The hippocampus converts dynamic entorhinal inputs into stable spatial maps

Summary

The medial entorhinal cortex (MEC)-hippocampal network plays a key role in the processing, storage, and recall of spatial information. However, how the spatial code provided by MEC inputs relates to spatial representations generated by principal cell assemblies within hippocampal subfields remains enigmatic. To investigate this coding relationship, we employed two-photon calcium imaging in mice navigating through dissimilar virtual environments. Imaging large MEC bouton populations revealed spatially tuned activity patterns. MEC inputs drastically changed their preferred spatial field locations between environments, whereas hippocampal cells showed lower levels of place field reconfiguration. Decoding analysis indicated that higher place field reliability and larger context-dependent activity-rate differences allow low numbers of principal cells, particularly in the DG and CA1, to provide information about location and context more accurately and rapidly than MEC inputs. Thus, conversion of dynamic MEC inputs into stable spatial hippocampal maps may enable fast encoding and efficient recall of spatio-contextual information.

Graphical abstract

Highlights

• •MEC inputs to the DG, CA3, and CA1 show different spatial coding properties
• •MEC inputs remap even more strongly than hippocampal principal cells
• •Hippocampal principal cell activity is more reliable and stable than their MEC inputs
• •Hippocampal principal cells allow improved spatial and contextual readout

Cholvin et al. imaged MEC input and principal cell activity in all three hippocampal subfields of mice navigating two different virtual environments. MEC projections more strongly discriminated between environments, whereas hippocampal principal cell activity was more stable and reliable, deriving a faster and more accurate readout of space and context.

Introduction

Memories of distinct episodes that take place at different spatial locations are fundamental building blocks to the narrative of our lives. Episodic memories associate events and items with the spatial environment and temporal circumstances in which they were experienced and, hence, make up their context. Hippocampal principal cell (PC) assemblies have been shown to encode such unique episodes, and their re-activation is thought to reinstate the associated memory (O’Keefe and Dostrovsky, 1971; Hafting et al., 2005; Jezek et al., 2011; Miller et al., 2013; Ramirez et al., 2013; Malvache et al., 2016; Josselyn and Tonegawa, 2020). Self-motion- and global-feature-based environmental information is largely provided to the hippocampus by the MEC, and it is thought that this input stream markedly contributes to the spatial and contextual representation by hippocampal PC assemblies (Fyhn et al., 2004; Hafting et al., 2005; Hargreaves et al., 2005; McNaughton et al., 2006; Buzsáki and Moser, 2013; Neunuebel et al., 2013; GoodSmith et al., 2019). Spatial and contextual information originates from multiple cellular MEC sources (Zhang et al., 2013), such as grid cells, which show regular hexagonal activity patterns that span the entire experienced environment (Hafting et al., 2005); non-grid spatial neurons, characterized by discharges in distinct locations (Diehl et al., 2017); and cue-specific cells (Casali et al., 2019). They transmit spatial information via the perforant path to hippocampal subfields and thereby contribute to the spatial tuning of hippocampal PCs (Zhang et al., 2020). Indeed, MEC lesion and activity modulation studies revealed a substantial influence of the MEC on the firing properties and spatial information content of dentate gyrus (DG), CA3, and CA1 cells (Lu et al., 2013; Hales et al., 2014; Pernía-Andrade and Jonas, 2014; Miao et al., 2015; Ormond and McNaughton, 2015; Schlesiger et al., 2015; Rueckemann et al., 2016; Kanter et al., 2017). However, how the spatial code provided by MEC inputs relates to spatial representations generated by PC assemblies in the hippocampal subfields is still not fully understood.

Prominent past work on the environmental representation in MEC-hippocampal networks largely examined the neuronal population within superficial MEC layers as a whole (Fyhn et al., 2007; Kitamura et al., 2015; Pérez-Escobar et al., 2016; Diehl et al., 2017) by recording somatic activity from various morphologically and neurochemically defined cell types. However, based on their unique axonal projections via the perforant path, it can be assumed that information about location and context is not uniformly distributed to PCs of the hippocampal subfields, i.e., pyramidal cells (PYRs) in CA1-3 and granule cells (GCs) in the DG (van Groen et al., 2003; Canto et al., 2008; Neunuebel et al., 2013). The two main PC types in the MEC are stellate cells and PYRs, which differ in morphology, intrinsic properties, neurochemical markers, and the specificity of their axonal projections to hippocampal subfields (Alonso and Llinás, 1989; Lingenhöhl and Finch, 1991; Canto and Witter, 2012). Layer II calbindin-positive PYRs seem to connect with cells in the ipsilateral CA1 (Kitamura et al., 2014) and the contralateral MEC (Varga et al., 2010) and appear to contain the largest proportion of grid cells. In contrast, calbindin-negative stellate cells primarily contact neurons within the DG and CA3 (Tang et al., 2014; Kitamura et al., 2015). Species-related specificities indicate that in mice individual stellate cells are unlikely to project to both the DG and CA3 (van Groen et al., 2003), and therefore, may fall into two distinct anatomical subgroups. Layer III PYRs are thought to mainly target CA1 (Suh et al., 2011; Kitamura et al., 2014). Moreover, MEC axons to the DG run ipsilateral whereas those to CA1 and CA3 project bilaterally (van Groen et al., 2003). Thus, to examine potential functional differences of MEC projections among hippocampal subfields and to determine the encoding of spatial information in hippocampal PCs provided by the MEC, recordings of hippocampal MEC synaptic inputs and PC outputs are essential.

Early theoretical studies suggested that a fundamental property of the MEC-hippocampal network is its ability to decorrelate overlapping input activity patterns, thereby allowing hippocampal networks to form more discriminatory representations of space and context than their MEC inputs (Marr, 1971; O’Reilly and McClelland, 1994; Treves and Rolls, 1994). These theories were later supported by experimental studies that compared firing patterns of PCs within superficial MEC layers representing spatial fields (Burgalossi et al., 2014; Kitamura et al., 2014, 2015; Leutgeb et al., 2005, 2007; Diehl et al., 2017) with the activity patterns of hippocampal place cells in response to environmental manipulations (Leutgeb et al., 2007; GoodSmith et al., 2017; Senzai and Buzsáki, 2017). Discriminatory representations of different environments in the hippocampus have since been thought to be characterized by substantial reorganization of activity patterns of spatially tuned PCs (Leutgeb et al., 2005; GoodSmith et al., 2017). However, only a few studies have compared the spatial characteristics of MEC and hippocampal PCs in response to similar environmental changes, and these studies indicate that the differential representation of spatial and feature information might already happen in the MEC (Fyhn et al., 2007; Diehl et al., 2017). Because reliability in axonal action potential conduction from cell bodies within the MEC over long distances to their release sites in the hippocampus has not been examined (see Rama et al., 2018), it remains difficult to predict what exact spatial and contextual information hippocampal PCs may receive from MEC cells. Thus, to understand how spatial and contextual information is processed within the entorhinal-hippocampal network, it is necessary to unveil input-output transformations by directly recording MEC axons and the output of PC ensembles across the different stages of the hippocampal circuit. However, such studies have been lacking so far.

To experimentally overcome these issues, we recorded activity patterns of populations of MEC boutons and PCs in the hippocampal DG, CA3, and CA1 using two-photon calcium imaging of head-fixed mice navigating through two familiar virtual environments characterized by different sets of visual cues and boundaries. Our results show that dorsal DG and CA3 networks receive a larger proportion of spatially modulated MEC inputs than CA1. Moreover, both MEC inputs and hippocampal PCs show characteristic place- and context-specific activity patterns within all subfields, drastically remapping between contexts. However, hippocampal spatial tuning responses are more stable and reliable over time than the ones provided by MEC inputs. Decoding analysis further revealed that hippocampal PCs, particularly in the DG and CA1, provide information about location and context more efficiently and accurately than their MEC inputs. Thus, we provide novel insights about the relationship between spatial properties of MEC inputs and hippocampal PC outputs in response to navigation through different environments and their potential role in spatial memory and recall.

Results

Imaging place- and context-modulated activity patterns of MEC boutons in hippocampal subfields

To obtain data on the activity of MEC projections targeting a respective hippocampal subfield, we performed two-photon calcium imaging of MEC inputs in mice running on a spherical treadmill on a 4-m-long virtual linear track in two distinct virtual environments (A and B) characterized by different wall patterns, visual cues, and floor textures (Figure 1A), which we refer to as contexts A and B. Mice collected a soymilk reward at the end of each track. After at least 10 days of familiarization to both tracks, imaging sessions started, in which mice ran on tracks A and B in alternating blocks of five runs (Figure 1B; Videos S1 and S2). Animals ran equally well on both tracks (Figure S1), indicating that they behaved similarly in both contexts. To measure neuronal calcium signals, mice were injected with adeno-associated viruses (AAVs) expressing the calcium indicator GCAMP6s panneuronally in layers II/III of the MEC (Figures 1C and S2), which contain the highest proportion of spatially and contextually modulated PCs projecting to the hippocampal subfields (Witter and Groenewegen, 1984; Tamamaki and Nojyo, 1993; Zhang et al., 2013; Kitamura et al., 2014). GCAMP6s-expressing axonal projections were reliably observed in the stratum lacunosum moleculare (slm) of CA1-3 and in the middle molecular layer (mml) of the DG, the classical termination zones of the perforant path (Figures 1D, S2B, and S2C). Putative MEC synapses were identified based on their bouton-like structure (Figure 1E) and imaged in the slm of the proximal CA1 at a depth of ∼350 μm or more laterally in the slm of the distal CA3 at ∼550 μm (CA3a; STAR Methods) and the mml of the suprapyramidal DG blade at a depth of ∼550 μm (Figures 1D–1G)—the hippocampal layers that receive predominant MEC inputs (Neunuebel et al., 2013). To record the activity of CA1-3 PYRs and DG GCs, we injected a separate cohort of animals with the same virus as in our MEC experiments (AAV1-GCaMP6s; Figure S3) and obtained neuronal activity within the same regions of interest as for MEC bouton recordings. Imaged GCs were located in the upper half of the granule cell layer to largely record from mature GCs and along the entire radial axis of the CA1-3 pyramidal cell layer. We used fast scanning to record from on average 522 ± 18 boutons and 145 ± 10 cells per imaging session (STAR Methods).

