Fos ensembles encode and shape stable spatial maps in the hippocampus – Nature

Mice

All experimental procedures were approved by the Harvard Medical School Institutional Animal Care and Use Committee and were performed in compliance with the Guide for Animal Care and Use of Laboratory Animals. For muscimol inactivation experiments, data were collected from four adult wild-type C57BL/6J male mice (Jackson Laboratory, stock no. 000664). For Fos staining experiments, data were collected from four adult wild-type C57BL/6J male mice (Jackson Laboratory, stock no. 000664) and four Npas4-FH^24,54 mice (Greenberg laboratory, Harvard Medical School), split evenly between the control and exposed conditions. Npas4 staining was not analysed. For Fos-GFP reporter imaging experiments, data were collected from four adult Thy1-jRGECO1a⁵⁵ × B6.Cg-Tg(Fos-tTA,Fos-EGFP*)1Mmay/J double-transgenic male mice and seven B6.Cg-Tg(Fos-tTA,Fos-EGFP*)1Mmay/J transgenic male mice (Jackson Laboratory, stock no. 018306) injected with AAV (serotype 2/1) encoding a CAG-jRGECO1a red-shifted calcium indicator⁵⁶. As the Fos–GFP transgene in these mice is not under the control of doxycycline and the Tet transactivator was not used, doxycycline was not administered in this study. GFP was localized to the nucleus owing to a nuclear localization sequence (M. Mayford, personal communication). For Fos-KO imaging experiments, data were collected from six adult Fos^fl/fl;Fosb^fl/fl;Junb^fl/fl male mice^24,57. For all behaviour and imaging experiments, mice were at least 12 weeks old before the first data collection. For ex vivo experiments, data were collected from six Fos^fl/fl;Fosb^fl/fl;Junb^fl/fl male and female mice 4–6 weeks of age and six B6.Cg-Tg(Fos-tTA,Fos-EGFP*)1Mmay/J male and female mice 4–6 weeks of age.

Virtual reality and behavioural hardware

We used a miniaturized modified version of a visual virtual reality system⁵⁸ that has been described previously⁴². Head-restrained mice ran on an air-supported spherical treadmill that was constrained with a yaw and roll blocker to rotate only in pitch (forwards and backwards relative to the mouse’s body). Ball movement was detected by two optical sensors (ADNS-9800, Avago Technologies) connected to a Teensy-3.2 microcontroller (PJRC.com) mounted to a custom printed circuit board. Forward translation in the virtual environment was controlled by rotation of the ball, with velocity gain adjusted such that distance travelled in the virtual environment equalled the distance travelled on the surface of the ball. The virtual environment was back-projected (laser pico-bit projector, Celluon) onto a parabolic screen surrounding ~180 degrees of the mouse in azimuth, with a minimum screen distance from the mouse of approximately 5 inches. Designs for the virtual reality and behaviour hardware can be found at https://github.com/HarveyLab/mouseVR. Water rewards were delivered through a spout, with a solenoid valve controlling reward timing and quantity. Licks were detected by an electrical circuit triggered by contact with the lick spout.

Virtual environment

The virtual track was constructed using the Virtual Reality Mouse Engine (ViRMEn)⁵⁹ in MATLAB (Mathworks). The end of the track was continuous with its beginning, such that the track repeated with a circular topology. The walls of the track were tiled with textures to serve as visual landmarks. In the no-task and remapping experiments, two additional linear-track environments were introduced that were visually distinct with an inter-trial interval of 5 s.

Behaviour task

Mice were transported from the housing facility in a light-blocked cart and kept in a dim experimental room throughout the day to minimize unintended Fos expression. Mice were trained to lick for water rewards in a hidden reward zone one-tenth the length of the track (20 cm). Before being exposed to the virtual environment, mice were habituated to the apparatus and trained to run and lick the water spout to receive rewards. Once mice transitioned into the virtual environment, the task contingency was fixed and water rewards were delivered after the first lick in the reward zone. Occasionally, manual rewards were delivered to ensure that the lick detection and reward delivery systems were working; trials with manual rewards were excluded from further analysis. In the final version of the behavioural task, mice were required to traverse the linear track and lick in a specific reward zone to receive water rewards^48,60,61. Three trial types were present within each session: standard, crutch and probe. In standard trials (60–70% of trials), a water reward was delivered after the first lick in the reward zone. In crutch trials (20–30% of trials), a water reward was delivered as soon as the mouse entered the reward zone, regardless of licking behaviour. In probe trials (less than 10% of trials), no rewards were delivered, regardless of the mouse’s licking behaviour. Probe trials allowed us to assess licking and running behaviour in the absence of rewards. For standard and crutch trials, licks that occurred in the reward zone after the delivery of reward were deemed ‘consumption licks’ and did not contribute to measures of licking selectivity or numbers of licks. All other licks were considered ‘test licks’ (non-consumption licks). Muscimol inactivation experiments were performed in an earlier version of the task where the size of the reward zone was 40 cm rather than 20 cm.

