Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 75 of the 75 references it cites were found in a registry or an index.

Subspace communication in the hippocampal–retrosplenial axis

Open access

Published version of a preprint

Abstract

The capacity of hippocampal circuits to transform inputs into downstream outputs is fundamental to navigation and memory, yet the circuit-level mechanisms that enable this flexibility in adapting to experience remain unclear. Here we approach this problem by performing large-scale (up to 1,024 channel) recordings across the hippocampal–retrosplenial cortex (RSC) circuit in behaving mice, enabling simultaneous access to spiking activity in dentate gyrus (DG), CA3, CA2, CA1 and RSC. On the basis of a linear dimensionality-reduction technique known as partial canonical correlation analysis, we identify low-dimensional communication subspaces1 between two regions while accounting for influences from a third area. These subspaces captured distinct input–output transformations in the CA1 region, linking upstream hippocampal activity (DG, CA3 and CA2) to downstream cortical targets (RSC). Intrinsic firing properties and anatomical location constrained subspace memberships—members were mapped to deep sublayers of the CA3–CA1–RSC axis during both spatial and non-spatial tasks. These subspaces could recombine overlapping neuronal pools to support distinct interareal interactions across changing experiences and brain states. Reactivation patterns of CA1–CA3 subspaces, but not those of CA1–RSC, during post-experience sleep correlated with replay, reflecting a plasticity–stability balance in the input–output transformation along the hippocampal–retrosplenial axis. Our findings suggest a model in which hippocampal–neocortical communication reconfigures predetermined circuit motifs to flexibly encode experiences. Hippocampal–neocortical communication reconfigures predetermined circuit motifs to flexibly encode experiences.

Overview

Integrity signals

Registry

Crossref · 10.1038/s41586-026-10481-z

Cites

75 of 75 references verified0 citation errors · 0 not checkable

Statistics

Recomputing p-values needs the paper’s full text.

