some, be considered landmarks or objects. How could this be reconciled or clarified?
We thank the Reviewer for raising this concern and acknowledge that the way in which we described Knierim's work on the MEC versus LEC distinction was misleading. There is considerable evidence now to support a division of labour within the rodent EC, based upon the type of information-to-be represented (e.g., objects/scenes, what/where) or the reference frame in which the spatial information is encoded (allocentric/egocentric). Our analysis in which we segment the EC into posterior and anterior portions, therefore, is justified in its attempt to gain greater insight into this division of labour in humans. We have now reformulated the Background section of the manuscript to emphasise the motivation to segment our ROIs into posterior and anterior portions.

As the authors themselves say, the "directional" signals here can be differentiated from the representation of the landmarks themselves (or in the case of the boundaries, the association between the boundaries and the distal landmarks, perhaps). This is quite challenging for the interpretation about goal
directions. The lack of varying/counterbalancing of landmark directions (arrangement) should be explained or justified or avoid the interpretation that this is not a directional representation but an object or landmark-related signal.
The Reviewer is correct that while the allocentric boundary direction samples different distal landmark cues, the allocentric goal condition is very much synonymous with this information, meaning that we cannot rule out this interpretation in the current study.
The reason that we maintain the same configuration of the landmarks is so that participants continue to have a coherent understanding of the layout of the environment. Given the sparsity of the environment, if we were to change the configuration of landmarks during the experiment, participants would most likely become confused, meaning that we would lose trials through incorrect responses and subsequently it would be more difficult to code trials according to allocentric direction.
We have now clarified this in the Discussion. We thank the Reviewer for raising these potential confounds, and we have now amended the manuscript to address these points. We have now included the analysis of egocentric boundary and egocentric goal condition. As can be seen in 16 Supplementary Figure 11, only the anterior subiculum appears to contain information regarding egocentric goal direction; in the remaining EC/subiculum ROIs it is not possible to decode egocentric properties, suggesting that this property is not driving the allocentric effects reported in the manuscript. Furthermore, as can be seen in the trial schematic in Supplementary Figure 2, when decoding either according to allocentric goal or boundary, the egocentric information is sampled equally across different trials. This means that, after the trial averaging over runs, there is no pure egocentric information remaining for these conditions. 6. The results show highly significant differences among the directions (e.g., north vs. east) and an effect of the relative difference between two views (e.g., the difficulty with 270 degrees etc). What is the explanation for these differences? Is there some inherent bias in the environment? Is there any brain activity that differentiates these and and explains the neural correlates of the behavioral differences?
The Reviewer is correct that there are differences in RT for the allocentric goal direction task according to direction. Previous studies using judgment of relative direction tasks have demonstrated that participants tend to impose a reference frame on environments, and that this often comprises a conceptual North (Mou & McNamara, 2002). In our study, judgments regarding the allocentric goal direction relative to these axes are faster, and these data could suggest that participants imposed a North-South axis on the environment reflecting the Mountain -Cathedral (as can be seen in Supplementary Figure 7). Importantly, these differences in RT did not explain the decoding performance (see our response to Reviewer 1, point 1), and during learning participants did not show a preference for a particular landmark identity. Why the participants showed poorer performance during learning for the 270 degree angle disparity is unclear. This could in part reflect exploration such that participants explored the environment in a clockwise fashion, and that the information was recapitulated in this way, explaining the increased latency for 270 degree judgments.
We have now also added an interpretation of these effects: Results (Lines: 497-501): "These differences in RT may reflect participants forming a reference frame in the environment, with the Mountain and Cathedral providing a conceptual North-South axis. Consequently, responses to allocentric goal judgments in these directions may be facilitated."