Remapping experiments

Remapping experiments were performed in the absence of a task. On each session, mice were exposed to one of three environments: A, B or C. Mice were only exposed to a single environment each day. Imaging sessions began after mice had already been exposed to environment A; thus, environment A can be considered ‘familiar’. After up to five sessions of imaging in environment A, mice were moved to environment B for up to two sessions and then to environment C for up to a single session.

Behaviour analysis

To quantify the degree to which mice remembered and licked selectively near the reward zone, we compared licking in an area 10 cm immediately before the start of the reward zone (where mice exhibited anticipatory licking) to an equally sized area on the opposite side of the track (100 cm removed, not associated with reward). Licking selectivity was defined as the difference in the number of licks within each of these zones divided by their sum: (licks_pre-reward – licks_unrewarded)/(licks_pre-reward + licks_unrewarded). Licking selectivity ranges from −1 to 1, with 1 indicating licking only in the pre-reward zone and 0 indicating chance levels. All analysis was performed on stable performance sessions, defined as the first session in which mice reached a licking selectivity of 0.6 and any sessions thereafter. Learning sessions (Fig. 1c) were defined as the first three sessions in the environment, regardless of performance (not exclusive with stable performance sessions).

Muscimol

To test whether hippocampal activity was required for memory of the reward location, we inactivated CA1 using bilateral injections of muscimol (Sigma). A group of four wild-type male mice was used for these experiments. Trained mice were injected on alternate sessions with either muscimol (1 ng nl^–1 in extracellular saline) or saline for ten consecutive sessions. Craniotomies were made before experimental testing and covered with Kwik-Sil. On the days of testing, mice were briefly anaesthetized and injections were made bilaterally into CA1 ~1.4 mm below the dura. The injection pipettes were slowly retracted to minimize back-flow. Mice were returned to their home cage and allowed to recover for at least 90 min before behavioural testing. The injection volumes used were 500 nl (two sessions), 100 nl (four sessions) and 50 nl (four sessions). Only sessions containing at least ten trials with a lick were analysed, to exclude sessions where mice may have become unmotivated to attempt the task (five sessions excluded).

Histology

Mice were anaesthetized with 10 mg ml^–1 ketamine and 1 mg ml^–1 xylazine in PBS through intraperitoneal injection. When fully anaesthetized, the animals were transcardially perfused with 5 ml of ice-cold PBS followed by 20 ml of cold 4% paraformaldehyde (PFA) in PBS. Brains were dissected and post-fixed for 1 h at 4 °C in 4% PFA, followed by three washes (each for 30 min) in cold PBS. Coronal sections (40 μm thick) were subsequently cut using a Leica VT1000 vibratome and stored in PBS at 4 °C. For immunostaining, slices were permeabilized for 30 min at room temperature in PBS containing 0.2% Triton X-100. Slices were blocked for 1 h at room temperature with PBS containing 0.2% Triton X-100, 2% normal donkey serum and 0.1% fish gelatin. Slices were incubated in primary antibodies diluted in blocking solution at 4 °C for 24 h: mouse anti-Fos (Abcam, ab208942; 1:1,000) or rabbit anti-Fos (Synaptic Systems, 226003; 1:3,000). Slices were then washed three times each with PBS for 10 min at room temperature, incubated for 2 h at room temperature with secondary antibodies conjugated to Alexa dye (Life Technologies; anti-rabbit Alexa 488 (A21206), anti-rabbit Alexa 555 (A31572), anti-mouse Alexa 488 (A21202) or anti-mouse Alexa 555 (A31570); 1:250) and washed three times with PBS. Slices were then mounted in DAPI Fluoromount-G (Southern Biotech) and imaged on a virtual slide microscope (Olympus, VS120). To quantify Fos expression across different brain regions, 200 μm × 200 μm non-overlapping regions of interest (ROIs) were manually selected tiling each of the four brain regions (dentate gyrus, CA1, retrosplenial area and primary somatosensory area), when present, in each slice. Slices that were excessively damaged, exhibited severe imaging artefacts or had extensive non-specific antibody labelling were excluded at this stage. The number of cells was then counted within these ROIs. Cells were counted in an average of 10.72 ROIs per brain region in each mouse (total of 343 ROIs). ROIs were selected pseudorandomly and presented to the experimenter during counting. The experimenter was blinded to mouse identity and experimental group during both ROI selection and cell counting.