Figures

Figure 1 | Subspace relationships reveal structured routing and rotation between hippocampal and cortical regions (a) Schematic showing SiNAPS probe targeting hippocampal subregions CA1, CA2, CA3, and DG. Right: example simultaneous recording from all regions. Only pyramidal cells are plotted. (b) Schematic showing Neuropixel 2.0 probe targeting CA3, CA1, and layer 5/6 RSC. Right: example simultaneous recording from all regions. Only pyramidal cells are plotted. (c) Illustration of pCCA identifying shared subspaces between regions. The solution yields linear combinations of neural activity (weights) in each region that maximize correlation. (d) Scatter plots showing for each cell its maximum absolute pCCA weight (across all the significant subspaces it participates). Dark dots represent the median at selected bins, and two solid curves represent 25% and 75% quantiles. Note the strong Pearson correlation (all R statistics were highly significant, P<10−20), indicating that the same neurons tend to participate in multiple hippocampal subspaces. (e) Example absolute subspace weights from the input (CA1-CA3) and output (CA1-RSC) subspaces in CA1. A total of 30 CA1 pyramidal cells and 2 significant components from each subspace (Dim1, Dim2) are shown. The thickness of the edges shows the weight magnitude, (f) Maximum principal angle between distinct subspaces (e.g., CA3–CA1 vs. CA1–RSC). For control in the left panel, the maximum principal angle was computed within each subspace by dividing the trials in half and calculating pCCA independently. Left panel: each dot represents the result from one session (n = 13 sessions from 6 mice, combining SiNAPS and NP experiments). For within subspace angles, the results from all within subspaces were averaged to yield a single angle. The same was done for the across subspace group. Right panel: populations across areas were matched for the number of neurons. Average rotations across 100 unit-matching samples per session are shown. N = 10 sessions in 4 mice for CA3 vs. CA1 or vs. RSC comparisons. N = 12 sessions in 4 mice for CA1 vs. RSC comparisons. P values were computed from the Wilcoxon signed-rank test.
Figure 1 | Subspace relationships reveal structured routing and rotation between hippocampal and cortical regions (a) Schematic showing SiNAPS probe targeting hippocampal subregions CA1, CA2, CA3, and DG. Right: example simultaneous recording from all regions. Only pyramidal cells are plotted. (b) Schematic showing Neuropixel 2.0 probe targeting CA3, CA1, and layer 5/6 RSC. Right: example simultaneous recording from all regions. Only pyramidal cells are plotted. (c) Illustration of pCCA identifying shared subspaces between regions. The solution yields linear combinations of neural activity (weights) in each region that maximize correlation. (d) Scatter plots showing for each cell its maximum absolute pCCA weight (across all the significant subspaces it participates). Dark dots represent the median at selected bins, and two solid curves represent 25% and 75% quantiles. Note the strong Pearson correlation (all R statistics were highly significant, P<10−20), indicating that the same neurons tend to participate in multiple hippocampal subspaces. (e) Example absolute subspace weights from the input (CA1-CA3) and output (CA1-RSC) subspaces in CA1. A total of 30 CA1 pyramidal cells and 2 significant components from each subspace (Dim1, Dim2) are shown. The thickness of the edges shows the weight magnitude, (f) Maximum principal angle between distinct subspaces (e.g., CA3–CA1 vs. CA1–RSC). For control in the left panel, the maximum principal angle was computed within each subspace by dividing the trials in half and calculating pCCA independently. Left panel: each dot represents the result from one session (n = 13 sessions from 6 mice, combining SiNAPS and NP experiments). For within subspace angles, the results from all within subspaces were averaged to yield a single angle. The same was done for the across subspace group. Right panel: populations across areas were matched for the number of neurons. Average rotations across 100 unit-matching samples per session are shown. N = 10 sessions in 4 mice for CA3 vs. CA1 or vs. RSC comparisons. N = 12 sessions in 4 mice for CA1 vs. RSC comparisons. P values were computed from the Wilcoxon signed-rank test.CC BY-NC-ND · PubMed Central
Figure 2 | Cross-regional subspaces adapt to distinct environments (a) Schematic of the behavioral protocol. Mice performed in Maze 1 for 30 minutes, followed by a <5-minute intermission, and then explored Maze 2 for an additional 30 minutes. Maze 1 was located within an enclosure in the recording room containing large X and O letters as distal cues. Linearized position trajectory over time is plotted below maze drawings. Traces below show the subspace projected for each area. Subspaces were computed independently on maze 1 (left panels) or maze 2 (right panels). Red and black traces show the subspace activity as a function of time projected on RSC (red) and CA1 (black). (b) Example absolute subspace weights from the two mazes in each area. A subset of 50 neurons and 2 significant components from each subspace (Dim1, Dim2) are illustrated. The thickness of the edges shows the magnitude of the weight. (c) Single neuron maximum subspace weight magnitudes in CA1, CA3, and RSC were correlated between the two mazes (Pearson correlation R=0.46). (d) Subspace correlation ρmax generalization from Maze 1 to Maze 2 or between the first and second half of the trials (Same Maze). All subspaces were computed under one condition and projected to the other. The difference to the MUA correlation is plotted. For the different mazes condition, each dot corresponds to the projection of the subspace computed in one maze to the other (n = 6: two subspaces per session (subspace maze 1 -> maze 2, subspace maze 2 -> maze 1), 3 sessions from 3 mice). For the same maze condition n = 12 sessions from 4 mice, subspace correlation between first and second half (from the subspace computed on the remaining half) were averaged for each session. CV: Cross validation. P values were computed from the Wilcoxon ranksum test.
