Anchoring of grid fields selectively enhances localisation by path integration

Reviewer #1 (Public Review):

In this study, Clark et. al. uncovered an association between the positional encoding of grid cell activity with good performance in spatial navigation tasks that requires path integration, highlighting the contribution of grid firing to behavior. Using electrophysiology approaches, the authors measured MEC neuron activity while mice performed a spatial memory task in one-dimensional (1D) virtual tracks, where the mice must stop in a specific reward zone for a reward. Individual trials either had a visual cue at the reward location (beaconed trials), no cue at the reward location (non-beaconed trials), or no cue and no reward regardless of stopping (probe trials). The authors identified that grid cells could encode track position or distance traveled, which were distinguished on a per-session or per-trial basis by calculating whether cell firing was periodic with respect to track length (firing at the same location on each trial) or periodic with respect to distance traveled (firing locations drift across trials), respectively. While some behavioral sessions had stable coding of either position or distances, other sessions exhibited coding schemes that switched between these two modes. The behavioral performance in beaconed trials was comparable when grid cells showed position or distance coding. In contrast, mice perform better on non-beaconed trials when grid cells showed position coding. The authors concluded that position coding in grid cells may enhance performance when tasks require path integration (non-beaconed condition).

The conclusions of this paper are mostly well supported by data, the finding about the association between grid cell encoding and behavior in spatial memory tasks is important. However, some aspects of the analysis need to be clarified or extended.

1. While the current dataset aims to demonstrate a "correlation" between grid cell encoding and task performance, the other variables that could confound this correlation should be carefully examined.
(1) The exact breakdown of the fraction of beaconed/non-beaconed/probe trials is never shown. if the session makeup has a significant effect on the coding scheme or other results, this variable should be accounted for.
(2) The manuscript did not provide information about whether individual mice experienced sessions with different combinations of the three trial types, and whether they show different preferences in position or distance encoding even in comparable sessions. This leads to the question of whether different behavior and activity encoding were dominated by experimental or natural differences between individual mice. Presenting the data per mouse will be helpful.
(3) Related to the above point, in Figure 5, the mice appeared to behave worse in probe trials than non-beaconed trials. If the mouse did not know if a trial is a probe or a non-beacon trial, they should behave equivalently until the reward location and thus should stop an equal amount. If this difference is because multiple probe trials are placed consecutively, did the mouse learn that it will not get a reward and then stop trying to get rewards? Did this affect switching between position and distance coding?
(4) It is not shown how the behaviors (e.g., running speed away from the reward zone, licking for reward) in beaconed/non-beaconed/probe trials were different and whether the difference in behaviors led to the different encoding schemes.
2. Regarding the behavior and activity encoding on a trial-by-trial basis, did the behavioral change occur first, or did the encoding switch occur first, or did they happen within the same trial? This analysis will potentially determine whether the encoding is causal for the behavior, or the other way around.
3. The author determined that the grid cell coding schemes were limited to distance encoding and position encoding. However, there could be other schemes, such as switching between different position encodings (with clear spatial fields but at different locations), as indicated by Low et. al., 2021, and switching between different distant encodings (with different distance periods). If these other schemes indeed existed in the data, they might contribute to the variation of the behaviors.
4. The percentage of neurons categorized in each coding scheme was similar between non-grid and grid cells. This implies that non-grid cells might switch coding schemes in sync with grid cells, which would mean the whole MEC network was switching between distance and position coding. This raises the question of whether the grid cell coding scheme was important per se, or just the MEC network coding scheme.
5. In Figure 2 there are several cell examples that are categorized as distance or position coding but have a high fraction of the other coding scheme on a per-trial basis. Given this variation, the full session data in F should be interpreted carefully, since this included all cells and not just "stable" coding cells. It will be cleaner to show the activity comparison only between the stable cells.
6. The manuscript is not well written. Throughout the manuscript, there are many unexplained concepts (especially in the introduction) and methods, mis-referenced figures, and unclear labels.