Surgery: virus injection

Virus injections were carried out in mice before they were put on a water schedule. Craniotomies were centred around 1.8 mm lateral to the midline (right hemisphere) and −2.3 mm posterior to bregma. The approximate locations of the three craniotomies were (1.55, −2.3), (1.93, −2.08) and (1.93, −2.52) mm (medial/lateral (ML) and anterior/posterior (AP) axes, respectively) from bregma. Virus injections were performed using bevelled glass micropipettes ~1.3 mm below the dura. For population imaging in Fos-TetTag mice, 60 nl of AAV2/1-CAG-jRGECO1a (1 × 10¹¹ genome copies per ml) was injected at each site. For sparse Cre infection, 100 nl of a 1:1 mixture of AAV2/1-CAG-Cre-GFP (1 × 10¹¹ genome copies per ml) and AAV2/1-CAG-jRGECO1a (1 × 10¹¹ genome copies per ml) was injected into each of the three locations. No virus injections were performed in the double-transgenic (Fos-tTA,Fos-shEGFP and Thy1-jRGECO1a) mice.

Surgery: cannula implantation

Cannula implantation for hippocampal imaging was performed on mice after water schedules had started, at approximately 90% of initial body weight. The hippocampal window and headplate implantation surgeries were carried out as described previously^43,62. During the cannula implantation surgery, dental cement was used to attach a titanium headplate to the skull parallel to the surface of the hippocampal window.

Two-photon imaging

Data were collected using a custom-built resonant-scanning two-photon microscope. An air-supported spherical treadmill was mounted on a three-axis translation stage (Dover Motion) to position the mouse with respect to the objective (Nikon ×16, 0.8 NA water immersion). Two-photon excitation of jRGECO1a was achieved using a mode-locked diode-pumped femtosecond laser at 1,040 nm (YBIX, Time-bandwidth) or 1,070 nm (Fidelty-2, Coherent). A Ti:sapphire laser (Coherent Chameleon Vision II) was used to deliver 920-nm pulsed excitation for GFP imaging. Imaging took place at either 920 nm for GFP or 1,040/1,070 nm for jRGECO1a. Emitted light was filtered and collected by a GaAsP photomultiplier tube (Hamamatsu). The microscope was controlled by ScanImage 2019 (Vidrio Technologies). Images were acquired at 30 Hz at a resolution of 512 × 512 pixels corresponding to a field of view of 448 × 448 μm² for two mice and 768 × 768 μm² for ten mice. To synchronize functional jRGECO1a imaging and behavioural data, the imaging frame clock and a subset of behavioural signals were recorded in pClamp (Molecular Devices) at 1 kHz. After recording, the full set of behavioural signals and task data collected in ViRMEn were synchronized with the imaging clock and downsampled to the imaging frame rate (30 Hz), using linear or nearest-neighbour interpolation when necessary. To measure either Fos promoter-driven GFP or Cre–GFP, image stacks were acquired centred in the z axis on the imaging plane. These volumes typically consisted of 40 planes (512 × 512 pixels, the same resolution as in jRGECO1a imaging) separated by 4 µm, with 200 frames per plane. Both green and red channels were acquired for post hoc registration of the volume to the functional imaging plane.

Maintaining the same field of view within and across imaging sessions

We used a custom headplate holder designed for reproducible day-to-day mounting of the mouse on the ball. Before imaging, the cannula and window were cleaned using rounds of filtered water and gentle vacuum suction to remove fine dust and debris. The mouse was positioned under the objective, and the field of view was manually aligned with a reference image taken on day 1 of the experiment. To maintain a consistent axial plane during imaging, a subset of recently acquired frames were registered online to a reference stack to estimate displacements from the target plane. These results were plotted throughout the experiment to guide periodic small manual adjustments countering axial and lateral drift. Post hoc assessment of drift and image quality was performed by manually examining sped-up and downsampled movies of the entire experiment after motion correction⁶³. Experiments with insufficiently stable imaging quality or imaging plane were excluded before the start of in-depth analysis.