Figure 2 | Cross-regional subspaces adapt to distinct environments (a) Schematic of the behavioral protocol. Mice performed in Maze 1 for 30 minutes, followed by a <5-minute intermission, and then explored Maze 2 for an additional 30 minutes. Maze 1 was located within an enclosure in the recording room containing large X and O letters as distal cues. Linearized position trajectory over time is plotted below maze drawings. Traces below show the subspace projected for each area. Subspaces were computed independently on maze 1 (left panels) or maze 2 (right panels). Red and black traces show the subspace activity as a function of time projected on RSC (red) and CA1 (black). (b) Example absolute subspace weights from the two mazes in each area. A subset of 50 neurons and 2 significant components from each subspace (Dim1, Dim2) are illustrated. The thickness of the edges shows the magnitude of the weight. (c) Single neuron maximum subspace weight magnitudes in CA1, CA3, and RSC were correlated between the two mazes (Pearson correlation R=0.46). (d) Subspace correlation ρmax generalization from Maze 1 to Maze 2 or between the first and second half of the trials (Same Maze). All subspaces were computed under one condition and projected to the other. The difference to the MUA correlation is plotted. For the different mazes condition, each dot corresponds to the projection of the subspace computed in one maze to the other (n = 6: two subspaces per session (subspace maze 1 -> maze 2, subspace maze 2 -> maze 1), 3 sessions from 3 mice). For the same maze condition n = 12 sessions from 4 mice, subspace correlation between first and second half (from the subspace computed on the remaining half) were averaged for each session. CV: Cross validation. P values were computed from the Wilcoxon ranksum test.CC BY-NC-ND · PubMed Central
Figure 3 | Physiological features of individual neurons predict subspace memberships. (a) CA1–CA3 subspace weights (best subspace, the pCCA component accounting for the strongest interregional correlation) for individual neurons. Subspace “members” (violet) are defined as neurons with absolute weights above the 75th percentile of the distribution (|weight| > 0.75 quantile), while non-member neurons (orange) are defined as having absolute weights below the 25th percentile (|weight| < 0.25 quantile). The dominant subspace (accounting for the largest correlations) was employed for this analysis. Similar results are obtained when using all significant subspaces (Extended Data Fig. 7). (b) Spiking characteristics of subspace members during pre-maze NREM. Left: Firing rate distributions (log scale) across RSC, CA1, and CA3. Right: Burst index distributions for the same areas. (c) Anatomical distribution of subspace members. Depth distribution of subspace members (blue) and non-members (orange) relative to each region’s midplane (defined by the average vertical position of all recorded pyramidal cells in that shank). P values were computed from the Wilcoxon ranksum test
Figure 3 | Physiological features of individual neurons predict subspace memberships. (a) CA1–CA3 subspace weights (best subspace, the pCCA component accounting for the strongest interregional correlation) for individual neurons. Subspace “members” (violet) are defined as neurons with absolute weights above the 75th percentile of the distribution (|weight| > 0.75 quantile), while non-member neurons (orange) are defined as having absolute weights below the 25th percentile (|weight| < 0.25 quantile). The dominant subspace (accounting for the largest correlations) was employed for this analysis. Similar results are obtained when using all significant subspaces (Extended Data Fig. 7). (b) Spiking characteristics of subspace members during pre-maze NREM. Left: Firing rate distributions (log scale) across RSC, CA1, and CA3. Right: Burst index distributions for the same areas. (c) Anatomical distribution of subspace members. Depth distribution of subspace members (blue) and non-members (orange) relative to each region’s midplane (defined by the average vertical position of all recorded pyramidal cells in that shank). P values were computed from the Wilcoxon ranksum testCC BY-NC-ND · PubMed Central
Figure 4 | Subspace members preserve maze-related manifold geometry better than non-members. (a) Top: Linearized position trajectory over time during a novel maze session. Middle: Z-scored firing rate heatmap for CA1 neurons classified as subspace members (purple) of the CA1–RSC subspace. Units were sorted based on the place field location. Bottom: Firing rates for an equal-sized subset of CA1 non-members (green). The dominant subspace was employed for this analysis. Similar results hold for all significant subspaces. 0.75 and 0.25 quantiles were chosen as thresholds for member and non-member definitions, respectively. Subspace members show more structured and coordinated activity over time compared to non-members. (b) 3D UMAP embeddings of population activity during the maze session. Color indicates position (left) or trial number (right). Top: Full CA1 population; Middle: CA1 members in CA1–CA3 subspace. Bottom: CA1 members in CA1–CA3 subspace. (c) Top: Manifold alignment score between UMAP embedding of members and non-members to the UMAP embedding of the full population (similarity between the UMAP embedding of the subset and the full population, see Methods). Middle: Spatial structure index quantified over the UMAP embeddings for members and non-members. Bottom: Temporal structure index quantified over the UMAP embeddings for members and non-members. P values were computed from the Wilcoxon signed-rank test. (d) Peer prediction schematic, a linear regression model was built to predict the activity of a downstream neuron based on the activity of an upstream neuron and/or the position of the animal along the track. 20% of the data was withheld from training and used to generate a cross-validated prediction. (e) Peer prediction R2-statistic results averaged for each session. **, P<0.01, ***, P<0.001, Wilcoxon signed-rank test.