Figure 1.

Two-photon calcium imaging of MEC boutons activity in the hippocampus in two virtual environments

(A) Experimental schematic of our virtual-reality setup for mice.

(B) Top, schematic of the two familiar virtual contexts; bottom, timeline of a recording session.

(C) GCaMP6s-labeled MEC neurons (green); tissue counterstained with DAPI (blue).

(D) Imaging window implantation site. Axonal projections from the MEC (GCaMP6s, green) over DAPI staining (blue). Red dotted lines, imaging planes in CA1, CA3, and DG.

(E) Calcium activity of axonal projections imaged in CA1, CA3, and DG.

(F) Raw calcium traces (gray) with significant transients (red) and linear-track position (blue) of a MEC-to-CA3 bouton showing place fields over time; left, context A; middle, context B; right, calcium activity over track distance of the same bouton in context A (top) and B (bottom).

(G) Same as (F) for a MEC-to-DG bouton having place fields (see STAR Methods).

(H) Top, fraction of active (>2 transients per min) boutons. Middle, boutons showing significant spatial information. Bottom, boutons with single or multiple place fields. Test for population overlap (χ² test).

(I and L) Mean calcium activity for all boutons (BT) and principal cells (PC).

(J and M) Spatial information of all active BTs and PCs.

(K and N) Width of place fields in context A for all place BTs and PCs.

(I–K) Rank-sum tests, per region (CA1/CA3/DG). (L–N) ANOVA on ranks, Dunn’s test, per functional domain (BT/PC). Boxes, 25th to 75th percentiles; bars, median; whiskers, 99% range. Values represent number of BTs/PCs. NS, not significant; ^∗p < 0.05; ^∗∗p < 0.01; ^∗∗∗p < 0.001. For exact p values, see Table S1.

In comparison to cell bodies, axonal boutons are small and more challenging to image in vivo. To test whether quality and information content of bouton signals in our recordings is comparable with that of somata, we expressed GCaMP6s in CA3 PYRs and imaged calcium transients at their soma and their Schaffer collateral outputs in CA1 (Figure S4). The advantage of this approach is that both somatic and bouton recordings can be obtained within the same animal during exploration of the virtual reality (STAR Methods). The mean baseline noise level and the peak amplitude of calcium transients were similar between somatic and bouton recording sites of CA3 PYRs (Figures S5A and S5B), indicating similar signal-to-noise ratios (see also similar signal-to-noise ratios between MEC boutons and hippocampal PYR somata; Figure S14A) and equal detection quality of calcium transients (Figure S4). The mean rise time and decay time were mildly slower for somatic than CA3 PYR bouton transients (rise time: somata 0.14 ± 0.003 s, boutons 0.12 ± 0.009 s; decay time: somata 1.48 ± 0.02 s, boutons 1.32 ± 0.007 s; p < 0.001 for both parameters; Figures S5C and S5D; STAR Methods). Conclusively, the mean activity and spatial information content were highly similar between bouton and somatic recordings from the same cell population (Figure S4), suggesting that bouton recordings yield a reliable assessment of presynaptic neuronal activity.

Spatial tuning of MEC inputs within hippocampal areas is context-dependent

We first analyzed activity levels and spatial information (STAR Methods) of individual MEC boutons in the two environments. The mean MEC activity level and spatial information content were not different between contexts on the level of individual mice (Figure S6), indicating that both environments were represented equally well by MEC boutons. The fraction of active MEC boutons was larger in CA3 and the DG compared with CA1 (Figure 1H, top row). While the majority of boutons were active in both contexts (Figure 1H, top row), many of them showed significant spatial information in only one context (Figure 1H, middle row; STAR Methods). This observation was even more pronounced when we analyzed boutons with significant place fields (Figures 1H, lower row, and 3C, left). To determine periodic potential grid-like activity patterns among MEC inputs and to test whether they appear context-related, we analyzed the structure of ΔF/F signals using a 1D-grid classifier (Yoon et al., 2016; Gu et al., 2018; STAR Methods). The majority of spatially tuned MEC boutons showed place cell-like activity patterns (75% in CA1, 66% in CA3, 74% in DG; Figures S7B and S7D). A further major population showed multiple periodic fields in one or both contexts (CA1: 9% in A, 11% in B, 5% in both contexts; CA3: 12% in A, 10% in B, 13% in both contexts; DG: 11% in A, 8% in B, 6% in both contexts; Figures S7C and S7D). The proportions of MEC boutons with periodic fields among those with a significant place field were similar between CA1 and the DG, but different between CA3 and CA1 or CA3 and the DG (p < 0.001). The numbers of fields among boutons with grid cell-like activity patterns were similar between hippocampal areas, with a majority of grid cell-like boutons showing 2 or 3 place fields along the 4-m-long tracks (Figure S7E). Overall, MEC boutons presumably originating from potential grid-cells rarely showed the same spatial arrangement of fields in both environments (Figures S7C and S7I). Taken together, our data suggest that the vast majority of MEC inputs with either place cell- or grid cell-like characteristics displayed context-dependent modulation of their spatial tuning.

Figure 3.

MEC boutons targeting hippocampal areas discriminate between contexts

(A and B) Activity-rate difference scores (see STAR Methods) between contexts A and B. Comparison of hippocampal regions (A) and functional domains (boutons/cells, B). BT, boutons; PC, principal cells.

(C) Left, activity maps of MEC-to-CA1 (top), MEC-to-CA3 (middle) or MEC-to-DG (bottom) BT having place fields in context A. Left, activity in context A; right, activity in context B; all blocks of runs in each context were merged. Right, similar to left for CA1 (top), CA3 (bottom), and DG (bottom) PC.

(D) Mean activity correlations between all runs in contexts A and B of boutons having place fields in context A.

(E) Mean activity correlations between first and second blocks of runs in context A (A-A′) and first block in A and first block in B (A-B) of place BT (left) and PC (right).

(A) Rank-sum tests, per region (CA1/CA3/DG). (B and E) ANOVA on ranks, Dunn’s test, per functional domain (BT/PC). (D) ANOVA on ranks, Dunn’s test. Boxes, 25th to 75th percentiles; bars, median; whiskers, 99% range. Values represent number of BTs/PCs. NS, not significant; ^∗p < 0.05; ^∗∗p < 0.01; ^∗∗∗p < 0.001. For exact p values, see Table S1.

Area-specific MEC inputs have lower spatial tuning than their respective target PCs

How does spatial information conveyed by MEC inputs relate to output activity patterns of downstream hippocampal PCs (Figures 1I–1N)? Consistent with the sparse activity of DG GCs (Kitamura et al., 2015; Danielson et al., 2016, Diamantaki et al., 2016; Pilz et al., 2016; GoodSmith et al., 2017; Senzai and Buzsáki, 2017; Pofahl et al., 2020; Zhang et al., 2020), we observed a markedly higher mean activity rate of MEC boutons in the mml compared with downstream GCs (Figure 1I). In contrast, average activity rate was almost identical between MEC inputs and PYR outputs in CA1 and CA3 (Figure 1I). The spatial information content was amplified in PCs of all hippocampal regions compared with the area-specific MEC inputs with the largest increase in CA1, followed by CA3 and the DG (Figure 1J), pointing to subfield-dependent hippocampal refinement of the spatial code provided by the MEC.

In accordance with a previous anatomical study in mice (van Groen et al., 2003), by injecting different color variants of fluorescently conjugated cholera toxin subunit B (CTB) retrograde tracers in the DG (CTB488), CA3 (CTB555), and CA1 (CTB647; STAR Methods), we revealed that the majority of MEC cells projecting to the DG were located within layer II, while CA1- and CA3-projecting neurons were almost exclusively located in layer III. Indeed, the overlap among cell groups projecting to CA3 and the DG was minimal (∼11% layer II, ∼6% layer III; Figure S8). Consistent with the different cellular origin, the average activity and spatial information content of MEC inputs differed among hippocampal subfields. They were highest within CA3, followed by the DG, and lowest in CA1 (Figures 1L and 1M, left). A different picture emerged on the level of active PCs within the respective hippocampal areas. The mean activity and spatial information content were the highest in CA3 PCs, followed by CA1 PCs, and the lowest in DG GCs (Figures 1L and 1M, right; see also Hainmueller and Bartos, 2018). We further observed that the mean place field width mildly increased for CA1 and DG place cells at constant place field numbers compared with the hippocampal subfield-related MEC inputs, whereas in CA3, place field size mildly declined and the number of place fields increased (MEC input ∼1.5 fields versus CA3 place cells ∼1.7 fields; Figures 1K, 1N and S3G–S3J). In conclusion, the spatial tuning of hippocampal PCs, particularly in CA1 and CA3, improved, while their spatial resolution declined in comparison to MEC boutons.