A true boundary-cell-like representation would not distinguish between different walls (the yellow brick wall to the north from the red brick wall to the north). Therefore, it is important to compare between the different "North"
boundaries to see if their representations are very similar.
The Reviewer is correct in that border and boundary vector cells are insensitive to the identity of the boundary and therefore neural responses should generalise across the two different boundaries for each cardinal direction. To classify accurately, the linear SVM used here was required to extrapolate over boundary identity, otherwise it would not be possible to distinguish between North/South or East/West allocentric boundary directions due to them sharing the same boundary. Furthermore, the classifier used a one-versus-all approach, meaning that to classify a given direction, it would have to distinguish it from the opposing direction of the same boundary, and generalise over the directional information across the two different walls. Finally, given the drop-out and noise in anterior temporal lobe regions in fMRI, we averaged our trials over runs to boost the signal-to-noise ratio, meaning that we do not have individual responses per boundary side, but instead an averaged response that boosts the condition of interest (e.g., allocentric boundary direction), and reduces the influence of lower-level visual differences (e.g., boundary texture).

In humans, in particular, there is quite a bit of lateralization reported in MTL function. Are there any differences when comparing left vs right hemispheres?
We agree that lateralization in medial temporal lobe function has been reported in numerous studies (e.g., Bellmund, Deuker, Schroeder, & Doeller, 2016). To address this concern, we reran the decoding analysis with the medial temporal lobe masks separated by hemisphere and entered the resulting values into an ANOVA comprising the factors Hemisphere, ROI, Anterior/posterior section and Condition. Although there was a significant ROI ´ Anterior/posterior interaction (F(1, 27) = 4.73, p = 0.039), there was no evidence of any effect of, or interaction with Hemisphere (ps > 0.81). 9. The "imagined" part of this task (which is the phase at which the fMRI signal is extracted) seems like an important aspect. Perhaps this should be highlighted from the start.

18
We thank the Reviewer for this comment. Unlike previous experiments in which participants have been actively instructed to imagine a trajectory (Horner, Bisby, Zotow, Bush, & Burgess, 2016) or the direction between two landmarks (Bellmund et al., 2016), we did not instruct our participants to imagine the boundaries during the presentation of the blank screen. The reason for taking the period of the trial in which there was no stimulus input was to match perceptual features across different classes of stimuli, and not to demonstrate evidence of a more abstract boundary representation. Consequently, we do not want to place too much emphasis on this "imagined" component, but we highlight clearly in the Methods that we use the portion of the trial in which there was no visual input.
Other comments: 10. The authors mention that this is the first demonstration of allocentric boundary direction representation. It seems important to clearly define what they mean by this and how this particular task design aims to distinguish this.

Some readers may not understand and, even if they do, they may not follow the logic of how this is not distinguished in other studies).
We thank the Reviewer for this comment, and have now added to the 'Background' We thank the Reviewer for this comment. Our anatomical ROIs were informed by the rodent literature, where we had clear predictions as to the involvement of the EC and subiculum in boundary coding, particularly with regard to the posterior extent of both of these regions. Consequently, we focussed our analyses in these medial temporal areas. In response to concerns from Reviewer 3 (point 2), we have removed the ROI ´ Anterior/posterior ´ Condition ANOVA (formerly Figure 5), and present our analysis by ROI only (e.g., Figure 4 -EC and subiculum). Regarding the PHC and CA1, additional analyses in response to Reviewer 3 (point 3) where we erode the masks to reduce the influence of neighbouring brain regions led to the effects in these ROIs no longer being significant; in contrast the effects in EC and subiculum remained relatively unchanged. We acknowledge that our effects are relatively small, and it is likely that other brain regions contribute to this boundary processing. Our statistical tests, however, were used to assess whether the information in a given ROI exceeded that expected by chance, not that decoding performance was significantly larger than would be expected in a different brain region. Consistent with example of grid cell-like representations in humans not being limited to the EC, we would not argue that EC and subiculum are the only regions in which one might observe boundary coding.