Motion correction

Motion correction of jRGECO1a calcium imaging movies was performed offline using a custom MATLAB pipeline (available at https://github.com/HarveyLab/Acquisition2P_class) as previously described^63,64. Motion correction accounted for non-rigid deformations taking place on subframe, full-frame and minute-long timescales. GFP stacks were registered in MATLAB using non-rigid registration (NoRMCorre)⁶⁵ for frames within each plane, followed by an fast Fourier transform (FFT)-based rigid registration across planes. Stacks were aligned to the imaging plane using the red channel, and GFP planes out of plane with the imaging target plane were discarded from further analysis.

Fluorescence source extraction and classification

After motion correction, spatial footprints of fluorescence sources in calcium movies were identified using Suite2p⁶⁶ (Python version, https://github.com/MouseLand/suite2p). The resulting sources were classified into two groups: putative cell body and non-cell body sources. Classification was performed using a simple convolutional neural network trained in MATLAB on manually labelled data, using a network architecture, hyperparameters and training procedure described previously⁶⁴ with the exception of two output classes rather than three. Non-cell sources were discarded from further analysis.

Fluorescence trace preprocessing

We monitored neural activity as calcium transients, which previous work has shown to be related to spiking in pyramidal neurons^34,35,36,56, using the red-shifted calcium indicator jRGECO1a. Raw traces extracted by Suite2p were further processed as follows: a baseline fluorescence estimate was computed as the 30th percentile in a 60-s moving window. ∆F/F traces were computed by subtracting and dividing the raw trace by the baseline. For most analyses, significant transient traces were used^52,67,68. Significant transients were identified using the following procedure, based on previous work⁶⁸ (Extended Data Fig. 4). ∆F/F traces were standardized by subtracting the median and dividing by the standard deviation. Threshold levels in units of standard deviation (σ) were chosen between 1 and 4 in 0.2σ increments. For every threshold level t, putative transients were identified as positive samples exceeding that threshold. For each putative transient n frames in length, we estimated a false positive rate as the number of negative-going transients (<–t) with at least n frames divided by the number of positive-going transients with at least n frames. Transients with a false positive rate of less than 0.001 (0.1%) were considered to be significant at that threshold level t. Frames that were significant at at least one threshold level were considered significant in the final output (union across threshold levels). Finally, transients separated by less than two frames were merged, and transients less than two frames in duration were removed. The final significant transient traces are original zero-baseline ∆F/F traces, where all frames without significant transients were set to zero.

Cross-day alignment of sources

For each mouse, a reference session was chosen for alignment (approximately halfway between the first and last days of imaging, to maximize similarity to other imaging sessions). All sessions were registered with the reference session using non-rigid alignment on the mean jRGECO1a images (NoRMCorre)⁶⁵. Source masks were transformed using this alignment to place all source masks for a mouse into a common reference space. CellReg⁶⁹ (a distance model) was used on these aligned images to match cell IDs across different sessions.

Fos induction measurements

Changes in Fos promoter-driven GFP expression were measured using time-lapse two-photon imaging^29,70. To compute the fold change in GFP fluorescence 2–4 h after the start of behaviour, the registered post-behaviour GFP image was divided by the pre-behaviour GFP image. To account for changes due to non-uniformities in brightness across the field of view, the resulting image was normalized by dividing by an estimated background image. The background image was obtained by two-dimensional median filtering the fold-induction image with a filter size of ~50 × 50 µm², approximately 30 times the area of individual cell sources. This normalization accounts for changes in fluorescence that could be due to changing imaging conditions rather than changing GFP expression, including non-uniform changes across the field of view, and it assumes that GFP fluorescence is relatively sparse and primarily localized to the nucleus, assumptions that are well supported by the literature. Regions of the fold-induction image that contained artefacts due to errors in registration or large changes in imaging quality were manually excluded from further analysis. For each cell source, fold-induction values were determined by averaging the pixel values of the fold-induction image within a circular ROI centred on the cell body, 10 µm in diameter. Fos-high cells were defined as cells with the 20% highest fold-induction values on a given session, and Fos-low cells were defined as cells with the 20% lowest fold-induction values on a given session. To determine the time course of induction (Extended Data Fig. 2), the same procedure was used with the addition that, before computing the fold-induction images, all images in the time course were standardized using the following procedure. Each image was standardized by its median and median absolute deviation and then rescaled by a multiplicative factor A and additive offset B, where A was the median absolute deviation of pixel values across all images in the series and B was the median pixel value of all images in the series. The median and median absolute values were chosen under the same rationale and set of assumptions as described above. The recovered time course was consistent with previous studies using this, and other, Fos-GFP reporters^17,29,30.