Improved stability of spatial representation in CA1 and DG PCs compared with MEC inputs

To determine the stability of spatial representations by MEC inputs and hippocampal PCs with a spatial field, we first measured the trial-by-trial reliability of place field locations on the linear track (Figures 2A–2C). Reliability of place field-related firing in MEC boutons was largely independent of the hippocampal area. However, reliability considerably improved from MEC inputs to downstream place cells in the DG and CA1 but remained unchanged in CA3 (Figure 2C). Likewise, when we compared the consistency of place-related firing of MEC inputs in CA1, CA3, and the DG between the first (A), middle (A′), and last (A′′) blocks of consecutive runs on the track (Figure 1A), we observed largely similar short- (A-A′) and long-term consistencies (A-A′′; Figure 2D). Similar to the trial-by-trial reliability, place field consistency improved from MEC inputs to downstream PCs on both short- (A-A′) and long-term (A-A′′) assessments in CA1 and the DG and markedly dropped in CA3 place cells (Figures 2F and 2G). Applying the same analytical approach to the data obtained in the second context (B) drew similar conclusions (Figure S9). Changes in place field correlations across all three sessions (Corr_(A-A′) − Corr_(A-A′′), Corr_(A-A′) − Corr_{(A′-A′′)}) were not different among hippocampal areas, neither on the level of MEC boutons nor hippocampal place cells (Figure S10). In summary, the consistency of the spatial code was substantially improved in downstream DG and CA1 but markedly diminished in CA3 place cells.

Figure 2.

MEC spatial boutons show low reliability over multiple exposures to the same context

(A) Activity maps of MEC-to-CA1 (top), MEC-to-CA3 (middle), and MEC-to-DG (bottom) place-modulated boutons during the first (left), second (middle), and third (right) blocks of runs in context A.

(B) Same as A for principal cells (PC) in CA1, CA3, and DG, respectively.

(C) Mean trial-to-trial reliability of place boutons (BT) and PC responses in context A.

(D) Mean activity correlations between first and second blocks of runs in context A (A-A′) and first and third blocks of runs in context A (A-A′′) for place boutons.

(E) Same as (D) for PC.

(F) Mean activity correlations between first and second blocks of runs in context A; comparison between place BTs and PCs for each hippocampal region.

(G) Same as (F) for the first and third blocks of runs.

(C–E) ANOVA on ranks, Dunn’s test. (F and G) Rank Sum-Tests, per region (CA1/CA3/DG). Values represent number of BTs/PCs. Boxes, 25th to 75th percentiles; bars, median; whiskers, 99% range. NS, not significant; ^∗p < 0.05; ^∗∗∗p < 0.001. For exact p values, see Table S1. For context B, see Figure S9.

In accordance with previous studies showing high temporal stability of spatial representations in DG GCs (Hainmueller and Bartos, 2018), we identified high short- and long-term consistencies in place field activity in the DG but significantly lower temporal consistencies for CA1 and CA3 place cells (Figure 2E). Indeed, long-term consistency showed the lowest values for place field discharges in CA3 (Figure 2E). Taken together, spatial representations provided by MEC inputs are less reliable and stable over time than the ones generated by CA1 and DG PCs but more reliable and stable than the ones in CA3. We were attentive to avoid selecting regions of interest (ROIs) surrounding boutons with high activity correlations and connected by an axon fiber suggesting their origin from the same axon (duplicates; STAR Methods). However, to test whether curation of our dataset in this manner might still contain potential duplicates we identified boutons with high activity correlations recorded in the same session and same hippocampal subfield, which would indicate that they may originate from the same axon (STAR Methods) and examined their potential influence on the obtained results (Figure S11). We observed that the fraction of highly correlated boutons was low (CA1, 2.04%; CA3, 3.13%; DG, 3.74%). Removal of this fraction of boutons from the original dataset had no influence on mean activity, reliability, or short-term and long-term consistency of the spatial representation. Similarly, different calcium transient analysis methods (full transient versus transient onset-only; Danielson et al., 2017) replicated our initial results (Figure S12), further emphasizing that differences in calcium transient kinetics did not have a major impact on our findings.

MEC inputs differentially represent contexts when compared with hippocampal PCs

Our results indicated that MEC inputs display differential spatial tuning across contexts (Figure 1H, middle and lower row). We therefore asked whether context differentiation by MEC inputs differs from representations by downstream hippocampal PCs. To quantify neuronal discrimination between environments, we first calculated the activity difference score for each bouton or cell, respectively, based on the normalized activity-rate difference between the two contexts (Figure 3; STAR Methods). The mean activity-rate difference score of active MEC boutons was markedly smaller than the one of downstream hippocampal PCs (Figure 3A). Notably, activity difference scores were the highest in DG GCs, followed by CA1, and the lowest in CA3 PYRs (Figures 3B and 3C). Moreover, in the DG, the distribution of activity difference scores was exceptionally broad (Figure 3A, bright red)—indicating that some GCs were active in only one context, pointing to pattern separation (Allegra et al., 2020)—whereas others were active in both contexts, though at different activity levels, suggesting their potential contribution to pattern generalization (Nakashiba et al., 2012; Hainmueller and Bartos, 2018). This broad distribution in activity patterns was unrelated to different neuron types such as GABAergic and mossy cells, as we selected only GCs based on small soma size and location within the granule cell layer (STAR Methods).

Next, we examined contextual discrimination by focusing on boutons and cells with place fields in context A and quantified how much their activity remapped in context B (Figure 3D). MEC inputs showed low activity map correlations in all hippocampal subfields, with lowest values in the DG (CA1: 0.17 ± 0.03; CA3: 0.1 ± 0.01; DG: 0.06 ± 0.01; Table S1), indicating highly discriminative MEC place fields. Interestingly, between-context correlations of place fields were similarly low for MEC inputs and PCs in CA1 and CA3 but higher for DG GCs than their MEC inputs (Figure 3D). These data indicate that many GC-mediated place fields might represent similarities between the two contexts rather than discriminating between them. This tendency was even more pronounced when we considered MEC boutons and GCs in the DG showing place fields in both contexts (Figure S13). Finally, place field consistency within a context (A-A′) was markedly higher than between contexts (A-B) for MEC inputs and place cells in all hippocampal subfields (Figure 3E), indicating a significant level of remapping in all place cell populations. The effect size of remapping, defined as the difference in place field stability within a context and remapping between contexts (Corr_(A-A′) − Corr_(A-B)), was only mildly different between MEC inputs and principal cells in the hippocampal areas (higher in CA1, equal in DG and lower for CA3 principal cells; p < 0.001, p = 0.31, p < 0.05, respectively; Figure S10).

To test whether differences in imaging conditions between cells and boutons could explain the observed differences between these populations, we first built a linear regression model to account for signal-to-noise ratio (SNR) and transient-amplitude stability as parameters for recording stability (Figure S14A). After regressing out these parameters, we still observed that place field stability was significantly higher in DG cells and significantly lower for CA3 PCs as compared with their inputs, respectively. Next, we determined changes in the SNR from the first (A) and the last (A′′) block of runs in the same context and compared them with the stability of place fields during the same time (A-A′′). Low recording stability would be associated with a decline of the SNR and result in a decrease in place field stability, leading to a correlation between these two variables. We found no such relationship (Figure S14B). Finally, our simultaneous recordings from CA3 somata and their Schaffer-Collateral boutons revealed no difference in trial-to-trial reliability or place field stability in separate blocks of runs in the same environment (A-A′; Figures S4J and S4K). This further argues against the idea that lower recording stability in axons could explain the lower spatial stability we observed in MEC axons. Taken together, these results strongly indicate that different imaging conditions did not explain the observed differences in place field stability of MEC boutons and hippocampal PCs.

In summary, hippocampal PCs show higher activity-rate differences between contexts than their upstream MEC boutons, suggesting improved context discrimination in the hippocampus. However, place field remapping between contexts was similar between MEC boutons and CA1–3 place cells and even reduced in DG GCs, raising the question of the relative contribution of each of these parameters in context discrimination.

Hippocampal output representations are more reliable than their MEC inputs

To directly address how these activity-dependent parameters could influence encoding of space and context, we applied an inverse approach and used the recorded cell or bouton activity to decode the animals’ concurrent location and environment (context) using a population-vector-based method (Meyers, 2013; Figure 4A; STAR Methods). For this analysis, we randomly selected subsamples from all recorded cells or boutons, respectively, in a given session. Decoding performance for both space and context increased when larger samples of cells or boutons, respectively, were used for decoding in all hippocampal subfields (Figure 4B). Notably, decoding errors for both space and context were always substantially smaller for hippocampal PCs than for an equally sized sample of MEC boutons (Figure 4C and 4D).

Figure 4.

Decoding of space and context from neuronal activity

(A) Decoding example. Top, activity of CA1 place cells over time; bottom, decoder output. White line denotes true position of the mouse, and green dots, the most likely decoded position.

(B) Spatial (colored lines) and contextual (gray lines) decoding error as a function of the number of neurons used simultaneously for decoding.

(C) Average spatial errors for ensembles of 50 cells or MEC boutons.

(D) Same as (C) for context errors.

(E) Ratio of spatial decoding errors from the half of the datasets with the highest values of each parameter divided by the lowest half. Open circles denote the mean of each region, filled circles the overall mean. The spatial error ratio is plotted for data split by activity-rate (rate), spatial information (SI), trial-to-trial reliability (rel), short-term consistency among sessions (A-A′), the inverse of the activity-map correlation between context A and B (−(A-B)), and the activity difference score between context A and B (DiffSc).

(F) Same as (E) for context errors.

(G) Ratio of context-decoding errors obtained using a template based on context-specific spatial maps divided by errors obtained using mean activity-rate per context only (Err_aSC/Err_aC; see also Figure S15B).

(H) Cumulative probability for decoding the correct context as a function of time and ensemble size for boutons (left) and cells (right).