Line 331-335: the averaging over three runs. Explain any effects across runs.
We thank the Reviewer for allowing us the opportunity to clarify this point. The data were averaged over the three runs to increase signal-to-noise ratio and therefore we did not analyse data per run; the analysis comprised the 96 averaged samples (i.e., 276 trials / 3 runs).

to-boundary' signal in humans other than the Shine et al study under consideration here. They do not study distance to boundary, but this is understandable in an initial study, and the Mosers, for instance, would say that nearly all the boundary cells in the entorhinal cortex fire directly adjacent to the
boundary. This is all to say that the current findings are novel, and reveal a signal that is arguably a more direct human analogue of the boundary cells found in rodent exploration than previous findings. This, in my view, is the paper's main claim to originality and significance. As mentioned above, the design allows analysis of separable direction-to-boundary and direction-to-goal signals, which offers an advance over the interesting study of Chadwick et al, We thank the Reviewer for raising this concern and, as suggested, have removed this analysis. The Reviewer is correct in that the anterior/posterior split that we have used is more basic than that used previously with 7T data (e.g., Maass, Berron, Libby, Ranganath, & Düzel, 2015), but this stems in part from the difficulty in implementing this segmentation at lower resolutions. As noted below in response to your comment number 3, although our in-plane resolution was the same as has been implemented previously at higher field strengths, our slice thickness was larger. This means that the transitional slices between anterior-lateral and posterior-medial EC are more likely to 24 contain a mixture of anatomical information from these subregions. Additional control analyses in which we eroded the masks (see response to your point number 3) to limit the influence of neighbouring ROIs, however, suggest that the results are robust, and the effects we present in the manuscript are not simply an artifact of the segmentation procedure employed. The Reviewer raises an important point regarding the specificity of our ROIs. We would first like to note that although the segmentation protocol used in the current manuscript was established for 7T MRI, the only difference between the resolution used for segmentation in our 3T study lies in the slice thickness; we used a slice 1.5mm thickness whereas the cited protocol used 1mm-thick slices. Importantly, for identifying anatomical boundaries within individual hippocampal subregions, the in-plane resolution is the same in both the protocol and our study (i.e., 0.4*0.4mm).

25
The Reviewer is correct that there is always a cost in anatomical precision whenever moving between the resolutions used for structural versus functional imaging, and this may lead to partial-volume effects, with voxels containing a mixture of signals from different anatomical regions. Although this situation is, at the current resolutions available, unavoidable in functional imaging, we tried to mitigate the 'leakage' of signal by performing the analysis on unsmoothed data, which we have also now clarified in the Methods (lines 328-330). Although we appreciate the Reviewer's point that weightmaps may provide more anatomical information and help resolve the issue of leakage, currently it is suggested that the presentation of weight maps is not optimal. It could be the case that a voxel with a large weight reflects the removal of a noise signal in the data allowing for the extraction of smaller, but more meaningful signal (Haufe et al., 2014). As a result, highly significant voxels may not actually reflect the neural computations of interest. Also, given the typically limited number of trials available in neuroscience research, many different brain maps will give rise to similar predictive outcomes (Varoquaux & Thirion, 2014). Furthermore, the choice of weight map to be visualized is difficult. As we use nested cross-validation, it is not entirely clear the correct approach to generate a summary weight map, given that each fold of the crossvalidation will result in a slightly different model. It is for these reasons that we have chosen not to report the weight maps. Rather, to address the question of leakage more directly, we have re-analysed the data by eroding the masks to reduce the influence of neighbouring anatomical structures. We thank the Reviewer for providing us with the opportunity to clarify the methods used to determine the significance level in our manuscript. In short, we carry out two different bootstrap procedures. The first is to demonstrate the distribution of our sample, the second is to determine the p-value.
The group-level decoding accuracy reflects the mean average decoding across all of our participants. As the Reviewer rightly points out, this is referred to as both "group-level decoding accuracy" and "group mean decoding accuracy". For consistency, in the manuscript we now just refer to this value as "group-level decoding accuracy". This is a single value, which is represented by the vertical black line in Figure 4. The pale blue histogram in these figures reflects the distribution of 10,000 means resulting from bootstrap resampling from the group's individual decoding accuracies. This was computed to demonstrate more clearly the distribution of the sample, rather than simply providing a single value representing the group-level decoding accuracy. The second bootstrap procedure was a Monte Carlo significance test used to determine the p-value associated with the group-level decoding accuracy, and to do this we first needed to generate a null distribution centred around chance performance. Accordingly, we subtracted the group-level decoding accuracy from every individual participant's decoding score (i.e., demeaning the sample) before adding to each participant's demeaned score chance performance (i.e., 25% when decoding four classes). This resulted in the group's decoding scores maintaining the same variance, but with a mean centred on chance. We then sampled from the null distribution 10,000 times and observed how many times the group-level decoding accuracy drawn from this null distribution exceeded the observed mean decoding (i.e., the vertical black line), and divided this number by the number of bootstrap permutations (i.e., 10,000) to obtain our p-value. Importantly, 1 is added to both the numerator and denominator of this calculation to correct for cases in which none of the null values exceed the mean decoding accuracy. We have now clarified this information in the Methods (Lines: 403-424). We thank the Reviewer for providing us the opportunity to explain in more detail the analysis methods used to generate the betas for the decoding analysis. First, with regards to the decoding of visual information, please see below our response to your point number 6. Second, the Reviewer is correct that there are a number of different models that can be used to generate the betas, which have been outlined previously 32 (Mumford, Turner, Ashby, & Poldrack, 2012). Some researchers have chosen to use separate GLMs for different trials convolved with the haemodynamic response function (HRF) to generate individual beta images. An alternate, and potentially more parsimonious approach in which no prior information regarding the shape of the HRF is fed into the GLM, is to fit an unconvolved boxcar regressor spanning a number of TRs around 4-6 seconds after the event of interest. This so-called "Add" model performed well in both simulations and with real data with short inter-stimulus intervals, which was attributed to its tolerance to the variability of the HRF (Mumford et al., 2012).