Reviewer #2 (Public Review):

Clark and Nolan's study aims to test whether the stability of grid cell firing fields is associated with better spatial behavior performance on a virtual task. Mice were trained to stop at a rewarded location along a virtual linear track. The rewarded location could be marked by distinct visual stimuli or be unmarked. When the rewarded location was unmarked, the animal had to estimate its distance run from the beginning of the trial to know where to stop. When the mouse reached the end of the virtual track, it was teleported back to the start of the virtual track.

The authors found that grid cells could fire in at least two modes. In the "virtual position" mode, grid firing fields had stable positions relative to the virtual track. In the "distance run" mode, grid fields were decoupled from the virtual cues and appeared to be located as a function of distance run on the running wheel. Importantly, on trials in which the rewarded location was unmarked, the behavioral performance of mice was better when grid cells fired in the "virtual position" mode.

This study is very timely as there is a pressing need to identify/delimitate the contribution of grid cells to spatial behaviors. More studies in which grid cell activity can be associated with navigational abilities are needed. The link proposed by Clark and Nolan between "virtual position" coding by grid cells and navigational performance is a significant step toward better understanding how grid cell activity might support behavior. It should be noted that the study by Clark and Nolan is correlative. Therefore, the effect of selective manipulations of grid cell activity on the virtual task will be needed to evaluate whether the activity of grid cells is causally linked to the behavioral performance on this task. In a previous study by the same research group, it was shown that inactivating the synaptic output of stellate cells of the medial entorhinal cortex affected mice's performance of the same virtual task (Tennant et al., 2018). Although this manipulation likely affects non-grid cells, it is still one of the most selective manipulations of grid cells that are currently available.

When interpreting the "position" and "distance" firing mode of grid cells, it is important to appreciate that the "position" code likely involves estimating distance. The visual cues on the virtual track appear to provide mainly optic flow to the animal. Thus, the animal has to estimate its position on the virtual track by estimating the distance run from the beginning of the track (or any other point in the virtual world).

It is also interesting to consider how grid cells could remain anchored to virtual cues. Recent work shows that grid cell activity spans the surface of a torus (Gardner et al., 2022). A run on the track can be mapped to a trajectory on the torus. Assuming that grid cell activity is updated primarily from self-motion cues on the track and that the grid cell period is unlikely to be an integer of the virtual track length, having stable firing fields on the virtual track likely requires a resetting mechanism taking place on each trial. The resetting means that a specific virtual track position is mapped to a constant position on the torus. Thus, the "virtual position" mode of grid cells may involve 1) a trial-by-trial resetting process anchoring the grid pattern to the virtual cues and 2) a path integration mechanism. Just like the "virtual position" mode of grid cell activity, successful behavioral performance on non-beaconed trials requires the animal to anchor its spatial behavior to VR cues.

One main conclusion of this study is that better performance on the VR task was observed when the grid cells were anchored to the reference frame that was the most behaviorally relevant.

Reviewer #3 (Public Review):

This study addresses the major question of 'whether and when grid cells contribute to behavior'. There is no doubt that this is a very important question. My major concern is that I'm not convinced that this study gives a significant contribution to this question, although this study is well-performed and potentially interesting. This is mainly due to the fact that the relation between grid cell properties and behavior is exclusively correlative and entirely based on single cell activity, although the introduction mentions quite often the grid cell network properties and dynamics. In general, this study gives the impression that grid cells exclusively support the cognitive processes involved in this task. This problem is in part related to the text. However, it would be interesting to look at the population level (even beyond grid cells) to test whether at the network level, the link between behavioral performance and neural activity is more straightforward compared to the single-cell level. This approach could reconcile the present results with those obtained in their previous study following MEC inactivation.