(I) Average time to 90% context-decoding accuracy for a fixed ensemble size of 50 cells or boutons, respectively.

(C, D, and I) Kruskal-Wallis, Dunn’s test. (E–G) Paired t test between upper and lower half (E and F) or template type (G), respectively. Boxes, 25th to 75th percentiles; bars, median; whiskers, 99% range. NS, not significant; ^∗p < 0.05; ^∗∗p < 0.01; ^∗∗∗p < 0.001. For exact p values, see Table S1.

Our results above revealed numerous differences in spatial and contextual coding properties between MEC inputs and hippocampal PCs (Figures 2 and 3). To systematically investigate which of these parameters account for the observed disparities in decoding quality, we sorted the cells or boutons in each dataset by their value of a given parameter and then compared the half of the cells/boutons with the higher values (upper half) to the lower half. We divided the decoding error obtained using the upper half by that obtained using the lower half for each dataset (Figures 4E and 4F). A ratio significantly below one would hence indicate that the investigated parameter had a positive impact on decoding performance. This approach revealed that trial-to-trial reliability (rel), spatial information (SI), and consistency (A-A′) of cellular responses were the strongest determining factors of spatial decoding performance across all hippocampal subfields (Figure 4E; see also Table S1 for statistical comparison of multiple parameters). For contextual decoding, activity-rate difference scores between environments had the strongest predictive value, followed by trial-to-trial reliability (Figure 4F). To interpret these results, we considered the level to which individual parameters were correlated with each other. Across datasets, activity-rate and spatial information had moderate correlations, as did stability-related parameters like trial-to-trial reliability and place field consistency, among each other (Figures S15D and S15E). However, activity levels, spatial stability, and place field remapping were not correlated, indicating that these parameter groups span largely orthogonal axes to classify cellular activity in our data.

Surprisingly, the degree by which spatial responses remapped between the two contexts (i.e., the inverse of the activity-map correlation between context A and B; −(A-B)) had a significantly lower predictive value for contextual decoding than activity-rate difference scores across all datasets (Figure 4F). To test whether rate differences might be the sole source of context discrimination in our datasets, we performed separate decoding sessions for context only, using decoding templates built from either the context-specific spatial activity maps (space x context; aSC) of each neuron, or just the average activity-rate for individual cells in each context (aC; Figure S15B). We then calculated a ratio of the mean decoding errors (Err_aSC/Err_aC; STAR Methods). For PCs in all hippocampal subfields, but not for boutons in any subfield, this ratio was significantly below one and largest for PCs in the DG and CA1 (Figure 4G), indicating that decoding performance was improved by using spatially stratified templates (i.e., aSC) for decoding. Hence, in hippocampal PCs, particularly in the DG and CA1, but not their entorhinal inputs, spatial remapping of place fields (global remapping; Leutgeb et al., 2005) provided information about context that was beyond the mere mean activity-rate difference between contexts (i.e., aC). Thus, context-dependent differences in the mean activity rate are key to contextual decoding performance (Figure 4F) and remapping of spatial fields adds moderately to decoding performance in PCs, but not in MEC boutons.

What could be the functional role of more reliable spatial and contextual representations in hippocampal PCs? The hippocampus is required for rapid recognition of behavioral contexts, but this effect can often be mitigated by prolonged context exposure (Matus-Amat et al., 2004; Brown et al., 2011). We therefore investigated how rapidly reliable contextual information could be read out from hippocampal and entorhinal ensembles of varying sizes. We quantified this as the average decision value for context obtained after observing a given neuronal ensemble over a progressively increasing time interval (Figure 4H). Decision accuracy progressively increased with both larger ensemble size and prolonged time of observation. Given an ensemble size of 50 cells or boutons, respectively, the median time to reach 90% contextual decoding accuracy was 13.8 s in CA1, 21.9 s in the DG, and 33.9 s for CA3 as compared with more than 57 s for MEC boutons targeting any hippocampal subfield (DG: 57.4 s; CA1: 59.0 s; CA3: 64.7 s; Figure 4I). Therefore, compared with MEC input synapses, relatively small ensembles of hippocampal PCs, particularly in CA1 and the DG, can rapidly provide a reliable representation of the environment in which the animal is navigating.

Discussion

Using in vivo two-photon calcium imaging, we assessed MEC input and PC output activity in CA1, CA3, and the DG of the hippocampus and found that the average firing rates and spatial activity patterns of both hippocampal MEC inputs and PCs markedly differ between environments, resulting in place maps specific to a given context (Figure 3). However, high trial-to-trial variability and low stability of these inputs compared to PC outputs in their target hippocampal areas (Figures 2 and S9) may preclude reliable contextual representations, which are required for efficient mnemonic associations (Yonelinas et al., 2019). Our data suggest that instead of making representations more dissimilar as previously assumed (Treves and Rolls, 1994; Leutgeb et al., 2007; Deng et al., 2010; Kitamura et al., 2015; Chavlis et al., 2017), the role of hippocampal PCs may rather consist in integrating individually weak and unreliable inputs into more consistent and reliable output patterns, allowing downstream networks to rapidly read out reliable information about space and context from a finite number of PCs.

Strengths and limits of the present study

The main strength of the study is that we obtained the first quantitative data on the spatial and contextual code provided by MEC inputs at the level of individual buttons to the three major hippocampal areas. This approach allowed us to reveal the following unique properties. (1) The proportion of spatially tuned MEC boutons was the highest in the DG, followed by CA3, but the lowest in CA1 (Figure 1). (2) The properties of spatial MEC input information including the mean activity, spatial tuning, and place field width were the highest in CA3, followed by the DG, and again the lowest in CA1 (Figure 1). (3) The location of MEC boutons’ place fields on the track was equally reliable on subsequent runs and similarly stable over subsequent recording sessions across hippocampal subfields, but less reliable and less stable than in downstream CA1 and DG place cells (Figures 2F and 2G). (4) Interestingly, this relationship in input-output stability of place fields was inverted in CA3, suggesting hippocampal area-specific differences in the reconfiguration of spatial representations from MEC inputs to hippocampal place cells. (5) Discrimination between contexts was similar among MEC inputs in the DG, CA3, and CA1. These data, together with the spatial information of PCs recorded under equal behavioral and technical conditions, constrained our decoder analysis on the predictive value of the MEC input and PC output contextual code. The decoder indicated that higher place field reliability and larger mean activity-rate differences between environments allow improved space and context predictions by hippocampal PCs, with the highest mean contextual predictive value observed in CA1 (Figure 4D) and highest mean spatial predictive value in the DG and CA1 (Figure 4C). These data fit to the previously observed long-term stability of DG GC assemblies representing familiar environments (Hainmueller and Bartos, 2018) and the important role of CA1 PCs in providing the spatial code to downstream cortical and subcortical regions (Lee et al., 2014), thereby assisting in associative memory formation.

The main limitation of our study is that calcium imaging experiments were performed in head-fixed mice exposed to a virtual one-dimensional (1D) linear spatial navigation task. Under these unavoidable recording conditions, vestibular inputs are lacking, and mice rely on proximal visual and somatosensory cues, while distal visual cues may appear stationary. It is possible that the spatial tuning of MEC inputs and hippocampal PCs is broader in the virtual reality compared with two-dimensional (2D) environments and more sensitive to local cues (Chen et al., 2018). Although activity of grid cells on 1D tracks might be interpreted as slices through a 2D lattice (Yoon et al., 2016; Pröll et al., 2018), grid-like activity and its spatial tuning on circular tracks seems to be more consistent with the integration of either path distance or time (Domnisoru et al., 2013, Kraus et al., 2015; Jacob et al., 2019) or landmarks (Campbell et al., 2018). Further investigations will be required to examine the influence of 1D versus 2D environments on the spatial properties and the spatial code relationship between MEC inputs and PC outputs. Finally, the familiarity of experienced environments may influence the activity of MEC inputs and their relationship to PC outputs (Hales et al., 2014; Allegra et al., 2020). Future work needs to address this question.

Possible cellular origin of MEC inputs

Where do the observed single (place-like) and periodic (grid-like) activity patterns of MEC inputs in the hippocampus originate? Place-like components could be generated by non-grid neurons, which represent two-thirds of spatially modulated cells in the superficial layers of the MEC (Diehl et al., 2017). The periodic component is likely to originate from grid cells (Hafting et al., 2005), which constitute the minority of superficial MEC cells involved in spatial coding (10%–20%; Sargolini et al., 2006; Sun et al., 2015). Consistent with previous studies, we observed that among the boutons having at least one place field, only 14%–25% depicted a grid-like activity pattern while the majority had single place fields (Figure S7D). Short linear tracks can prevent the detection of grid-like activity (Gu et al., 2018). However, our linear track had a length of 4 m. It is therefore unlikely that the single place field tuning originated from grid cells with a low spatial frequency. Conduction failures of propagating action potentials over long distances (Rama et al., 2018) from MEC cell bodies to hippocampal synaptic release sites might contribute to failures of field representations and thus low reliability on individual runs (Figures 2C and S9C) but are unlikely to result in a lack of fields of periodic spatial activity. Conclusively, we suggest that non-grid neurons represented the major component of spatially tuned MEC signals that we recorded in the hippocampal areas. A minor proportion of GABAergic inhibitory cells within superficial MEC layers form long-range projections to the hippocampus (∼0.9% of all retrogradely labeled cells in layer II/III; Melzer et al., 2012). MEC interneurons exhibit low spatially selective activity and lack spatial periodicity (Buetfering et al., 2014), arguing against a substantial contribution of interneurons in our sample of spatially tuned MEC boutons. Thus, our data indicate that environment-specific information is provided by MEC inputs and is then further refined within the hippocampal circuitry.