Place cell definition

For all place field analyses, trials were only considered if they met the following criteria: at least three licks, duration between 4 and 60 s, no experimenter-triggered rewards and occurring before 1.2 ml of cumulative reward had been delivered. In addition, following convention in the field, place cell characteristics were computed using only frames during running, defined as when speed exceeded 5 cm s^–1. To assess significant spatial tuning, the track was divided into 40 bins, each 5 cm in size. For each cell, the average activity within each bin was computed, and the resulting binned activity was smoothed with a Gaussian kernel with a standard deviation of 1 bin, or 5 cm, with circular padding, consistent with the circular topology of the track. Significant peaks in the spatially binned activity were determined by a shuffle test. On each shuffle (n = 1,000 shuffles), the behaviour was circularly shifted by a random number of frames and then divided into six blocks of roughly equal duration and the order of those blocks was randomly permuted. This procedure perturbs the relationship between neural activity and behaviour while maintaining the temporal and autocorrelation structure of the activity and behaviour. For each shuffle, the same spatial binning and smoothing were performed. Peaks in the true binned activity that exceeded the 99th percentile of the shuffle distribution for at least three consecutive spatial bins were deemed significant peaks. Apart from the lower bound on the size of significant fields (imposed to limit false positives), we remained agnostic to the shape, amplitude and number of fields: any cell with significant peaks is referred to as a place cell, with the significant peaks of that cell corresponding to its place field.

Spatial information

In addition to defining place fields, we computed spatial information for each cell in the population, regardless of whether it had a significant field. Spatial information was defined as the mutual information H between neural activity and position, using the following formula:

where i is the ith spatial bin, a is the overall mean activity, a_i is the mean activity in the ith bin and p_i is the fraction of time spent in the ith bin. For each cell, we computed the real value and the value for each of 1,000 shuffles (see ‘Place cell definition’). We report the normalized spatial information for each cell as the real information divided by the mean across shuffles⁷¹.

Place field properties

The trial-to-trial correlation of a place cell was defined as the mean of off-diagonal elements in the n trial × n trial Pearson correlation matrix of a cell’s trial-wise spatially binned activity. Entries comparing pairs of trials that lacked activity were set to zero. To compute the fraction of trials in which a place field was active, trials with at least one significant transient within that cell’s place field were considered ‘active’. To compute the selectivity of place fields, we measured the average spatially binned activity within a cell’s significant fields (in-field activity) and the average spatially binned activity outside a cell’s significant fields (out-of-field activity). Selectivity was defined as the difference between these values divided by their sum: (in-field activity – out-of-field activity)/(in-field activity + out-of-field activity).

Position decoder

Naive Bayes’ decoders were used to decode position^72,73 from the activity of groups of neurons within each session. The decoder assumed Poisson firing and independence between neurons and adopted a uniform prior for all spatial bins. The inclusion criteria for frames used for either training or testing were the same as those used for place field calculation. Even-numbered trials were always used for training (computing place field activity templates) while odd-numbered trials were used for testing. Decoding error was defined as the absolute difference between the true spatial bin and the decoded bin, ranging between 0 and 100 cm owing to the circular structure of the track. For each frame, the posterior probability of the mouse being in a given position bin ‘pos’ was computed as

where a_group is the activity of all cells in the group being used for decoding, C is the normalization constant, τ is the temporal bin size of one frame (one-thirtieth of a second), N is the total number of cells and for each cell f_i(pos) is the spatially binned activity template and a_i is the activity on the frame. The bin with the maximum posterior probability was selected as the decoded position on that frame. To compare decoder performance across cell groups, equally sized groups were used for decoding. For Fos-GFP imaging experiments, a decoder was trained and tested on each induction decile group independently. Fos-low and Fos-high decoding performance was the average performance of deciles 1–2 and 9–10, respectively. To compare Cre⁺ and wild-type groups, the number of cells used for decoding was selected on the basis of the size of the smaller cell group, with a minimum of 10 cells up to a maximum of 100 cells used for decoding. Each group was then randomly downsampled to this number of cells, and the decoder was trained and tested using these equally sized populations. The downsampling, training and testing procedure was repeated 100 times, and decoder performance was averaged for each group across these repetitions. Sessions were only included if they contained at least ten cells in the smaller cell group. The code for the naive Bayes’ decoder was modified from code available on the Buzsaki laboratory’s GIT repository (https://github.com/buzsakilab/buzcode).