The authors used a statistical method based on the computation of the frequency spectrum of the spatial periodicity of the neural firing to classify grid cells as 'position-coding' (with fields anchored to the virtual track) and 'distance-coding' (with fields repeating at regular intervals across trials). This is an interesting approach that has nonetheless the default to be based exclusively on autocorelograms. It would be interesting to compare with a different method based on the similarities between raw maps. Beyond this minor point, cell categorization is performed using all trial types. Each trial type (i.e. beacon or non-beacon) is supposed to force mice to use different strategies and should induce different spatial representations within the entorhinal-hippocampal circuit (and not only in the grid cell system). In that context, since all trials are mixed, it is difficult to extrapolate general information. On page 5 the authors state that 'Since only position representations should reliably predict the reward location, ..., we reasoned that the presence of positional coding could be used to assess whether grid firing contributes to the ongoing behaviour'.

I do not agree with this statement. First of all, position coding should be more informative only in a cue-guided trial. Second, distance coding could be as informative as position coding since at the network level may provide information relevant to the task (such as distance from the reward). This possibility is not tested here. Third, position-coding is interpreted as more relevant because it predominates in correct trials. However, this does not imply that this coding scheme is indeed used to perform correct trials. It could be more informative to push forward the correlative analysis by looking at whether behavioral performance can be predicted by the coding scheme on a trial-by-trial basis. This analysis would not provide a causal relation between cell activity and behavior, but could strengthen the correlation between the two.

Author Response

Reviewer #1 (Public Review):

1. While the current dataset aims to demonstrate a "correlation" between grid cell encoding and task performance, the other variables that could confound this correlation should be carefully examined.

(1) The exact breakdown of the fraction of beaconed/non-beaconed/probe trials is never shown. if the session makeup has a significant effect on the coding scheme or other results, this variable should be accounted for.

(2) The manuscript did not provide information about whether individual mice experienced sessions with different combinations of the three trial types, and whether they show different preferences in position or distance encoding even in comparable sessions. This leads to the question of whether different behaviour and activity encoding were dominated by experimental or natural differences between individual mice. Presenting the data per mouse will be helpful.

(3) Related to the above point, in Figure 5, the mice appeared to behave worse in probe trials than non-beaconed trials. If the mouse did not know if a trial is a probe or a non-beacon trial, they should behave equivalently until the reward location and thus should stop an equal amount. If this difference is because multiple probe trials are placed consecutively, did the mouse learn that it will not get a reward and then stop trying to get rewards? Did this affect switching between position and distance coding?

(4) It is not shown how the behaviours (e.g., running speed away from the reward zone, licking for reward) in beaconed/non-beaconed/probe trials were different and whether the difference in behaviours led to the different encoding schemes.

We appreciate these suggestions and will add all of the requested analyses in a revised manuscript. We note here that while the proportion of trial types differed between sessions, in all sessions trial types were varied in a repeating sequence, so blocks of behaviour where grid firing is anchored (or not anchored) to the track coordinates can not be explained as a consequence of a particular trial type. We will make this clearer in a revised manuscript.

1. Regarding the behaviour and activity encoding on a trial-by-trial basis, did the behavioural change occur first, or did the encoding switch occur first, or did they happen within the same trial? This analysis will potentially determine whether the encoding is causal for the behaviour, or the other way around.

We agree this is an important point and the corresponding analyses will be reported in a revised manuscript.

1. The author determined that the grid cell coding schemes were limited to distance encoding and position encoding. However, there could be other schemes, such as switching between different position encodings (with clear spatial fields but at different locations), as indicated by Low et. al., 2021, and switching between different distant encodings (with different distance periods). If these other schemes indeed existed in the data, they might contribute to the variation of the behaviours.

We did not observe switching between coding schemes of the same type within our dataset and so did not document this. We agree it is important to do so and will provide additional analyses in the revised manuscript

1. The percentage of neurons categorised in each coding scheme was similar between non-grid and grid cells. This implies that non-grid cells might switch coding schemes in sync with grid cells, which would mean the whole MEC network was switching between distance and position coding. This raises the question of whether the grid cell coding scheme was important per se, or just the MEC network coding scheme.

We appreciate the suggestion and very much agree that looking at cells outside of just grid cells is important in determining which cells are functionally relevant in spatial behaviours. We will provide additional analyses in a revised manuscript.