It is possible that object cells within the lateral entorhinal cortex, which might be locally or periodically active on the linear track (Deshmukh and Knierim, 2011), could contribute to the spatial tuning of hippocampal PCs. We cannot exclude that DG mossy cells, which express spatially modulated activity patterns and target GCs (Danielson et al., 2017; GoodSmith et al., 2017; Senzai and Buzsáki, 2017), as well as the tri-synaptic path containing mossy fiber synapses contacting CA3 PYRs and Schaffer collaterals targeting CA1 PYRs contribute to the improved reliability and stability of PC spatial tuning under our experimental conditions. However, transgenic inhibition of mature GC-mediated mossy fiber synapses, which form the majority of DG inputs to CA3 PCs, indicates that this spatial information is dispensable for forming neuronal representations of already experienced familiar environments (Nakashiba et al., 2012). Moreover, blockade of CA3 outputs indicates mild influences on place field tuning and no influence on the stability of place fields in familiar spatial contexts (after >3 days of contextual experience), suggesting that the entorhinal-CA1 system is sufficient for the recollection of spatial memory (Brun et al., 2002) and can even improve CA1 place cell tuning (Nakashiba et al., 2008). Further work combining retrograde tracing with selective manipulation of MEC cell types will be required to disentangle the activity of these different input types in the hippocampal subfields and to examine their influence on the characteristics of spatially tuned firing patterns of hippocampal PC assemblies.

Reconfiguration of MEC boutons and hippocampal PC activity patterns following the experience of different environments

Previous theoretical studies assumed that changes of MEC-mediated input activity patterns following the experience of different environments should be substantially amplified in the hippocampal network, thereby reducing the overlap in the representation of locations and contexts—a process called orthogonalization (O’Reilly and McClelland, 1994; Treves and Rolls, 1994; Leutgeb et al., 2007; Kitamura et al., 2015; Yonelinas et al., 2019). In line with this theory, grid axons recorded in the mml of the DG did not exhibit detectable changes in their firing field location in response to marked changes of the environment but caused decorrelations of place field activity in DG cells (Leutgeb et al., 2007). Our data are not entirely consistent with this idea. We provide evidence that MEC inputs drastically change their preferred firing locations in mice navigating through different environments, whereas DG GCs showed higher, and CA1-3 PYRs similar, place field correlations compared with their upstream MEC inputs (Figures 3D and S13B). The DG-specific higher place field correlations are consistent with their ability to generalize between two distinct environments (Hainmueller and Bartos, 2018) and suggest that DG granule cells remap less than their MEC inputs. However, the influence of contextual changes on the reconfiguration of activity patterns of spatially tuned MEC cells is complex and depends on several factors—in particular, the degree of contextual modification, the behavioral conditions, and the MEC cell type. Under conditions of marked decorrelation of place fields in CA1–3, clear shifts in the activity vertices of grid patterns have been detected for MEC cells (Fyhn et al., 2007; Diehl et al., 2017) but not during recordings of the perforant path in the DG (Leutgeb et al., 2007). Spatially tuned non-grid MEC neurons decorrelate their spatial place fields more strongly than grid cells in response to similar degrees of spatial modifications (Diehl et al., 2017). The majority of our recordings depicted MEC inputs with place cell-like properties consistent with their numeric dominance in superficial MEC layers (Diehl et al., 2017). Taken together, our results suggest that contextual changes on the linear track caused substantial remapping of MEC inputs as a whole, supporting the view that orthogonalization processes may already take place at synapses upstream of hippocampal PCs (Marozzi et al., 2015).

Functional implications of enhanced stability of the spatial code provided by the hippocampus to downstream brain areas

Biological readout of information by individual downstream neurons puts extraordinary demands on the synaptic inputs, as integration times are limited to a few tens of milliseconds by the neuronal membrane time constant (hippocampal PCs ∼20 ms; Magee, 1998; Buzsáki, 2010; Kowalski et al., 2016). These time windows are considerably shorter than those required to achieve accurate context-decoding performance in our decoding experiments (Figure 4H). Biological networks are likely to achieve faster “decoding” of input patterns by integrating inputs from substantially larger “upstream” neuronal ensembles of a few thousand cells (Patton and McNaughton, 1995). However, our results indicate that for comparable decoding performance in the same amount of time, this number would need to be as much as an order of magnitude higher when sampling from MEC axon terminals as compared with hippocampal cells in the CA1 area (Figure 4H). Therefore, neurons in networks downstream of the temporal lobe, which must integrate contextual information concomitantly with other inputs (representing e.g., events, persons, or objects) might rely on highly reliable context representation because only a fraction of their input synapses can be allocated to contextual information in order to not exceed their total input capacity. Thus, our present data support the essential role of the hippocampus in rapid context recognition (Matus-Amat et al., 2004; Brown et al., 2011) and association of events with locations and contexts (Yonelinas et al., 2019), by converting context-specific but dynamic MEC inputs into dependable and stable output representations. In line with observations on input-output relations in sensory cortices (e.g., visual cortex; Chen et al., 2013, Wilson et al., 2016), our work indicates that feature selectivity is less precise at dendritic inputs than at their PC somatic outputs, potentially suggesting general organizational principles to ensure feature selectivity and coding flexibility of PCs in cortical networks. Whether plasticity mechanisms such as increase in synaptic efficacy (Kentros et al., 1998; McHugh et al., 2007; Pelkey and McBain, 2007; Bittner et al., 2017), changes in intrinsic membrane properties (Pignatelli et al., 2019) or nonlinear dendritic input integrations (Bittner et al., 2015) may contribute to the MEC input to hippocampal output stabilization of the spatial code remains to be determined.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Rabbit anti-PCP4 primary antibody	Sigma-Aldrich	Cat#HPA005792; RRID: AB_1855086
Goat anti-Rabbit IgG (H+L) Highly Cross-Adsorbed Secondary Antibody, Alexa Fluor 568	Invitrogen	Cat#A11036; RRID: AB_10563566
Bacterial and virus strains
AAV1.Syn.GCaMP6s.WPRE.SV40	Chen et al. Nature. 2013 Jul 18;499(7458):295-300. https://doi.org/10.1038/nature12354.	Addgene, Cat#100843-AAV1; RRID:Addgene_100843
Chemicals, peptides, and recombinant proteins
Cholera Toxin Subunit B (Recombinant), Alexa Fluor™ 488 Conjugate	ThermoFisher	Cat#C34775
Cholera Toxin Subunit B (Recombinant), Alexa Fluor™ 555 Conjugate	ThermoFisher	Cat#C34776
Cholera Toxin Subunit B (Recombinant), Alexa Fluor™ 647 Conjugate	ThermoFisher	Cat#C34778
NaCl	Roth	Cat#P029.1
NaHCO3	Roth	Cat#6885.2
KCl	VWR (Merck)	Cat#26764.232
Na2HPO4	Roth	Cat#4984.2
NaH2PO4	VWR (Merck)	Cat#1.06346.0500
CaCl2	VWR (Merck)	Cat#1.02382.1000
MgCl2	Roth	Cat#A537.4
Glucose	Roth	Cat#6887.1
DAPI	Sigma-Aldrich	Cat#D9542
Triton X-100	Sigma-Aldrich	Cat#T8787
Mowiol 4-88	Roth	Cat#0713.2
Normal Goat Serum	Dianova	Cat#005-000-121
Paraformaldehyde	Roth	Cat#0335.3
Ketamine/Xylazine	Sigma-Aldrich	Cat#K113
Super-Bond Universal Polymer Radiopaque	Hentschel Dental	Cat#40305-01
Super-Bond Catalyst V	Hentschel Dental	Cat#40305-03
Super-Bond Quick Monomer	Hentschel Dental	Cat#40305-06
Experimental models: Organisms/strains
C57BL6/J wild-type mice	Charles River (from The Jackson Laboratory)	RRID: IMSR_JAX:000664
Software and algorithms
Python 3	Python Software Foundation	https://www.python.org
MATLAB 2019, 2020	Mathworks	https://www.mathworks.com
Blender	The Blender Foundation	https://www.blender.org
Suite2p	Carsen Stringer, Marius Pachitariu	https://github.com/MouseLand/suite2p
Zen 2012	ZEISS	https://www.zeiss.com/corporate/int/home.html
Sigmaplot 13	Systat Software	https://systatsoftware.com/sigmaplot
Coreldraw 2020	Corel Corporation	https://www.coreldraw.com/en/
Scanbox	Neurolabware	https://scanbox.org
Two-photon calcium imaging pipeline (MATLAB code)	Bartos Lab	https://doi.org/10.5281/zenodo.5410473

Resource availability

Lead Contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Marlene Bartos (marlene.bartos@physiologie.uni-freiburg.de).

Materials Availability

This study did not generate new unique reagents.

Experimental model and subject details

Animal experiments

All experiments involving animals were carried out according to national and institutional guidelines and approved by the ‘Tierversuchskommission’ of the Regierungspräsidium Freiburg (license no. G18/071) in accordance with national legislation. For in vivo two-photon calcium imaging of MEC boutons and principal cells in the hippocampus, we used 16 C57BL/6J wild-type male mice aged 9-12 weeks at the beginning of the experiments. Animals were all recorded in more than one hippocampal region, allowing us to reduce the number of animals used as follows: recordings of MEC inputs in hippocampal areas: n_total = 5 animals, with n = 5 for CA1, n = 4 for CA3, n = 4 for DG; hippocampal principal cells recordings: n_total = 7 animals, with n = 7 for CA1, n = 5 for CA3, n = 6 for DG; CA3 principal cells and Schaffer collaterals (CA1) recordings: n_total = 4 animals. For the retrograde labeling of MEC cells projecting to the hippocampus using Cholera Toxin Subunit B, we used 6 C57BL/6J male mice aged 9-12 weeks at the beginning of the experiments. Mice were housed on a 12-h light-dark cycle in groups of 2-3 mice. No statistical methods were used to predetermine sample size. The experiments were not randomized and the investigators were not blind to allocation during experiments and outcome assessment.