Figure 4 | Subspace members preserve maze-related manifold geometry better than non-members. (a) Top: Linearized position trajectory over time during a novel maze session. Middle: Z-scored firing rate heatmap for CA1 neurons classified as subspace members (purple) of the CA1–RSC subspace. Units were sorted based on the place field location. Bottom: Firing rates for an equal-sized subset of CA1 non-members (green). The dominant subspace was employed for this analysis. Similar results hold for all significant subspaces. 0.75 and 0.25 quantiles were chosen as thresholds for member and non-member definitions, respectively. Subspace members show more structured and coordinated activity over time compared to non-members. (b) 3D UMAP embeddings of population activity during the maze session. Color indicates position (left) or trial number (right). Top: Full CA1 population; Middle: CA1 members in CA1–CA3 subspace. Bottom: CA1 members in CA1–CA3 subspace. (c) Top: Manifold alignment score between UMAP embedding of members and non-members to the UMAP embedding of the full population (similarity between the UMAP embedding of the subset and the full population, see Methods). Middle: Spatial structure index quantified over the UMAP embeddings for members and non-members. Bottom: Temporal structure index quantified over the UMAP embeddings for members and non-members. P values were computed from the Wilcoxon signed-rank test. (d) Peer prediction schematic, a linear regression model was built to predict the activity of a downstream neuron based on the activity of an upstream neuron and/or the position of the animal along the track. 20% of the data was withheld from training and used to generate a cross-validated prediction. (e) Peer prediction R2-statistic results averaged for each session. **, P<0.01, ***, P<0.001, Wilcoxon signed-rank test.CC BY-NC-ND · PubMed Central
Figure 5 | Sleep reactivation of task-defined dominant subspaces predicts replay and spike correlations (a) Schematic showing the projection of maze-defined CA1 subspaces (independently for CA1-CA3 and CA1-RSC) to sleep epochs before (blue) and after (orange) maze exploration. (b) Subspace correlations between maze behavior versus NREM and SPW-R events. The best subspace was used for comparisons, but similar results were obtained for all significant subspaces (Extended Data Fig. 12). (c) Spike rasters and decoded posterior probabilities for representative SPW-R events36,37. values indicate subspace correlations. Note that ρmax for CA1-RSC could be low even when the sequence was regularly ordered (last column). Sequence score was computed for each candidate replay event. P values were computed from the Wilcoxon signed-rank test. (d) Left: Cumulative distribution functions of replay scores for SPW-R events with high vs. low subspace correlation values. P values were computed from a Kolmogorov-Smirnov test. Right: Mean sequence scores for each session for high vs. low subspace correlation events. Only the sessions with a median position decoded error <20 cm were included in the analysis. P values were computed from the Wilcoxon signed-rank test. (e) Normalized population cross-correlograms. All pairwise cross-correlograms (n = ~10,000 pairs) were pooled together across all pairs and sessions. Normalization consisted of subtracting the slow cross-correlogram fit and dividing the difference by the number of spikes emitted by the source population (see Methods). (f) Spike cross-correlations between pre-maze and post-maze sleep sessions. Each line shows a single session. For each session, the normalized cross-correlograms were integrated in a 0–30 ms (CA3-CA1) or 0–50 ms (CA1-RSC) window (shaded area in panel e) and then the results were averaged (n ~ 1000 correlograms). The same 75th and 25th percentile threshold was employed to define members and non-members. P values were computed from the Wilcoxon signed-rank test
Figure 5 | Sleep reactivation of task-defined dominant subspaces predicts replay and spike correlations (a) Schematic showing the projection of maze-defined CA1 subspaces (independently for CA1-CA3 and CA1-RSC) to sleep epochs before (blue) and after (orange) maze exploration. (b) Subspace correlations between maze behavior versus NREM and SPW-R events. The best subspace was used for comparisons, but similar results were obtained for all significant subspaces (Extended Data Fig. 12). (c) Spike rasters and decoded posterior probabilities for representative SPW-R events36,37. values indicate subspace correlations. Note that ρmax for CA1-RSC could be low even when the sequence was regularly ordered (last column). Sequence score was computed for each candidate replay event. P values were computed from the Wilcoxon signed-rank test. (d) Left: Cumulative distribution functions of replay scores for SPW-R events with high vs. low subspace correlation values. P values were computed from a Kolmogorov-Smirnov test. Right: Mean sequence scores for each session for high vs. low subspace correlation events. Only the sessions with a median position decoded error <20 cm were included in the analysis. P values were computed from the Wilcoxon signed-rank test. (e) Normalized population cross-correlograms. All pairwise cross-correlograms (n = ~10,000 pairs) were pooled together across all pairs and sessions. Normalization consisted of subtracting the slow cross-correlogram fit and dividing the difference by the number of spikes emitted by the source population (see Methods). (f) Spike cross-correlations between pre-maze and post-maze sleep sessions. Each line shows a single session. For each session, the normalized cross-correlograms were integrated in a 0–30 ms (CA3-CA1) or 0–50 ms (CA1-RSC) window (shaded area in panel e) and then the results were averaged (n ~ 1000 correlograms). The same 75th and 25th percentile threshold was employed to define members and non-members. P values were computed from the Wilcoxon signed-rank testCC BY-NC-ND · PubMed Central