1. In Figure 2 there are several cell examples that are categorised as distance or position coding but have a high fraction of the other coding scheme on a per-trial basis. Given this variation, the full session data in F should be interpreted carefully, since this included all cells and not just "stable" coding cells. It will be cleaner to show the activity comparison only between the stable cells.

We agree that showing stable examples before introducing examples that switch on a per-trial basis will be helpful. We will amend this in a revised manuscript.

1. The manuscript is not well written. Throughout the manuscript, there are many unexplained concepts (especially in the introduction) and methods, mis-referenced figures, and unclear labels.

We appreciate the feedback and will work to address the concerns in a revised manuscript.

Reviewer #2 (Public Review):

This study is very timely as there is a pressing need to identify/delimitate the contribution of grid cells to spatial behaviors. More studies in which grid cell activity can be associated with navigational abilities are needed. The link proposed by Clark and Nolan between "virtual position" coding by grid cells and navigational performance is a significant step toward better understanding how grid cell activity might support behavior. It should be noted that the study by Clark and Nolan is correlative. Therefore, the effect of selective manipulations of grid cell activity on the virtual task will be needed to evaluate whether the activity of grid cells is causally linked to the behavioral performance on this task. In a previous study by the same research group, it was shown that inactivating the synaptic output of stellate cells of the medial entorhinal cortex affected mice's performance of the same virtual task (Tennant et al., 2018). Although this manipulation likely affects non-grid cells, it is still one of the most selective manipulations of grid cells that are currently available.

We appreciate this additional context provided here. In our view, it is critical to narrow down the space of possible behaviours that grid cells might contribute to. As the reviewer notes, our previous work provided evidence that speaks to this question by targeting genetic manipulations (Tennat et al., 2018), but while this approach was specific to stellate cells it does not discriminate grid from non-grid cells and so does not tell us specifically about roles for grid cells. As far as we are aware there is currently no manipulation that will do this. In the experiments here, we take a complementary approach, leveraging the variability inherent in behaviour and the fact that in our location memory task animals will perform many trials in a session. By showing that spatially anchored grid firing does not predict behavioural success on cued trials, but does predict success on trials that are solved by path integration, we substantially narrow the space of behaviours that grid cells could contribute to. Importantly, stellate cells appear necessary for both cued and uncued behaviour in the task (Tennant et al., 2018), suggesting that their roles are more general than the grid cell population, which is likely to be only a subset of stellate cells. We will more carefully address this point in a revised manuscript.

When interpreting the "position" and "distance" firing mode of grid cells, it is important to appreciate that the "position" code likely involves estimating distance. The visual cues on the virtual track appear to provide mainly optic flow to the animal. Thus, the animal has to estimate its position on the virtual track by estimating the distance run from the beginning of the track (or any other point in the virtual world).

We agree this terminology has the potential for causing confusion. A simpler descriptive definition would be track-anchored and track-independent rather than position and distance coding. We will consider this and other alternatives for a revised manuscript.

Reviewer #3 (Public Review):

This study addresses the major question of 'whether and when grid cells contribute to behaviour'. There is no doubt that this is a very important question. My major concern is that I'm not convinced that this study gives a significant contribution to this question, although this study is well-performed and potentially interesting. This is mainly due to the fact that the relation between grid cell properties and behaviour is exclusively correlative and entirely based on single cell activity, although the introduction mentions quite often the grid cell network properties and dynamics. In general, this study gives the impression that grid cells exclusively support the cognitive processes involved in this task. This problem is in part related to the text. However, it would be interesting to look at the population level (even beyond grid cells) to test whether at the network level, the link between behavioural performance and neural activity is more straightforward compared to the single-cell level.

We appreciate the feedback and suggestions. As we note in our response to Reviewer #2, there is currently no method for selective manipulation of grid cells, while testing correlation is a critical step on the path to establishing causation. Our study contributes by reducing the space of possible functions of grid cells to exclude behaviours in which local cues are available, while providing evidence for a clear relationship between anchoring of grid cells and successful outcomes when path integration is used for localisation. We’re unclear here about what the reviewer means by ‘more straightforward’ as the relationships we establish do not appear overly complicated, and as strong relationships between activity of single grid cells and populations of grid cells are already well established (Gardner et al., 2021; Waaga et al., 2021; Yoon et al., 2013).