Methods details

Virus injections and head plate implantation

All surgical procedures were performed in a stereotactic apparatus (Kopf instruments) under anesthesia with 1.5%–2% isoflurane and analgesia using 0.1 mg.kg⁻¹ buprenorphine, and postoperative analgesic treatment consisted of Carprofen administration (5 mg.kg⁻¹ of body weight) provided during 3 days after surgery. An eye lubricant ointment (Bepanthen, Bayer) was applied to protect corneal membranes during surgery. The skin was disinfected with 70% ethanol and Povidone-iodine (Betadine, Avrio Health L.P.) before surgical incision. For GCAMP6s expression in MEC boutons, a small craniotomy (diameter 0.5-1 mm) was made over the entorhinal cortex (A/P −0.8 mm from the lambdoid suture, M/L 3.25 mm from Lambda). A glass micropipette was lowered at an angle of −11° (pointing anterior) and 120 nL of AAV1.Syn.GCaMP6s.WPRE.SV40 (titer 2.5 × 10¹² vg (viral genomes) per ml; Addgene plasmid #100843 (Chen et al., 2013)) were slowly injected while progressively moving the tip of the needle from D/V −1.75 to −1.35 mm (from brain surface) in order to allow infection of layers II and III over the ventro-dorsal extend of the dorsal MEC. For GCAMP6s expression in hippocampal PCs (main experiment), 2-times 400 nL of AAV1.Syn.GCaMP6s.WPRE.SV40 (titer 1 × 10¹² vg per ml) were injected at the following coordinates: A/P −2.0 mm from Bregma, M/L 1.45 mm from Bregma, D/V −1.35 mm or −1.75 mm from brain surface, ensuring CA1, CA3 and DG infection. For GCAMP6s expression in CA3 PCs (for CA3 cells and Schaffer collaterals recordings), 150 nL of AAV1.Syn.GCaMP6s.WPRE.SV40 (titer 1 × 10¹² vg per ml) were injected at the following coordinates: A/P −2.0 mm from Bregma, M/L 2.3 mm from Bregma, D/V −1.9 mm from brain surface. For every viral injection, the target volume was slowly injected over ∼2 min and the glass micropipette was further left in situ for 7 min to ensure complete diffusion of the viral vector in the parenchyma. In the same surgery session, mice were implanted with a stainless-steel head plate (25 × 10 × 0.8 mm with 8 mm-wide central aperture) installed horizontally over the hippocampus, centered on A/P = −1.9 mm and M/L = −1.8 mm from Bregma. All animals were injected in and implanted above the left hemisphere. Mice were allowed to recover from surgery for at least 5 days before any further experiment.

Imaging window implantation

During a second surgery, taking place at least 7 days after the first one described in the previous section, an imaging window was implanted. A craniotomy (diameter 3 mm) was made at A/P −1.9 mm, M/L −1.8 mm. Under continuous irrigation with chilled saline, part of the somatosensory cortex and posterior parietal association cortex located above the hippocampus were progressively aspirated until the external capsule was exposed. The outer part of the external capsule was then gently peeled away using fine forceps, leaving the inner capsule and the hippocampus optically accessible, yet undamaged. The imaging window implant consisted of a 3-mm diameter coverslip (CS-3R, Warner Instruments) glued to the bottom of a stainless-steel cannula (3-mm diameter, 1.3-mm height). This window was gently lowered into the craniotomy using forceps until the coverslip was sitting on the external capsule. The implant was then fixed to the surrounding skull using cyanoacrylate. Mice were allowed to recover from window implantation for at least 4 days before any further experiment.

Virtual environment setup

As previously described (Dombeck et al., 2010; Hainmueller and Bartos, 2018; Sheffield and Dombeck, 2015), our custom virtual environment setup consisted of an air-supported polystyrene ball (20-cm diameter) attached at one side with a small metal axle, which constrained the ball motion to the forward–backward direction. Ball movement was monitored using an optical sensor (G-500, Logitech) and translated into forward motion through the virtual environment. The forward gain was set such that 4 m of distance traveled along the circumference of the ball equaled one full traversal of the linear track. When the mouse reached the end of the track, screens were blanked for 5 s and the mouse was ‘teleported’ back to the start of the linear track. The virtual environment was displayed on four TFT monitors (19" screen diagonal, Dell) arranged in a hexagonal arc around the mouse and placed ∼25 cm away from the head, thereby covering ∼260° of the horizontal and ∼60° of the vertical visual field of the mouse. The virtual environment was created and simulated using the open-source 3D rendering software Blender (available at www.blender.org). The two linear tracks consisted of distinct arrangements of textured walls, floors and other 3D rendered objects placed along the tracks sides as visual cues. Reward locations were marked with a visual cue on the floor (similar to a platform), and 3 μl of soy milk were dispensed through a spout in front of the mouse when it reached the rewarded spot located at the end of each track (Videos S1 and S2).

Behavioral training

Five to 7 days after head plate implantation, mice were placed in the virtual environment setup for 10 to 30 min daily, with gradually increasing time spans over days. For the very first days (3 sessions), only the first familiar context (A) was available to the mice. After this delay, both tracks were displayed to the mice alternatingly over the course of each training session in a pseudo-randomized manner. Once the mice showed signs of habituation to this behavioral task (i.e., appropriate position on the ball and consistent voluntary running, usually after 5 to 10 days of training), food restriction was initiated with a goal of ∼85% of the ad libitum body weight. Training in both virtual environments was maintained for 30–60 min daily until consistent reward licking was observed in all animals, and familiarization to both contexts had been achieved (i.e., at least 10 days of exposure to both contexts before any imaging session).

Behavioral paradigm for imaging sessions

For each imaging session, mice alternatingly ran on the two tracks for a total of 30 runs, by blocks of 5 runs on each track (starting with context A), for a total of 3 blocks and 15 recordings for each track. In many of the mice, the visible area under the imaging window was sufficiently large to select several imaging fields of view that contained different populations of neurons, in the same hippocampal area or in a different one, allowing for example to record CA1 cells, but also CA3 cells using another field of view. In these cases, we repeated the imaging session on another day (only one imaging session per animal and per day), alternating between the hippocampal areas that could be imaged (depending on each animal). In this manner, we performed, in total: for MEC boutons, 12 experiments in 5 animals in CA1 (5593 boutons), 14 experiments in 4 animals in CA3 (7818 boutons), and 14 experiments in 4 animals in the DG (7455 boutons); for hippocampal PCs, 12 experiments in 7 animals in CA1 (1955 cells), 14 experiments in 5 animals in CA3 (2019 cells), and 23 experiments in 6 animals in the DG (3133 cells).

In vivo two-photon calcium imaging

Imaging was performed using a resonant / galvo high-speed laser scanning two-photon microscope (Neurolabware) through a 16x objective (Nikon, 0.8 N.A., 3 mm WD) with a frame rate of 15.5 Hz and using a single plane for imaging. GCaMP6s was excited at 930 nm with a femtosecond-pulsed two-photon laser (Mai Tai DeepSee, Spectra-Physics). To block ambient light from reaching the photodetectors, the animal’s head-plate was attached to the bottom of an opaque imaging chamber before each experiment, and the mouse was then affixed to the virtual environment setup using this head chamber. A ring of black foam rubber was placed between the imaging chamber and the microscope objective, and a metallic collar surrounding the objective was sitting on the imaging chamber, blocking any remaining stray light. MEC boutons were imaged in the stratum lacunosum moleculare of CA1 at a depth of ∼350 μm, more laterally in the stratum lacunosum moleculare of the distal CA3 at ∼550 μm, and in the middle molecular layer of the DG at a depth of ∼550 μm. Hippocampal PCs were imaged at a depth of ∼150 μm in CA1, ∼650 μm in CA3 and ∼650 μm in the DG. Thus, by imaging MEC projections to CA3 as well as CA3 PCs at a depth similar to the depths used for their DG counterparts, we ensured that most of the CA3 BTs and PCs we recorded were actually located in CA3, and unlikely to be part of CA2 network. To define as precisely as possible the CA3 imaging area, we applied the following measures: first, we experienced that our viral injections into the hippocampus (preparation for PC recordings) resulted in highest GCAMP6s expression levels in CA2, leading to the consistent observation of a strongly labeled population of CA2 cells under the 2-Photon microscope. We used this as a landmark to obtain imaging data from PCs below this critical region. We confirmed that this excessive expression of GCaMP was indeed confined to CA2 by antibody labeling against PCP4, a specific marker of CA2 PYRs (Fernandez-Lamo et al., 2019) (Figure S16) in post-mortem sections from our imaged animals. Second, we took the characteristic orientation of basal PYR dendrites as an additional orientation point along the dorsal-ventral axis (from CA2 to CA3b). In CA2, basal dendrites are oriented upward (i.e., more superficial as compared to the somata) while the apical dendrites are oriented downward (i.e., going down in the parenchyma). In CA3a, apical and basal dendrites are within the same plan as the somata under the 2-Photon microscope (see Video S7). In CA3b, apical and basal dendrites show the opposite orientation as compared to CA2 (basal dendrites are oriented downward and apical dendrites are oriented upward in relation to the soma). We used these 3 factors (imaging depth, intense CA2 PYR labeling and dendrite orientation) to ensure that the vast majority of the CA3 cells that we imaged were located in CA3a. Laser power and photomultiplier (PMT) detectors (Hamamatsu H11706-40 GaAsP) were compensated appropriately in Z throughout the stack (laser power range: 12.7-39.3 mW; PMTs’ gain range: 47%–60%) and was kept similar for boutons and cells within a hippocampal area. The three hippocampal areas within animals were chosen in a randomized manner for repetitive imaging. Thus, bleaching or different laser setting cannot account for the observed differences in spatial coding properties of boutons among hippocampal subfields or differences in between boutons and PCs within a given hippocampal area.