Visual after-effects and
For our analysis, we used an "Add" model, and mirrored the analysis methods of Bellmund et al. (2016) in which multivariate analysis methods were used to examine the neural response in the entorhinal cortex during an fMRI spatial navigation task. Bellmund et al. (2016) first regressed out the movement parameters from their data before fitting an "Add" model in the residuals resulting from this regression. Consistent with previous research, in a bid to boost the signal-to-noise ratio for the decoding analysis (Isik, Meyers, Leibo, & Poggio, 2013;Nau, Navarro Schröder, Bellmund, & Doeller, 2018), we then averaged the individual trial-specific betas over the three runs and used these resulting estimates for our decoding analyses with nested crossvalidation. We have now provided more details in the Methods section and included Supplementary Figure 5 that outlines more clearly this analysis pipeline.
Data analysis (Lines: 361-372): "Given its high performance in decoding using eventrelated functional imaging data with short inter-stimulus intervals, the "Add" 37 model was implemented here. This model aims to capture the putative peak of the haemodynamic response function occurring 4-6 seconds after the onset of the event of interest. Since we wanted to capture activity associated with the stationary period of the trial, which occupied the period 2-6 seconds after trial onset (see Figure 2C), we took the estimates from an unconvolved boxcar regressor that spanned three TRs occurring 4-6 seconds after the stationary phase 36,38 (i.e., 6-12 seconds after trial onset), in separate models comprising one regressor representing the trial of interest, and a second regressor modelling all other trials in the scan run" 6. Again justification is one thing, but what might be particularly reassuring is to add an analysis from a visual control region (such as V1), even if, given the relatively small slab of brain that they sampled, they cannot sample the whole region. It would be worrying for the interpretation of the results if visual cortex were able to outperform the MTL regions of interest. Simply put, every single region they have looked at, if we ignore multiple corrections for a moment, produces either a significant boundary direction or goal direction signal. It would be reassuring to report results for an additional control region or two.
As the Reviewer correctly points out, we have only a small slab of brain from which to choose a control region, and recent evidence suggests that regions such as V1, which were previously thought to have little involvement in navigation, also show spatiallymodulated responses in the absence of visual input in the rodent brain (Pakan, Currie, Fischer, & Rochefort, 2018) which even includes positional signals (Saleem, Diamanti, Fournier, Harris, & Carandini, 2018). Furthermore, there is increasing evidence of a dynamic interplay between the medial temporal lobe and V1 as evidenced in recent fMRI work (Hindy, Ng, & Turk-Browne, 2016). It becomes difficult, therefore, to be entirely sure that an ROI should show no response to a given task manipulation.
Critically, however, our control analysis on egocentric boundary direction provides reassurance that we are not simply decoding lower-level visual features in EC and subiculum (please see above the response to Reviewer 1, point 2).
Specifically, when explicitly introducing a visual confound by classifying trials according to the position of the boundary in the visual field, we could decode egocentric boundary direction in V1. In contrast, decoding accuracy was at chance for both EC and subiculum.
With regards to a significant boundary direction and/or goal direction signal being evident in every ROI, additional analyses highlighted in Supplementary Figures   8 and 10 demonstrate that although our EC and subiculum effects remain relatively consistent using the eroded masks, it is no longer possible to decode either allocentric property in parahippocampal cortex or CA1. These analyses suggest that the effects observed in key EC and subiculum ROIs are robust and not an artifact of our analysis method. We thank the Reviewer for these helpful suggestions. In the legend to Figure 1a we have now included a key making clear that N, S, E, and W refer to the boundary directions, and have changed the proportions of the schematic to reflect the square environment. We have increased the size of Figure 1B and adjusted the legend in 1C and the Methods to highlight that there were four global landmarks. Furthermore, we have adjusted the schematic of the environment to make it square. We thank the Reviewer for this comment. We have now increased the number of decimal places in the reporting of the p-value so that the precise number is reported. 9. The behavioural performance paragraphs on pages 20-21 are hard to read.