Verification

Claims the paper makes

We couldn’t retrieve this paper’s full text, so its claims and self-check weren’t run. The registry and citation checks still apply.

Citations

Cites

Show all 73 references

Research works printed but not linked: 1

  1. O’Keefe, J. & Nadel, L. The Hippocampus as a Cognitive Map (Oxford Univ. Press, 1978).

Cited by

See all 14

Metadata

Authors
Gonzalez, J., Vöröslakos, M., Aykan, D., Soto, N., Nitzan, N., Swanson, R., Karadas, M., Chen, Z. S., & Buzsáki, G.
Affiliations
  • New York University (US)
  • NYU Langone Health (US)
Corresponding author
  • Chen, Zhe Sage
  • Buzsáki, György
Venue
Nature
Published
2026-05-13
Volume
655
Issue
8121
Pages
192-201
ISSN
  • 0028-0836 (ISSN-L)
  • 1476-4687
OpenAlex
W7161001422
Cited by
7
References
73
Citations by year
2026: 6
FWCI
24.78
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • PubMed Central (submittedVersion)
Funders
  • National Institutes of Health (1rf1da056394-01)
  • National Institute of Mental Health (R01MH122391, P50MH132642, R01MH139216, R01MH139352)
  • National Institute of Neurological Disorders and Stroke (U19NS107616)
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Memory and Neural Mechanisms
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Neural dynamics and brain function
  • Life Sciences › Neuroscience › Cellular and Molecular Neuroscience › Neuroscience and Neuropharmacology Research
Keywords
input-output transformation, CA1, CA3, dentate gyrus, CA2, retrosplenial cortex, memory reactivation, spatial navigation, memory encoding, large-scale neural recording

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Subspace communication in the hippocampal–retrosplenial axis. https://app.cochat.ai/papers/10.1038/s41586-026-10481-z
The paper
Gonzalez, J., Vöröslakos, M., Aykan, D., Soto, N., Nitzan, N., Swanson, R., Karadas, M., Chen, Z. S., & Buzsáki, G. (2026). Subspace communication in the hippocampal–retrosplenial axis. Nature. https://doi.org/10.1038/s41586-026-10481-z

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.