Histology

At the end of experiments, mice were deeply anaesthetized using a mixture of ketamine / xylazine (Sigma Aldrich), then intracardially perfused with 0.1 M phosphate-buffered saline (PBS) for 5 min followed by 4% paraformaldehyde (PFA) in PBS for 10 min. Brains were immersed 3 h in 4% PFA, then kept in PBS until they were cut into 80-μm-tick sagittal slices containing the MEC and hippocampal areas. On a subset of slices, PCP4 antibody labeling was performed to reveal CA2 PYRs (Figure S16). Slices were first permeabilized in PBS containing 0.5% Triton X-100 (PBS-TX), then incubated in a blocking solution (10% Normal Goat Serum (NGS) in PBS-TX) for 2 h at room temperature (RT, ∼21°C). Slices were incubated with the primary antibody rabbit anti-PCP4 (1:300, Sigma-Aldrich, HPA005792) diluted in a solution of PBS-TX and 3% NGS overnight 2 h at RT. The day after, sections were washed 3 × 10 min in PBS-TX at RT and incubated in the secondary antibody goat anti-rabbit (1:500, Alexa Fluor 568, Invitrogen #A11036) diluted in a solution of PBS-TX and 3% NGS for 2 h. Slices were finally rinsed in PBS 3 × 10 min, counterstained with DAPI and mounted in Mowiol. Image stacks of GCaMP6s (and when applicable, PCP4) were acquired from slices with a confocal microscope (LSM 710, Zeiss). In MEC-injected animals, confocal image stacks were used to confirm that GCaMP6s infection was centered and restricted to layers II and III of the MEC (Figure S2). In all animals, the locations of the in vivo imaged hippocampal regions were confirmed by comparing averaged two-photon calcium images with confocal images.

Quantification and statistical analysis

Calcium imaging data processing and ROI extraction

The processing of all raw calcium data was done using the Python-based toolbox Suite2p, a free automated pipeline for processing two-photon calcium imaging recordings (available at https://github.com/Mouseland/suite2p). Briefly, Suite2p first aligns all frames of a calcium movie using two-dimensional rigid registration based on regularized phase correlation, subpixel interpolation, and kriging (Pachitariu et al., 2016; Stringer et al., 2019). This toolbox then allows visual inspection of the registered movie. One video of a typical example for each area (CA1, CA3 and DG) and population (cells / boutons) is available online (Videos S3, S4, S5, S6, S7, and S8). Only datasets in which consistent alignment was achieved were kept for further processing. Next, Suite2p performs automated cells / boutons detection and neuropil correction. To detect cells / boutons, Suite2p computes a low-dimensional decomposition of the data, which is used to run a clustering algorithm that finds regions of interest (ROIs) based on the correlation of the pixels inside them. Suite2p also allows to compensate cells / boutons fluorescence traces for the surrounding neuropil signal. This contamination is removed by subtracting a scaled-down version of the neuropil signal around the ROI out from the ROI signal (scaling factor set to 0.7 for all data). All ROIs were manually curated to ensure the most accurate selection of boutons / cells. Cells / boutons were inspected for each run to ensure that segmented entities were clearly visible throughout the entire experiment. In case of bouton-related data, special attention was given to keep only round shaped ROIs representing isolated boutons, while axonal segments were rejected (Figure S17). We were attentive to avoid selecting ROIs surrounding boutons with high activity correlations and connected by an axon fiber suggesting their origin from the same axon (duplicates). The software Suite2p allows for the display of ROIs with high activity correlations, thereby greatly facilitating this procedure (Pachitariu et al., 2016). However, we further tested whether our datasets curated in this manner still contained ROIs showing excessively high calcium-signal correlations, which would indicate that they may originate from the same axon. To determine the range of correlation values that would be expected from independent cells, we consulted our PC datasets and determined for each cell the maximum of the signal correlation values with any other PC in the same recording (Figure S11A). We then turned to our bouton recordings, reasoning that boutons whose pairwise signal correlation substantially exceeded the distribution of correlations in the PC datasets could be a potential duplicate and therefore excluded all boutons whose pairwise correlations lay more than two standard distributions (threshold 2 σ; R = 0.88) above the distribution of maximum correlations for PCs. We observed that the fraction of such highly correlated boutons (CA1, 2.04%; CA3, 3.13%; DG, 3.74%) and PYR somata (CA1, 1.38%; CA3, 1.24%; DG, 1.53%) was low and did not influence our obtained data (e.g., mean activity, reliability, short as well as long-term consistency). Somata were identified based on size and location. In the DG, we disambiguated GCs from other cell types by their small soma size and their location in the granule cell layer. Neurons with unusually large somata or locations within the hilus were discarded. In CA1 and CA3, only neurons in the pyramidal cell layers were included, and the frequency and shape of the calcium transients was used to discard putative interneurons.

We identified significant calcium transients as described previously (Dombeck et al., 2007, 2010). This technique has been used in a number of subsequent hippocampal in vivo calcium imaging studies in other (Rajasethupathy et al., 2015, Sheffield and Dombeck, 2015; Danielson et al., 2016) and our lab (Hainmueller and Bartos, 2018). We restricted analysis to running periods with a speed of at least 5 cm s⁻¹. In brief, calcium traces were corrected for slow changes in fluorescence by subtracting the 8^th percentile value of the fluorescence-value distribution in a window of 20 s around each time point from the raw fluorescence trace. We obtained an initial estimate on baseline fluorescence by calculating the mean and standard deviation (SD) of all points of the fluorescence signal that did not exceed 2.3 SD of the total signal. We then divided the raw fluorescence trace by this value to obtain the ΔF/F trace. This trace was used to determine the parameters for transient detection that yielded a false positive rate (defined as the ratio of negative to positive oriented transients) < 5% and extracted all significant transients from the raw ΔF/F trace (Dombeck et al., 2010). Definitive values for baseline fluorescence and baseline SD were calculated from all points of the trace that did not contain significant transients. A transient mask was created, and for further analysis, all values of this ΔF/F trace that did not contain significant calcium transients were set to zero (Dombeck et al., 2010), to improve the signal to noise ratio. Using this method, ΔF/F is expressed in units of SD (the standard deviation of the baseline fluorescence). In a subset of experiments, we used a second calcium transient detection method to select only the onsets of calcium transients (Danielson et al., 2017). Rising parts of the significant transients were identified by selecting calcium signals having a positive derivative in time for a minimum duration of 0.3 s (Figure S12). We used this method to re-analyze the mean activity, the activity difference score between context A and B, trial-by-trial reliability, place field correlations in context A and between context A and B, short- (A versus A’) and long-term (A versus A’’) consistencies of active putative MEC boutons and pyramidal cell somata in all three hippocampal regions (Figure S12D-G). We observed no systematic differences on the aforementioned measures between the two detection methods.

To determine the kinetic properties of calcium transients of CA3 PYRs and their projections to CA1 (Figure S4 and S5), we fit the fluorescence trace of each response with an autoregressive model of order 2 using the OASIS method (Friedrich et al., 2017) provided in the Caiman package (Giovannucci et al., 2019). We obtained the coefficients in the autoregression, then simulated an impulse response to extract the rise time in which the amplitude rises to 1-e^-1 (∼63%) of the maximal amplitude and the decay time during which the amplitude decays to e^-1 (∼37%) of the maximal amplitude.

Activity differences and spatial information

Activity-rate difference scores (Figure 3A-B) were calculated for each cell / bouton using the following formula: | (activity_CtxA – activity_CtxB) | / (activity_CtxA + activity_CtxB). To calculate a measure for spatial information (SI) content, we adapted a common method of SI assessment (Skaggs et al., 1993) from calcium imaging data. The average calcium activity (mean ΔF/F) was computed for each 5-cm-wide bin along the linear track and used as an approximation for the neurons’ average firing rate in that location. As described previously (Hainmueller and Bartos, 2018), spatial information was calculated as $\text{SI}=\sum _{i=1}^N{\text{λ}}_i\ln \frac{{\text{λ}}_i}{\text{λ}}{p}_i$ in which ${\text{λ}}_i$ and $p_i$ are the average calcium activity and fraction of time spent in the i^th bin, respectively, $\text{λ}$ is the overall calcium activity averaged over the entire track, and $N$ is the number of bins on the track (80 bins in total). Therefore, SI content is inferred from differences in the calcium activity and expressed as bits ^∗ s^-1. For each bouton / cell, significant SI was assessed by shuffling y traces (position of the animal along the track) of the original dataset and computing the SI score of the resulting shuffled dataset. This procedure was repeated 1000 times, and the p value was determined as the fraction of shuffled datasets in which the SI score was higher than the SI score of the original dataset. Spatial information was considered significant if p < 0.05.