The authors used a statistical method based on the computation of the frequency spectrum of the spatial periodicity of the neural firing to classify grid cells as 'position-coding' (with fields anchored to the virtual track) and 'distance-coding' (with fields repeating at regular intervals across trials). This is an interesting approach that has nonetheless the default to be based exclusively on autocorrelograms. It would be interesting to compare with a different method based on the similarities between raw maps.

We’re not sure we understand the point here. The manuscript provides analyses comparing rate maps for activity periods in which grid cells are / are not anchored to the task environment (e.g. Figure 2A-C, Figure 3B-E); when grid cells are anchored the rate maps are clearly spatial, when they are not anchored we show that spatial information (in the track reference frame) is very substantially reduced.

Beyond this minor point, cell categorization is performed using all trial types. Each trial type (i.e. beacon or non-beacon) is supposed to force mice to use different strategies and should induce different spatial representations within the entorhinal-hippocampal circuit (and not only in the grid cell system). In that context, since all trials are mixed, it is difficult to extrapolate general information.

Again, we’re not sure we understand the point. We appreciate this likely reflects a lack of clarity on our part in the writing of the manuscript. As noted in our response to Reviewer #1, we will include additional details about the organisation of trials and relationships between trials, behavioural outcomes and neural codes observed. We should note here that mice are not ‘forced’ to adopt any particular strategy. Rather, on uncued trials a path integration strategy is the most efficient way to solve the task. Mice could instead use a less efficient strategy of stopping at short intervals and still obtain rewards, although the behavioural evidence suggests they do not choose to do this after learning the task.

On page 5 the authors state that 'Since only position representations should reliably predict the reward location, ..., we reasoned that the presence of positional coding could be used to assess whether grid firing contributes to the ongoing behaviour'. I do not agree with this statement. First of all, position coding should be more informative only in a cue-guided trial. Second, distance coding could be as informative as position coding since at the network level may provide information relevant to the task (such as distance from the reward).

Again, this point perhaps reflects a lack of clarity on our part in writing the manuscript. When grid cells are anchored to the track reference frame (position encoding in the manuscript), then the location of the rate peaks in grid firing is reliable from trial to trial. This is the case whether or not the trial is cued. When grid cells are independent of the track reference frame (distance encoding in the manuscript, but we now appreciate this is a poor choice of words), then the location of the firing rate peaks vary from trial to trial; thus position can not be read out directly from trial to trial. In principle, when grid cells are not anchored to the track the mouse could read out track position by storing the grid network configuration at the start of each trial and then subtracting this from readouts of distance as mice move along the track. If mice do use this computation we would expect them to do so equally well on cued and uncued trials, whereas our results clearly show a dissociation between trial types in the relationship between grid firing and behavioural outcome. We will highlight this possibility in a revised manuscript.

Third, position-coding is interpreted as more relevant because it predominates in correct trials. However, this does not imply that this coding scheme is indeed used to perform correct trials.

As we note above, our analyses reduce the space of behaviours to which grid cells might contribute, by providing evidence that anchoring of grid firing is associated with successful outcomes specifically when mice adopt a path integration strategy. We agree that alternative models remain plausible, for example perhaps the behaviourally relevant computations are implemented elsewhere in the brain with grid anchoring to the track as an indirect consequence. Nevertheless, the space of alternative models is substantially reduced given our experiments and analyses, while our approach complements tests of grid-behaviour functions that rely on manipulations which leave open alternative explanations based on off target effects. We expect that inclusion in a revised manuscript of the further analyses suggested above should provide further tests of the grid-behaviour relationship.

It could be more informative to push forward the correlative analysis by looking at whether behavioural performance can be predicted by the coding scheme on a trial-by-trial basis.

Figure 5E shows the recommended analysis.