Many factors can contribute to variability in the mean decoding accuracy across different positions of the track, including running speed, the configuration of visual cues⁷⁴ and behavioural variability. In particular, higher running speeds in the region of the track away from the reward zone, combined with the relatively slow kinetics of the calcium indicator, probably contributed to the increased decoding error there. We therefore focused on comparing groups of neurons recorded simultaneously and treated identically during analysis. Decoding analyses were always performed on size-matched subpopulations, and testing and training frames were identical for each subpopulation tested.

Activity matching

For activity-matched analyses, activity matching was always performed between Fos-high and Fos-low neurons on a given session. Activity matching was carried out on integrated ∆F/F (henceforth α), which is proportional to mean ∆F/F activity. First, activity bin edges were determined by binning log₁₀(α) from the entire population into ten bins (using MATLAB’s histcounts function). Log-transformed activity was chosen to determine bin edges because the raw activity distribution has a long positive tail that would otherwise lead to low binning resolution for the majority of α values and relatively sparse counts in the bins with higher α values. Within each activity bin, Fos-high and/or Fos-low cells were subsampled to equal amounts pseudorandomly using the following iterative procedure. Fos-high and Fos-low neurons within the bin were randomly drawn from either the most active or least active half (for that bin) depending on whether the mean α of all subsampled Fos-high neurons up to that iteration was higher or lower than the mean α of all subsampled Fos-low neurons. For example, if the mean subsampled Fos-high α was higher than the Fos-low α, the next iteration would draw from the least active Fos-high neurons in that bin and the most active Fos-low neurons in that bin. This procedure continued until either all the Fos-high or all the Fos-low neurons in that bin were drawn, whichever came first. This had the effect of not only matching the binned activity distributions to be identical, but also balancing mean activity between Fos-high and Fos-low populations.

Classification of Fos-KO cells

After alignment across days, individual cells in mice with conditional Fos knockout were manually examined for Cre–GFP expression. Cre probability maps were produced from aligned Cre–GFP images using Cellpose⁷⁵. Cells were manually examined taking into account their source mask as well as jRGECO1a, Cre–GFP and Cre probability images. Cells were manually labelled as Cre⁺ (Fos-KO), Cre^– (wild type) or ambiguous. Ambiguous cells were excluded from further analysis.

Ensemble analyses

The following analyses were performed only on cells with significant place fields. To identify groups of place cells that had correlated activity across trials, independently of the positioning of their place fields in the track, we quantified the level of activity of each cell within its place field(s) on each eligible trial (see ‘Place field definition’) within the session: we refer to this as the trial-wise place field activation. Within-group correlation was quantified as the mean pairwise correlation of trial-wise place field activation within each group. To visualize and quantify group structure within the trial-to-trial variability in the population, we clustered cells with affinity propagation⁷⁶, using as similarity their pairwise correlations in trial-wise place field activation (preference parameter, −1; max iterations, 5,000). To analyse within-group correlations across days, we examined pairs of sessions within the same animal separated by the variable ∆sessions. We will refer to the earlier session as the reference session and the subsequent session as the target session: for example, comparing sessions 3 and 4 corresponds to a ∆+1 pair, where the reference session is 3 and the target session is 4. Because we only considered pairs with a positive ∆ value, each unique pair of sessions was considered at most once in the analysis. For every pair of sessions with a positive ∆ value where the reference session had Fos induction data, the cell groups (Fos-high and Fos-low) were defined by induction quantile (top 20% and bottom 20%, respectively) on the reference session. We then computed the within-group correlation between these cells in the target session, as described above.

Place field stability analyses

To analyse the relationship between Fos induction or Fos disruption and place field stability, we took the approach of comparing pairs of sessions separated by differing numbers of sessions within the same animal. For clarity, we will refer to one session as the reference session and the other as the target session, where the difference between them (∆sessions) was measured as the target minus reference. We only considered pairs with a positive ∆sessions value, such that any pair of sessions was considered at most once, and the target session always occurred after the reference session. Fos induction cell groups were defined by induction values on the reference session. For each pair of sessions, we examined the stability of cells that were members of a specific cell group (Fos-high, Fos-low, Fos-KO, Fos-WT), had significant place fields on the reference session and were present on both sessions. At least 20 cells within each group were required for a session pair to be considered for stability analysis. For each cell group within each valid session pair, we computed stability maps that described place field stability as a function of track position. First, we computed for each cell the Pearson correlation between spatially binned activity on the reference and target sessions, producing the place field correlation vector x. Stability at each bin i along the track was computed as a weighted mean of the vector x, where the weight vector w for bin i was the normalized activity of each place cell in that bin. Thus, if a place cell exhibits no activity in a particular bin, it will not contribute to the mean stability of that bin while it will contribute strongly to the stability of bins around the peak of its place field. Before computing the weights, the spatially binned activity of each place field was normalized to sum to one so that the contribution of each place field to the overall stability map was equal regardless of the cell’s activity level.