Place field identification

Place fields were identified according to published methods (Dombeck et al., 2010; Sheffield and Dombeck, 2015; Hainmueller and Bartos, 2018). In brief, the mean ΔF/F was computed from significant calcium transients for each 5-cm-wide bin along the linear track (80 bins) and this mean fluorescence over distance was then smoothed by averaging over the three adjacent points for each bin. Potential place fields were initially identified as contiguous regions of this ΔF/F over distance plot in which all of the points were greater than 25% of the difference between the bin with the highest ΔF/F value and the baseline value (mean of the lowest 20 out of 80 bins’ ΔF/F values). In addition, the candidate place fields had to fulfil the following criteria: (1) the width of the potential field had to be of at least 3 bins (corresponding to 15 cm running distance); (2) the mean ΔF/F value inside the field had to be at least seven times the mean of the ΔF/F value outside the field; and (3) significant calcium transients had to be present at least 20% of the time in which the mouse was moving in the field. Potential place fields that fulfilled these criteria were accepted if their P value from bootstrapping exceeded 0.05. For bootstrapping, the ΔF/F trace for each experiment was broken into segments of 50 consecutive imaging frames and randomly shuffled, and this was performed 1000 times. Then the place field detection procedure described above was performed on each of the shuffled ΔF/F traces, and the P value of the place field was defined as the number of these randomly shuffled traces on which a place field was detected (according to the outlined criteria divided by the number of shuffles, i.e., 1000). Note, that the same exact parameters were used for cells and boutons in order to maximize comparability. These criteria for place field identification could be considered as relatively conservative and, thus, may underestimate the fraction of detected place cells / boutons among the active ones (Figure 1H; Figure S3C).

Place field consistency, similarity between contexts and trial-to-trial reliability

To assess the similarity of a place cell / boutons spatial representation in different contexts, we calculated the mean ΔF/F value for each of the 80 bins on the track, based on all significant calcium transients (activity map) for each cell / bouton, and for both contexts. As each recording session consisted of 3 blocks of runs in each context (5 runs per block, for a total of 6 blocks and 30 runs), the stability of place fields was measured as the cross-correlation of the mean activity maps for runs in the same context on two different blocks of runs, i.e., the cross-correlation between the average activity of the first and the second block of five consecutive runs on the same track and session (short-term consistency; A versus A’; Figure 2D-F; Figure 3E; B versus B’ Figure S9D-F) or the first and the third (and last) block of five consecutive runs on the same track and session (long-term consistency, A versus A’’, Figure 2D, E, G; B versus B’’ Figure S9D, E, G) or the second and third block of five consecutive runs on the same track (A’ versus A’’, B’ versus B’’, late-term consistency, Figure S10F,H). The similarity of place fields between contexts (i.e., the inverse of remapping) was quantified by the correlation of mean activity maps for runs in context A and runs in context B, either for all the runs in each context (Figure 3D) or for the first and second blocks of 5 runs (taking place in contexts A and B, respectively; Figure 3E, comparison A-B). Finally, the trial-to-trial reliability (Figure 2C; Figure S9C) was measured by calculating the pairwise cross-correlations between the calcium signals of all individual runs in one session on the same track and then averaging the obtained values for each cell / bouton.

Population vector-based decoding

To decode position and context from neuronal activity data. We first split every dataset in two interleaved halves of template- and testing runs, respectively. Templates for population-vector based decoding were generated using the template runs by calculating the mean activity for each 5 cm bin on linear track A and B. The first and last bins on each track were omitted to mitigate potential errors due to small inhomogeneities in running speed, resulting in a total of 144 bins across both contexts. We then averaged neuronal activity from the testing data and calculated population activity vectors for each 100 ms bin. We then computed the Pearson correlation value for each of those population vectors with the template population vectors for each position. The most likely decoded location was then determined as the spatial bin that had the highest correlation with the population activity at a given time (Meyers, 2013). To obtain the mean spatial error, we calculated the absolute distance between the most likely decoded spatial location (irrespective of the decoded context) and the true location at each time point and averaged this distance across all time points. Context error for each time bin was either 0 if the decoded location was in the correct context or 1 otherwise. The mean context error was obtained by averaging over all time bins.

Cumulative context-decoding performance

Contextual decoding was often incorrect in individual 100 ms time bins, although on average it significantly predicted context in all recordings (Figure 4D). We hence reasoned that averaging over an increasingly larger number of 100 ms bins should increase the likelihood of decoding the correct context. To get a robust estimate for the time course over which such improvement might happen, we randomly drew 50 cells or boutons, respectively, from each recording to perform context-decoding as outlined above. We then calculated a decision value for the current context for increasingly longer time intervals (Δt). For each Δt, we divided the test data into non-overlapping segments of length Δt and calculated the average, binarized decision value for the current context across all 100 ms bins within this segment. If the mean context-decision fell closer to the correct context, that value for that individual segment was 0, otherwise it was 1. We then averaged over all segments of length Δt to obtain the mean accuracy of context decoding for segments of this size. The procedure was repeated for 10 different, random ensembles in each individual recording and an average accuracy curve was obtained for that recording. The time to 90% decoding accuracy was then determined as the first Δt for which the mean context-decoding accuracy exceeded 90% (Figure 4I).

Decoding using spatial versus mean-rate templates

To determine whether context-specific spatial maps carried contextual information (Figure 4G), we either constructed location- and context-stratified decoding templates (Figure S15B, left), as outlined above, or a simplified context-only template where solely the mean activity-rates across the entire linear track A and B, respectively, was used to construct population vectors for the decoding template (Figure S15B, right). We then compared decoding performance on the test runs for each of these templates per dataset by building a ratio between the two obtained error values. A ratio below 1 would indicate that decoding performance of the template using spatial and contextual information was superior to one that used context-dependent firing rate differences only (Figure 4G).

Identification of boutons showing grid-like patterns (Figure S7)

Among boutons having at least one place field (see STAR Methods and Figure 1), we looked for grid-like activity patterns. To do so, we applied the same approach as in Gu et al. (2018), which relies on 5 criteria: (1) a grid bouton must have at least two spatial fields on a track; (2) For a track length of 400 cm, as used here in both contexts, grid boutons must have a number of transitions between an in-field and out-of-field period larger than 400 / 5w, where w is the mean field width of a given bouton, in cm; (3) The widest field of a bouton must be smaller than 5w; (4) 30% or more of the bins must be assigned to either in-field or out-of-field periods; (5) The mean ΔF/F of in-field periods divided by the mean ΔF/F of out-of-field periods must be larger than 2. As this analysis was done on the fraction of boutons having at least one place field in one context, all boutons entering the analysis but not fulfilling these grid-like criteria were assigned to the place category (Figure S7D). We then looked for the maximal number of place fields found in either context A or B for the boutons identified as grid-like by this approach (Figure S7E). Finally, we compared place-field reliability, short and long-term consistencies and place field correlation between contexts of grid-like versus place boutons (Figure S7F-I).

Retrograde labeling of MEC cells projecting to the hippocampus using Cholera Toxin Subunit B (Figure S8)

Alexa Fluor 488-, 555- and 647-conjugated Cholera Toxin Subunit B (CTB488, CTB555 and CTB647, respectively) from Molecular Probes, Eugene, OR, USA were used at a concentration of 0.2% (dissolved in phosphate-buffered saline (PBS)). CTB injections were performed in a stereotactic apparatus (Kopf instruments) under anesthesia with 1.5%–2% isoflurane and analgesia using 0.1 mg.kg−1 buprenorphine. An eye lubricant ointment (Bepanthen, Bayer) was applied to protect corneal membranes during surgery. The skin was disinfected with 70% ethanol and Povidone-iodine (Betadine, Avrio Health L.P.) before surgical incision. Three small craniotomies (diameter 0.5-1 mm) were drilled over the hippocampus. Using glass micropipettes, each animal (n = 6 mice) was injected with ∼70-80 nL of each of the 3 CTB retrograde tracers in the DG (CTB488), CA3 (CTB555) and CA1 (CTB647) using the following coordinates: A/P = −2.00 / −2.00 / −2.00 mm from Bregma, M/L = 1.12 / 2.30 / 1.23 mm from midline, D/V = −1.87 / −1.92 / −1.23 mm from brain surface, respectively. After each slow injection of a retrograde tracer (∼2 min), the glass micropipette was left in situ for 10 min to ensure complete diffusion of the tracer in the parenchyma. 7 days after CTB injections, mice were deeply anaesthetized using a mixture of ketamine / xylazine (Sigma Aldrich), then perfused transcardially with 4% paraformaldehyde in PBS. Brains were cut into 80-μm-tick sagittal slices containing the MEC and hippocampus. Image stacks (single plane, 20x objective, N.A. = 0.8, 35291 × 44185 pixels) were acquired using a confocal microscope (Zeiss LSM900). For CTB-labeled cells counting, individual channels were exported and overlayed (registered in x-y), and MEC layers were drawn independently of the retrograde tracing by using DAPI labeling. Then, each channel in turn was viewed, and small colored dots according to each channel were applied to indicate the presence of retrograde labeling in a given cell (blind to cortical layer and the other channels). Overlapping (colocalized) retrograde labeling was calculated by viewing the dot labels for two channels at the same time, and the original images were also checked for verification. Cell numbers for each tracer (or combination) in each layer were then counted (and expressed as a percentage of total cells in that layer).

Statistics

The reported n indicate numbers of cells / boutons and exclude missing (‘NaN’) values. Unless otherwise stated, error bars are showing standard errors of the mean. All statistical tests are described in the corresponding figure legends. Unless otherwise stated, statistical comparisons were made between cells / boutons fulfilling the specific criteria as indicated in the figure legends. All comparisons were two-sided and all ANOVA tests were one-way tests (ANOVA on ranks, Dunn’s test). The null hypothesis was rejected at the p < 0.05 level.

Published: October 6, 2021

Data and code availability

• •Original data reported in this paper will be shared by the lead contact upon reasonable request.
• •All original code has been deposited at Zenodo.org and is publicly available as of the date of publication. Corresponding DOI is listed in the key resources table.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon reasonable request.