Distance Mapping and Variable-Specific Geometry of Goal-Relevant Frames in the Retrosplenial Cortex

Reviewer #1 (Public review):

Summary:

Chen and colleagues utilize in vivo electrophysiology to characterize distance-encoding neurons in the retrosplenial cortex of rats during both random foraging and goal-oriented navigation. They observe a subset of RSC neurons that encode distance to a hidden goal location and that HD coding is enhanced during goal-directed navigation when compared to foraging. They also demonstrate that goal distance coding is preserved in the dark. Distance encoding has been shown in RSC in prior publications, but examining it with respect to a behaviorally relevant location will be of interest to the field. That said, the manuscript needs substantially more methodological detail before I am convinced that goal-distance-to-goal (DTG) coding is not an artifact of self-motion or of other spatial tuning already known to exist in the area. The authors also introduce several new machine-learning approaches that are hard to interpret without clearer justification or a demonstrated need. Addressing the points below would provide more convincing evidence for DTG coding and enhance readability.

Strengths:

The task is useful for determining whether the goal distance is encoded in neural populations.

Retrosplenial cortex is an excellent candidate region for examining representations related to goal distance.

The analytical framework is state-of-the-art and is useful for determining the contribution of goal distance to complex activation in retrosplenial cortex that possesses mixed selectivity.

Weaknesses:

The analyses/simulations intended to ascertain the relationship between other known spatial/self-motion codes in RSC and DTG coding are insufficient. I struggle with how DTG and speed can be convincingly disentangled given task structure. I suspect many DTG cells are in fact speed-modulated, and that if the task was flipped such that the animal had to run through the goal location rather than stop, DTG tuning curves would be mirrored. There are a couple of simple things that could help:

(1) Show significantly more examples of DTG neurons (perhaps all of them) alongside their corresponding spatial (e.g., EBC and HD) and self-motion tuning curves (e.g., speed and angular speed) for multiple goal locations.

(2) The relationship between the DTG tuning curve and the speed tuning curve should be presented.

(3) Report what fraction of DTG cells are tuned to a random, non-goal location, what fraction qualify as DTG by chance, and how much better goal decoding is than decoding to a random location (Figure 2a).

(4) We need visualizations of DTG reliability both within and across goal locations (see below).

The GLM analyses are meant to address the unique contribution of distance to goal, but there are issues with this approach.

(1) The five behavioral variables included in the analysis are not independent and will covary strongly. Forward selection is greedy, so once one member of a correlated set is admitted, the remaining members' unique contribution to held-out log-likelihood may fall below the 0.01 bits/spike threshold even if they are genuinely encoded. The pairwise correlation (or mutual information) structure among the five variables should be reported for all sessions, as well as the Δllh values of the variables rejected at each step, so readers can judge how close the near-misses were.

(2) What is the L1 penalty and how was it selected? L1 shrinkage lowers each candidate's Δllh, so a stronger penalty yields smaller selected models. This is also important when considering that the basis sets for each variable have different dimensionality. Is a single L1 penalty shared across candidate models?

Reliability of DTG responses.

(1) The 1D DTG tuning curves lack error bars, which should be presented to convey reliability. These should be shown for individual DTG neurons for multiple goal locations.

(2) The 2D DTG ratemaps in Figure 1 are not especially compelling; it would be helpful to see all examples in the supplement.

Several observations suggest that some DTG neurons may actually be encoding boundaries.

(1) The distribution of DTG peaks is strongly bimodal, with peaks either at the goal or at the boundary. Together with the mixed-selectivity results, this suggests that some DTG cells may show a boundary response or activity at the goal related to speed or acceleration. The manuscript at present does not effectively rule out this possibility. This could be addressed by showing that DTG coding is preserved across different goal locations.

(2) An arena-expansion condition would clearly distinguish DTG responses from boundary responses.

(3) It would be useful to see the peak sorted plot (1E) cross-validated within and across goal locations. I suspect that DTGs with intermediate-distance peaks shift with goal location while the others do not. If they remain fixed, it would solidify the presence of the phenomenon.

More characterization of the neurons that are 'important' for goal distance decoding that are not DTG cells is needed (i.e., the IMP population). As I understand it, IMP and DTG cells were grouped for decoding analyses, but the two populations do not overlap. The IMP population is much larger than the DTG cells, and it is unclear what these neurons are doing. Moreover, it seems that the IMP sub-class could alone be used to decode distance to goal. This is difficult to reconcile with the framing of DTG cells as the substrate of goal-distance coding and deserves comment. Decoding should be conducted with the IMP cells alone to show what the DTG cells actually add.

Reviewer #2 (Public review):

Summary:

Chen et al. consider the activity of retrosplenial cortex (RS) neurons during performance of an open-field navigation task in mice. Using a Ca-++ transient imaging approach to examine activity, the authors claim to find tuning to distance of the animal to a hidden reward location. The question of tuning to distance in RS is of much interest of late, with other works making claims. In this respect, the present work is interesting in that it utilizes an actual navigational task that does not explicitly demand encoding of distance and does consider an open-field environment. I do have reservations concerning the robustness of distance tuning.

Strengths:

Testing of distance coding in open fields during performance of an actual navigational task.

Weaknesses:

Lack of robust evidence for distance coding and head direction coding.