Ex vivo whole-cell electrophysiology and calcium imaging

AAV injection (Fos
^fl/fl;Fosb
^fl/fl;Junb
^fl/fl mice)

Before ex vivo electrophysiological recordings, Fos^fl/fl;Fosb^fl/fl;Junb^fl/fl mice were injected with AAV-Cre-GFP to conditionally disrupt the Fos complex in a sparse subset of CA1 neurons, as previously described²⁴. Surgical preparation and procedures up to virus injection were carried out as described in the surgery methods for in vivo experiments. Mice at 3–4 weeks of age were anaesthetized by isoflurane inhalation and placed in a stereotaxic frame (Kopf, model 1900). A small craniotomy was made on the dorsal surface of the skull, and virus injection was performed bilaterally using bevelled glass micropipettes targeted to CA1 (ML: ±3.0 mm; AP: −2.4 mm; dorsal/ventral, DV: −2.8 mm). Approximately 1 μl of AAV2/1-CAG-Cre-GFP (titre of 1.751 × 10¹¹ genome copies per ml) was injected at a rate of 150 nl min^–1, after which the pipette was left in place for approximately 5 min to allow the virus to diffuse and pressure to dissipate before retraction. Mice were allowed to recover for at least 7 days to allow for virus expression and recombination to occur.

Acute slice preparation

Mice were placed in an enriched environment before recording (at least 3 h for Fos reporter mice and at least 2 days for Fos^fl/fl;Fosb^fl/fl;Junb^fl/fl mice). The enriched environment set-up consisted of a large opaque cage (0.66 m × 0.46 m × 0.38 m) containing an assortment of toys and objects such as a running wheel, tubes, ladders, platforms, huts and different kinds of animal bedding. Food pellets were scattered throughout to encourage exploration.

Acute slice preparation was carried out as previously described²⁴. Mice aged 4–5 weeks were anaesthetized by ketamine/xylazine injection and transcardially perfused with ice-cold oxygenated (95% O₂/5% CO₂) choline-based artificial cerebrospinal fluid (choline-ACSF) consisting of (in mM) 110 choline chloride, 25 NaHCO₃, 1.25 NaH₂PO₄, 2.5 KCl, 7 MgCl₂, 25 glucose, 0.5 CaCl₂, 11.6 sodium l-ascorbate and 3.1 sodium pyruvate. Transverse 300-μm hippocampal slices were prepared from both hemispheres using a vibratome (Leica, VT1000). The dorsal hippocampal slices were collected in a holding chamber containing oxygenated ACSF consisting of (in mM) 127 NaCl, 25 NaHCO₃, 1.25 NaH₂PO₄, 2.5 KCl, 1 MgCl₂, 10 glucose and 2 CaCl₂. The slices were then incubated at 32 °C for 20 min followed by 30 min at room temperature before recordings. Recordings were carried out at room temperature. For all solutions, pH was set to 7.2 and osmolarity was set to 280–290 mOsm.

Ex vivo electrophysiology and calcium imaging

For whole-cell current-clamp recordings, a K⁺-based internal solution consisting of (in mM) 142 potassium gluconate, 4 KCl, 10 HEPES, 4 magnesium ATP, 0.3 sodium GTP, 10 sodium phosphocreatinine and 1.1 EGTA (pH 7.2, 280 mOsm) was used. To record calcium transients, cells were filled with the synthetic red calcium indicator Cal-590 dextran (molecular weight, 4,000) at a 100 μM concentration. In all recordings, neurons were held at −70 mV with patch pipettes made with borosilicate glass with filament (Sutter, BF150-86-7.5) with open pipette resistance of 2–4 MΩ.

Simultaneous electrophysiological recordings and calcium imaging were performed on an upright Olympus BX51 WI microscope with an sCMOS camera (Zyla 5.5 sCMOS, Oxford Instruments), ×60 water-immersion objective (Olympus Lumplan Fl/IR ×60/0.90 NA) and light-emitting diode (Excelitas XCite LED120). Electrophysiological data were acquired using Clampex 10.6 (Molecular Devices). Data were low-pass filtered at 4 kHz, sampled at 10 kHz with an Axon Multiclamp 700B amplifier and digitized with an Axon Digidata 1440A data acquisition system. Experiments were discarded if the holding current exceeded −200 pA or if the series resistance was greater than 30 MΩ. Image acquisition was performed using MicroManager. Timestamps of frame acquisition were recorded on Clampex to allow for synchronization of electrophysiology to calcium images.