Minor
They would be better as tables or graphs. There should be some interpretation of the results, regarding faster reaction times for certain judgements, e.g. for boundaries and goals to the North.
We thank the Reviewer for this comment and note that the accuracy data are displayed in Figure 3C, and the reaction time data in Supplementary Figure 7. Given that the participants performed the allocentric goal direction task, we have provided a brief interpretation for this effect in the Results section. Specifically, in judgment of relative direction tasks, it has been demonstrated that participants impose a reference frame when encoding positional information, and that this is often aligned with geometric cues of the environment, such as room structure, or initial facing direction (Mou & McNamara, 2002). Participants may have interpreted the mountain as a conceptual North, which may have facilitated reaction times for allocentric goal judgements relative to this North-South axis. Importantly, during learning there was no evidence of a landmark preference , and there was no evidence that classifier performance was modulated by these differences in RT (see response to Reviewer 1, comment 1). : 497-503): "These differences in RT may reflect participants forming a reference frame in the environment, with the Mountain and Cathedral providing a conceptual North-South axis. Consequently, responses to allocentric goal judgments in these directions may be facilitated. Consequently, responses to allocentric goal judgments in these directions may be facilitated 48,49 . Importantly, however, these differences in RT did not influence subsequent decoding performance (see Supplementary Information)". We have now added a brief discussion of this work. We would like to note that although Hodgetts et al. (2017) report scene-selectivity in anterior subiculum, our results do not argue against a univariate scene-preference in anterior subiculum, rather that the 36 multivariate signal in this region is not informative regarding allocentric boundary direction. ,: "Scene-specific responses have been reported also in the human anterior subiculum 64 . Although these data may seem at odds with our posterior subiculum boundary effects, it is possible that anterior subiculum shows a univariate scene response, whereas the multivariate pattern in posterior subiculum is informative of allocentric boundary information in the absence of greater scene-related activity. Future studies will be necessary to elucidate the nature of scene-sensitivity in the subiculum, and the precise perceptual features driving these effects." 11. P27 -the authors should mention other functions for boundaries than error correction for path integration, e.g. defining where objects are relative to

boundaries -see TMS work of Julian and Epstein in Current Biology.
We thank the Reviewer for this suggestion and have now incorporated this work in the Background section.
Background : "In humans, boundaries have been shown to be behaviourally salient, aiding reorientation 13 , and being used to define object locations 20,: "Furthermore, the occipital place area has been shown to be causally involved in memory for object locations relative to boundaries but not landmarks"