To measure spike-evoked calcium influx, each protocol consisted of five sweeps each with 1, 3, 5, 7 or 9 current injection pulses. Current was injected with a 4-ms square-wave pulse every 20 ms to evoke 50-Hz spike trains. The amplitude of the current injection was set by the experimenter to be slightly greater than the minimum value required to reliably evoke spikes with a 4-ms pulse, measured in 50-pA increments. Sweeps in which the number of spikes did not equal the number of current pulses were excluded from analysis. Between one and five repetitions of the protocol were averaged together to produce the spike-evoked fluorescence for each cell.

To measure the intrinsic electrophysiological properties of cells, a current–voltage (IV) curve protocol was carried out in a subset of cells. This protocol consisted of 500-ms current injections between 150 and 600 pA in 50-pA increments. Steady-state voltage used to construct IV curves was computed as the average membrane voltage in the last 100 ms of current injection. Input resistance was computed from the slope of the IV curve on current steps before the first current step in which a spike was evoked. The membrane time constant was computed using a single exponential fit to the membrane voltage polarization in response to the 50-pA negative current injection, using MATLAB’s lsqcurvefit function. Action potential waveform properties were computed on the first evoked action potential in the protocol. Action potential width was measured as the full width halfway in amplitude between the action potential threshold and the action potential peak.

Ex vivo electrophysiology and calcium imaging analysis

All ex vivo electrophysiology and calcium imaging analyses were carried out using custom scripts in MATLAB. Calcium imaging data were synchronized with traces using frame exposure trigger signals collected in the same time base as intracellular voltage and current injection data. To extract fluorescence traces from each cell, one rectangular ROI was drawn over the soma of the cell and another larger rectangular ROI was drawn outside the soma to estimate background fluorescence. Pixels within these ROIs were averaged on each frame to yield raw somatic and background fluorescence traces. To compute ∆F/F, the background fluorescence trace was subtracted from the somatic trace to yield a fluorescence trace F. F₀ was computed for each sweep as the average F in the 200 ms before current injection. ∆F/F was then computed for each sweep as (F – F₀)/F_0. Between one and five repetitions of the protocol were averaged together to yield a set of five sweep traces per cell corresponding to 1, 3, 5, 7 and 9 action potentials triggered at 50 Hz. We observed that the decay kinetics appeared to have two components: fast and slow. This is consistent with our imaging picking up a mixture of somatic calcium dynamics, which tend to be lower amplitude and slower, and dendritic calcium dynamics, which tend to be higher amplitude and faster. Nuclear calcium dynamics may also be present as the dye is not restricted from the nucleus. This is expected because epifluorescence imaging does not allow excitation to be primarily restricted to either the soma or the dendrite. Decay kinetics were therefore fit with a double-exponential decay using MATLAB’s lsqcurvefit function.

Data analysis, statistics and reproducibility

Data were analysed in MATLAB (2019a and 2021b) together with the DataJoint toolbox^77,78 (version 3.3.2). No statistical method was used to predetermine sample size. Sample sizes in terms of mice, sessions and neurons are similar to those in other contemporary studies in the field¹⁶. For in vivo imaging experiments, comparisons were made between simultaneously recorded neurons within mice that were subjected to the same experimental conditions; randomization across subjects and blinding to experimental conditions were not necessary and did not take place during calcium imaging experiments or analysis. During manual classification of Cre⁺ cells, the experimenter did not have access to information about the activity or functional properties of the cells. For manual cell counting of labelling with anti-Fos antibody, the experimenter was blinded to mouse identity and experimental group during both brain region ROI selection and cell counting (see ‘Histology’). For hypothesis testing, we used permutation shuffle tests, two-sided paired-sample t tests and two-sided two-sample t tests, showing both individual data points and difference histograms when possible. In Fig. 1i, representative images are shown from one session of the 27 sessions in 6 mice that were analysed in Fig. 1j–n. In Fig. 3b, representative images (cropped from the full field of view for the purposes of visualization) are shown from one mouse of the 6 mice analysed in Fig. 3. In Fig. 5a, two representative fields of view are shown from one mouse of the 6 mice analysed in Fig. 5.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.